Apr 18, 2013, 7:02:21 PM (6 years ago)

Initial translation. Special characters, figures, tables, bib, to go.

1 edited


  • docs/Balisage13/Bal2013came0601/Bal2013came0601.xml

    r3038 r3039  
    44<article xmlns="http://docbook.org/ns/docbook" version="5.0-subset Balisage-1.3"
    55  xml:id="HR-23632987-8973">
    6    <title>Parallel Bit Stream Technology as a Foundation for XML Parsing Performance</title>
     6   <title></title>
    77   <info>
    1515      <abstract>
    16          <para>By first transforming the octets (bytes) of XML texts into eight parallel bit
    17             streams, the SIMD features of commodity processors can be exploited for parallel
    18             processing of blocks of 128 input bytes at a time. Established transcoding and parsing
    19             techniques are reviewed followed by new techniques including parsing with bitstream
    20             addition. Further opportunities are discussed in light of expected advances in CPU
    21             architecture and compiler technology. Implications for various APIs and information
    22             models are presented as well opportunities for collaborative open-source
    23          development.</para>
     16         <para>Prior research on the acceleration of XML processing
     17using SIMD and multi-core parallelism has lead to
     18a number of interesting research prototypes.  This work
     19investigates the extent to which the techniques underlying
     20these prototypes could result in systematic performance
     21benefits when fully integrated into a commercial XML parser.
     22The widely used Xerces-C++ parser of the Apache Software
     23Foundation was chosen as the foundation for the study.
     24A systematic restructuring of the parser was undertaken,
     25while maintaining the existing API for application programmers.
     26Using SIMD techniques alone, an increase in parsing speed
     27of at least 50% was observed in a range of applications.
     28When coupled with pipeline parallelism on dual core processors,
     29improvements of 2x and beyond were realized.
    2431      </abstract>
     32      <author>
     33         <personname>
     34            <firstname>Nigel</firstname>
     35            <surname>Medforth</surname>
     36         </personname>
     37         <personblurb>
     38            <para></para>
     39         </personblurb>
     40         <affiliation>
     41            <jobtitle></jobtitle>
     42            <orgname></orgname>
     43         </affiliation>
     44         <email></email>
     45      </author>
     46      <author>
     47         <personname>
     48            <firstname>Dan</firstname>
     49            <surname>Lin</surname>
     50         </personname>
     51         <personblurb>
     52            <para></para>
     53         </personblurb>
     54         <affiliation>
     55            <jobtitle></jobtitle>
     56            <orgname></orgname>
     57         </affiliation>
     58         <email></email>
     59      </author>
     60      <author>
     61         <personname>
     62            <firstname>Kenneth</firstname>
     63            <surname>Herdy</surname>
     64         </personname>
     65         <personblurb>
     66            <para> Ken Herdy completed an Advanced Diploma of Technology in Geographical Information
     67               Systems at the British Columbia Institute of Technology in 2003 and earned a Bachelor
     68               of Science in Computing Science with a Certificate in Spatial Information Systems at
     69               Simon Fraser University in 2005.
     70                                                </para>
     71            <para> Ken is currently pursuing graduate studies in Computing Science at Simon Fraser
     72               University with industrial scholarship support from the Natural Sciences and
     73               Engineering Research Council of Canada, the Mathematics of Information Technology and
     74               Complex Systems NCE, and the BC Innovation Council. His research focus is an analysis
     75               of the principal techniques that may be used to improve XML processing performance in
     76               the context of the Geography Markup Language (GML).
     77                                                </para>
     78         </personblurb>
     79         <affiliation>
     80            <jobtitle>Graduate Student, School of Computing Science</jobtitle>
     81            <orgname>Simon Fraser University </orgname>
     82         </affiliation>
     83         <email>ksherdy@sfu.ca</email>
     84      </author>
    2585      <author>
    2686         <personname>
    3696               patentleft evangelist, advocating university-based technology transfer models
    3797               dedicated to free use in open source. </para>
    3998         </personblurb>
    4099         <affiliation>
    46105      <author>
    47106         <personname>
    48             <firstname>Ken</firstname>
    49             <surname>Herdy</surname>
     107            <firstname>Arrvindh</firstname>
     108            <surname>Shriraman</surname>
    50109         </personname>
    51110         <personblurb>
    52             <para> Ken Herdy completed an Advanced Diploma of Technology in Geographical Information
    53                Systems at the British Columbia Institute of Technology in 2003 and earned a Bachelor
    54                of Science in Computing Science with a Certificate in Spatial Information Systems at
    55                Simon Fraser University in 2005. </para>
    56             <para> Ken is currently pursuing graduate studies in Computing Science at Simon Fraser
    57                University with industrial scholarship support from the Natural Sciences and
    58                Engineering Research Council of Canada, the Mathematics of Information Technology and
    59                Complex Systems NCE, and the BC Innovation Council. His research focus is an analysis
    60                of the principal techniques that may be used to improve XML processing performance in
    61                the context of the Geography Markup Language (GML). </para>
     111            <para></para>
    63112         </personblurb>
    64113         <affiliation>
    65             <jobtitle>Graduate Student, School of Computing Science</jobtitle>
    66             <orgname>Simon Fraser University </orgname>
     114            <jobtitle></jobtitle>
     115            <orgname></orgname>
    67116         </affiliation>
    68          <email>ksherdy@cs.sfu.ca</email>
    69       </author>
    70       <author>
    71          <personname>
    72             <firstname>Ehsan</firstname>
    73             <surname>Amiri</surname>
    74          </personname>
    75          <personblurb>
    76             <para>Ehsan Amiri is a PhD student of Computer Science at Simon Fraser University.
    77                Before that he studied at Sharif University of Technology, Tehran, Iran. While his
    78                graduate research has been focused on theoretical problems like fingerprinting, Ehsan
    79                has worked on some software projects like development of a multi-node firewall as
    80                well. More recently he has been developing compiler technology for automatic
    81                generation of bit stream processing code. </para>
    83          </personblurb>
    84          <affiliation>
    85             <jobtitle>Graduate Student, School of Computing Science</jobtitle>
    86             <orgname>Simon Fraser University</orgname>
    87          </affiliation>
    88          <email>eamiri@cs.sfu.ca</email>
     117         <email></email>
    89118      </author>
    97126      <keywordset role="author">
    98127         <keyword/>
    99          <keyword/>
    100          <keyword/>
    101128      </keywordset>
    102130   </info>
    103131   <section>
    104132      <title>Introduction</title>
    105       <para> While particular XML applications may benefit from special-purpose hardware such as XML
    106          chips [<xref linkend="XMLChip09"/>] or appliances [<xref linkend="Datapower09"/>], the bulk
    107          of the world's XML processing workload will continue to be handled by XML software stacks
    108          on commodity processors. Exploiting the SIMD capabilities of such processors such as the
    109          SSE instructions of x86 chips, parallel bit stream technology offers the potential of
    110          dramatic improvement over byte-at-a-time processing for a variety of XML processing tasks.
    111          Character set issues such as Unicode validation and transcoding [<xref linkend="PPoPP08"
    112          />], normalization of line breaks and white space and XML character validation can be
    113          handled fully in parallel using this representation. Lexical item streams, such as the bit
    114          stream marking the positions of opening angle brackets, can also be formed in parallel.
    115          Bit-scan instructions of commodity processors may then be used on lexical item streams to
    116          implement rapid single-instruction scanning across variable-length multi-byte text blocks
    117          as in the Parabix XML parser [<xref linkend="CASCON08"/>]. Overall, these techniques may be
    118          combined to yield end-to-end performance that may be 1.5X to 15X faster than alternatives
    119             [<xref linkend="SVGOpen08"/>].</para>
    120       <para>Continued research in parallel bit stream techniques as well as more conventional
    121          application of SIMD techniques in XML processing offers further prospects for improvement
    122          of core XML components as well as for tackling performance-critical tasks further up the
    123          stack. A newly prototyped technique for parallel tag parsing using bitstream addition is
    124          expected to improve parsing performance even beyond that achieved using sequential bit
    125          scans. Several techniques for improved symbol table performance are being investigated,
    126          including parallel hash value calculation and length-based sorting using the cheap length
    127          determination afforded by bit scans. To deliver the benefits of parallel bit stream
    128          technology to the Java world, we are developing Array Set Model (ASM) representations of
    129          XML Infoset and other XML information models for efficient transmission across the JNI
    130          boundary.</para>
    132       <para>Amplifying these software advances, continuing hardware advances in commodity processors
    133          increase the relative advantage of parallel bit stream techniques over traditional
    134          byte-at-a-time processors. For example, the Intel Core architecture improved SSE processing
    135          to give superscalar execution of bitwise logic operations (3 instructions per cycle vs. 1
    136          in Pentium 4). Upcoming 256-bit AVX technology extends the register set and replaces
    137          destructive two-operand instructions with a nondestructive three-operand form. General
    138          purpose programming on graphic processing units (GPGPU) such as the upcoming 512-bit
    139          Larrabee processor may also be useful for XML applications using parallel bit streams. New
    140          instruction set architectures may also offer dramatic improvements in core algorithms.
    141          Using the relatively simple extensions to support the principle of inductive doubling, a 3X
    142          improvement in several core parallel bit stream algorithms may be achieved [<xref
    143             linkend="ASPLOS09"/>]. Other possibilities include direct implementation of parallel
    144          extract and parallel deposit (pex/pdep) instructions [<xref linkend="Pex06"/>], and
    145          bit-level interleave operations as in Larrabee, each of which would have important
    146          application to parallel bit stream processing.</para>
    148       <para>Further prospects for XML performance improvement arise from leveraging the
    149          intraregister parallelism of parallel bit stream technology to exploit the interchip
    150          parallelism of multicore computing. Parallel bit stream techniques can support multicore
    151          parallelism in both data partitioning and task partitioning models. For example, the
    152          datasection partitioning approach of Wu, Zhang, Yu and Li may be used to partition blocks
    153          for speculative parallel parsing on separate cores followed by a postprocessing step to
    154          join partial S-trees [<xref linkend="Wu08"/>].</para>
    156       <para>In our view, the established and expected performance advantages of parallel bit stream
    157          technology over traditional byte-at-a-time processing are so compelling that parallel bit
    158          stream technology should ultimately form the foundation of every high-performance XML
    159          software stack. We envision a common high-performance XML kernel that may be customized to
    160          a variety of processor architectures and that supports a wide range of existing and new XML
    161          APIs. Widespread deployment of this technology should greatly benefit the XML community in
    162          addressing both the deserved and undeserved criticism of XML on performance grounds. A
    163          further benefit of improved performance is a substantial greening of XML technologies.</para>
    165       <para>To complement our research program investigating fundamental algorithms and issues in
    166          high-performance XML processing, our work also involves development of open source software
    167          implementing these algorithms, with a goal of full conformance to relevant specifications.
    168          From the research perspective, this approach is valuable in ensuring that the full
    169          complexity of required XML processing is addressed in reporting and assessing processing
    170          results. However, our goal is also to use this open source software as a basis of
    171          technology transfer. A Simon Fraser University spin-off company, called International
    172          Characters, Inc., has been created to commercialize the results of this work using a
    173          patent-based open source model.</para>
    175       <para>To date, we have not yet been successful in establishing a broader community of
    176          participation with our open source code base. Within open-source communities, there is
    177          often a general antipathy towards software patents; this may limit engagement with our
    178          technology, even though it has been dedicated for free use in open source. </para>
    180       <para>A further complication is the inherent difficulty of SIMD programming in general, and
    181          parallel bit stream programming in particular. Considerable work is required with each new
    182          algorithmic technique being investigated as well as in retargetting our techniques for each
    183          new development in SIMD and multicore processor technologies. To address these concerns, we
    184          have increasingly shifted the emphasis of our research program towards compiler technology
    185          capable of generating parallel bit stream code from higher-level specifications.</para>
     133      <para></para>
     134      <para></para>
     135      <para></para>
     136      <para></para>
    186137   </section>
    188139   <section>
    189       <title>A Catalog of Parallel Bit Streams for XML</title>
     140      <title>Background</title>
    190141      <section>
    191          <title>Introduction</title>
    192          <para>In this section, we introduce the fundamental concepts of parallel bit stream
    193             technology and present a comprehensive catalog of parallel bit streams for use in XML
    194             processing. In presenting this catalog, the focus is on the specification of the bit
    195             streams as data streams in one-to-one correspondence with the character code units of an
    196             input XML stream. The goal is to define these bit streams in the abstract without
    197             initially considering memory layouts, register widths or other issues related to
    198             particular target architectures. In cataloging these techniques, we also hope to convey
    199             a sense of the breadth of applications of parallel bit stream technology to XML
    200             processing tasks. </para>
    201       </section>
    203       <section>
    204          <title>Basis Bit Streams</title>
    205          <para>Given a byte-oriented text stream represented in UTF-8, for example, we define a
    206             transform representation of this text consisting of a set of eight parallel bit streams
    207             for the individual bits of each byte. Thus, the <code>Bit0</code> stream is the stream
    208             of bits consisting of bit 0 of each byte in the input byte stream, <code>Bit1</code> is
    209             the bit stream consisting of bit 1 of each byte in the input stream and so on. The set
    210             of streams <code>Bit0</code> through <code>Bit7</code> are known as the <emphasis>basis
    211                streams</emphasis> of the parallel bit stream representation. The following table
    212             shows an example XML character stream together with its representation as a set of 8
    213             basis streams. <table>
    214                <caption>
    215                   <para>XML Character Stream Transposition.</para>
    216                </caption>
    217                <colgroup>
    218                   <col align="left" valign="top"/>
    219                   <col align="left" valign="top"/>
    220                   <col align="left" valign="top"/>
    221                   <col align="left" valign="top"/>
    222                   <col align="left" valign="top"/>
    223                   <col align="left" valign="top"/>
    224                </colgroup>
    225                <tbody>
    226                   <tr valign="top">
    227                      <td>Input Data</td>
    228                      <td>
    229                         <code>&lt;</code>
    230                      </td>
    231                      <td>
    232                         <code>t</code>
    233                      </td>
    234                      <td>
    235                         <code>a</code>
    236                      </td>
    237                      <td>
    238                         <code>g</code>
    239                      </td>
    240                      <td>
    241                         <code>/</code>
    242                      </td>
    243                      <td>
    244                         <code>&gt;</code>
    245                      </td>
    246                   </tr>
    247                   <tr valign="top">
    248                      <td>ASCII</td>
    249                      <td>
    250                         <code>00111100</code>
    251                      </td>
    252                      <td>
    253                         <code>01110100</code>
    254                      </td>
    255                      <td>
    256                         <code>01100001</code>
    257                      </td>
    258                      <td>
    259                         <code>01100111</code>
    260                      </td>
    261                      <td>
    262                         <code>00101111</code>
    263                      </td>
    264                      <td>
    265                         <code>00111110</code>
    266                      </td>
    267                   </tr>
    268                   <tr valign="top">
    269                      <td>Bit0</td>
    270                      <td>
    271                         <code>0</code>
    272                      </td>
    273                      <td>
    274                         <code>0</code>
    275                      </td>
    276                      <td>
    277                         <code>0</code>
    278                      </td>
    279                      <td>
    280                         <code>0</code>
    281                      </td>
    282                      <td>
    283                         <code>0</code>
    284                      </td>
    285                      <td>
    286                         <code>0</code>
    287                      </td>
    288                   </tr>
    289                   <tr valign="top">
    290                      <td>Bit1</td>
    291                      <td>
    292                         <code>0</code>
    293                      </td>
    294                      <td>
    295                         <code>1</code>
    296                      </td>
    297                      <td>
    298                         <code>1</code>
    299                      </td>
    300                      <td>
    301                         <code>1</code>
    302                      </td>
    303                      <td>
    304                         <code>0</code>
    305                      </td>
    306                      <td>
    307                         <code>0</code>
    308                      </td>
    309                   </tr>
    310                   <tr valign="top">
    311                      <td>Bit2</td>
    312                      <td>
    313                         <code>1</code>
    314                      </td>
    315                      <td>
    316                         <code>1</code>
    317                      </td>
    318                      <td>
    319                         <code>1</code>
    320                      </td>
    321                      <td>
    322                         <code>1</code>
    323                      </td>
    324                      <td>
    325                         <code>1</code>
    326                      </td>
    327                      <td>
    328                         <code>1</code>
    329                      </td>
    330                   </tr>
    331                   <tr valign="top">
    332                      <td>Bit3</td>
    333                      <td>
    334                         <code>1</code>
    335                      </td>
    336                      <td>
    337                         <code>1</code>
    338                      </td>
    339                      <td>
    340                         <code>0</code>
    341                      </td>
    342                      <td>
    343                         <code>0</code>
    344                      </td>
    345                      <td>
    346                         <code>0</code>
    347                      </td>
    348                      <td>
    349                         <code>1</code>
    350                      </td>
    351                   </tr>
    352                   <tr valign="top">
    353                      <td>Bit4</td>
    354                      <td>
    355                         <code>1</code>
    356                      </td>
    357                      <td>
    358                         <code>0</code>
    359                      </td>
    360                      <td>
    361                         <code>0</code>
    362                      </td>
    363                      <td>
    364                         <code>0</code>
    365                      </td>
    366                      <td>
    367                         <code>1</code>
    368                      </td>
    369                      <td>
    370                         <code>1</code>
    371                      </td>
    372                   </tr>
    373                   <tr valign="top">
    374                      <td>Bit5</td>
    375                      <td>
    376                         <code>1</code>
    377                      </td>
    378                      <td>
    379                         <code>1</code>
    380                      </td>
    381                      <td>
    382                         <code>0</code>
    383                      </td>
    384                      <td>
    385                         <code>1</code>
    386                      </td>
    387                      <td>
    388                         <code>1</code>
    389                      </td>
    390                      <td>
    391                         <code>1</code>
    392                      </td>
    393                   </tr>
    394                   <tr valign="top">
    395                      <td>Bit6</td>
    396                      <td>
    397                         <code>0</code>
    398                      </td>
    399                      <td>
    400                         <code>0</code>
    401                      </td>
    402                      <td>
    403                         <code>0</code>
    404                      </td>
    405                      <td>
    406                         <code>1</code>
    407                      </td>
    408                      <td>
    409                         <code>1</code>
    410                      </td>
    411                      <td>
    412                         <code>1</code>
    413                      </td>
    414                   </tr>
    415                   <tr valign="top">
    416                      <td>Bit7</td>
    417                      <td>
    418                         <code>0</code>
    419                      </td>
    420                      <td>
    421                         <code>0</code>
    422                      </td>
    423                      <td>
    424                         <code>1</code>
    425                      </td>
    426                      <td>
    427                         <code>1</code>
    428                      </td>
    429                      <td>
    430                         <code>1</code>
    431                      </td>
    432                      <td>
    433                         <code>0</code>
    434                      </td>
    435                   </tr>
    436                </tbody>
    437             </table>
    438          </para>
    439          <para> Depending on the features of a particular processor architecture, there are a number
    440             of algorithms for transposition to parallel bit stream form. Several of these algorithms
    441             employ a three-stage structure. In the first stage, the input byte stream is divided
    442             into a pair of half-length streams consisting of four bits for each byte, for example,
    443             one stream for the high nybble of each byte and another for the low nybble of each byte.
    444             In the second stage, these streams of four bits per byte are each divided into streams
    445             consisting of two bits per original byte, for example streams for the
    446             <code>Bit0/Bit1</code>, <code>Bit2/Bit3</code>, <code>Bit4/Bit5</code>, and
    447                <code>Bit6/Bit7</code> pairs. In the final stage, the streams are further subdivided
    448             in the individual bit streams. </para>
    449          <para> Using SIMD capabilities, this process is quite efficient, with an amortized cost of
    450             1.1 CPU cycles per input byte on Intel Core 2 with SSE, or 0.6 CPU cycles per input byte
    451             on Power PC G4 with Altivec. With future advances in processor technology, this
    452             transposition overhead is expected to reduce, possibly taking advantage of upcoming
    453             parallel extract (pex) instructions on Intel technology. In the ideal, only 24
    454             instructions are needed to transform a block of 128 input bytes using 128-bit SSE
    455             registers using the inductive doubling instruction set architecture, representing an
    456             overhead of less than 0.2 instructions per input byte. </para>
    457       </section>
    459       <section>
    460          <title>General Streams</title>
    461          <para>This section describes bit streams which support basic processing operations.</para>
    463          <section>
    464             <title>Deletion Mask Streams</title>
    465             <para>DelMask (deletion mask) streams marks character code unit positions for deletion.
    466                Since the deletion operation is dependency free across many stages of XML processing,
    467                it is possible to simply mark and record deletion positions as deletion mask streams for future processing. A single
    468                invocation of a SIMD based parallel deletion algorithm can then perform the deletion of
    469                positions accumulated across a number of stages through a bitwise ORing of deletion
    470                masks. For example, deletion arises in the replacement of predefined entities with a
    471                single character, such as in the replacement of the &amp;amp; entity, with the
    472                &amp; character. Deletion also arises in XML
    473                end-of-line handling, and CDATA section delimeter processing. Several algorithms to
    474                delete bits at positions marked by DelMask are possible [<xref linkend="u8u16"/>]. </para>
    475             <para>The following table provides an example of generating a DelMask in the context of
    476                bit stream based parsing of well-formed character references and predefined entities.
    477                The result is the generation of a DelMask stream. <table>
    478                   <caption>
    479                      <para>DelMask Stream Generation</para>
    480                   </caption>
    481                   <colgroup>
    482                      <col align="left" valign="top"/>
    483                      <col align="left" valign="top"/>
    484                   </colgroup>
    485                   <tbody>
    486                      <tr valign="top">
    487                         <td>Input Data</td>
    488                         <td>
    489                            <code>&amp;gt; &amp;#13; &amp;#x0a;</code>
    490                         </td>
    491                      </tr>
    492                      <tr valign="top">
    493                         <td>GenRefs</td>
    494                         <td>
    495                            <code>_11______________</code>
    496                         </td>
    497                      </tr>
    499                      <tr valign="top">
    500                         <td>DecRefs</td>
    501                         <td>
    502                            <code>_______11________</code>
    503                         </td>
    504                      </tr>
    505                      <tr valign="top">
    506                         <td>HexRefs</td>
    507                         <td>
    508                            <code>______________11_</code>
    509                         </td>
    510                      </tr>
    511                      <tr valign="top">
    512                         <td>DelMask</td>
    513                         <td>
    514                            <code>111__1111__11111_</code>
    515                         </td>
    516                      </tr>
    517                      <tr valign="top">
    518                         <td>ErrorFlag</td>
    519                         <td>
    520                            <code>_________________</code>
    521                         </td>
    522                      </tr>
    523                   </tbody>
    524                </table>
    525             </para>
    526          </section>
    528          <section>
    529             <title>Error Flag Streams </title>
    530             <para>Error flag streams indicates the character code unit positions of syntactical
    531                errors. XML processing examples which benefit from the marking of error positions
    532                include UTF-8 character sequence validation and XML parsing [<xref linkend="u8u16"
    533                />].</para>
    534             <para>The following table provides an example of using bit streams to parse character
    535                references and predefined entities which fail to meet the XML 1.0 well-formedness
    536                constraints. The result is the generation of an error flag stream that marks the
    537                positions of mal-formed decimal and hexical character references respectively. <table>
    538                   <caption>
    539                      <para>Error Flag Stream Generation</para>
    540                   </caption>
    541                   <colgroup>
    542                      <col align="left" valign="top"/>
    543                      <col align="left" valign="top"/>
    544                   </colgroup>
    545                   <tbody>
    546                      <tr valign="top">
    547                         <td>Input Data</td>
    548                         <td>
    549                            <code>&amp;gt; &amp;#, &amp;#x; </code>
    550                         </td>
    551                      </tr>
    552                      <tr valign="top">
    553                         <td>GenRefs</td>
    554                         <td>
    555                            <code>_11___________</code>
    556                         </td>
    557                      </tr>
    558                      <tr valign="top">
    559                         <td>DecRefs</td>
    560                         <td>
    561                            <code>______________</code>
    562                         </td>
    563                      </tr>
    564                      <tr valign="top">
    565                         <td>HexRefs</td>
    566                         <td>
    567                            <code>______________</code>
    568                         </td>
    569                      </tr>
    570                      <tr valign="top">
    571                         <td>DelMask</td>
    572                         <td>
    573                            <code>111__11__111__</code>
    574                         </td>
    575                      </tr>
    576                      <tr valign="top">
    577                         <td>ErrorFlag</td>
    578                         <td>
    579                            <code>_______1____1_</code>
    580                         </td>
    581                      </tr>
    582                   </tbody>
    583                </table>
    584             </para>
    585          </section>
    587       </section>
    589       <section>
    590          <title>Lexical Item Streams</title>
    591          <para>Lexical item streams differ from traditional streams of tokens in that they are bit
    592             streams that mark the positions of tokens, whitespace or delimiters. Additional bit
    593             streams, such as the reference streams and callout streams, are subsequently constructed
    594             based on the information held within the set of lexical items streams. Differentiation
    595             between the actual tokens that may occur at a particular point (e.g., the different XML
    596             tokens that begin “&lt;”) may be performed using multicharacter recognizers on the
    597             bytestream representation [<xref linkend="CASCON08"/>].</para>
    598          <para>A key role of lexical item streams in XML parsing is to facilitate fast scanning
    599             operations. For example, a left angle bracket lexical item stream may be formed to
    600             identify those character code unit positions at which a “&lt;” character occurs.
    601             Hardware register bit scan operations may then be used by the XML parser on the left
    602             angle bracket stream to efficiently identify the position of the next “&lt;”. Based
    603             on the capabilities of current commodity processors, a single register bit scan
    604             operation may effectively scan up to 64 byte positions with a single instruction.</para>
    605          <para>Overall, the construction of the full set of lexical item stream computations
    606             requires approximately 1.0 CPU cycles per byte when implemented for 128 positions at a
    607             time using 128-bit SSE registers on Intel Core2 processors [<xref linkend="CASCON08"/>].
    608             The following table defines the core lexical item streams defined by the Parabix XML
    609             parser.</para>
    610          <para>
    611             <table>
    612                <caption>
    613                   <para>Lexical item stream descriptions.</para>
    614                </caption>
    615                <tbody>
    616                   <tr>
    617                      <td align="left"> LAngle </td>
    618                      <td align="left"> Marks the position of any left angle bracket character.</td>
    619                   </tr>
    620                   <tr>
    621                      <td align="left"> RAngle </td>
    622                      <td align="left"> Marks the position of any right angle bracket character.</td>
    623                   </tr>
    624                   <tr>
    625                      <td align="left"> LBracket </td>
    626                      <td align="left"> Marks the position of any left square bracker character.</td>
    627                   </tr>
    628                   <tr>
    629                      <td align="left"> RBracket </td>
    630                      <td align="left"> Marks the position of any right square bracket
    631                      character.</td>
    632                   </tr>
    633                   <tr>
    634                      <td align="left"> Exclam </td>
    635                      <td align="left"> Marks the position of any exclamation mark character.</td>
    636                   </tr>
    637                   <tr>
    638                      <td align="left"> QMark </td>
    639                      <td align="left"> Marks the position of any question mark character.</td>
    640                   </tr>
    641                   <tr>
    642                      <td align="left"> Hyphen </td>
    643                      <td align="left"> Marks the position of any hyphen character.</td>
    644                   </tr>
    645                   <tr>
    646                      <td align="left"> Equals </td>
    647                      <td align="left"> Marks the position of any equal sign character.</td>
    648                   </tr>
    649                   <tr>
    650                      <td align="left"> SQuote </td>
    651                      <td align="left"> Marks the position of any single quote character.</td>
    652                   </tr>
    653                   <tr>
    654                      <td align="left"> DQuote </td>
    655                      <td align="left"> Marks the position of any double quote character.</td>
    656                   </tr>
    657                   <tr>
    658                      <td align="left"> Slash </td>
    659                      <td align="left"> Marks the position of any forward slash character</td>
    660                   </tr>
    661                   <tr>
    662                      <td align="left"> NameScan </td>
    663                      <td align="left"> Marks the position of any XML name character.</td>
    664                   </tr>
    665                   <tr>
    666                      <td align="left"> WS </td>
    667                      <td align="left"> Marks the position of any XML 1.0 whitespace character.</td>
    668                   </tr>
    669                   <tr>
    670                      <td align="left"> PI_start </td>
    671                      <td align="left"> Marks the position of the start of any processing instruction
    672                         at the '?' character position.</td>
    673                   </tr>
    674                   <tr>
    675                      <td align="left"> PI_end </td>
    676                      <td align="left"> Marks the position of any end of any processing instruction
    677                         at the '>' character position.</td>
    678                   </tr>
    679                   <tr>
    680                      <td align="left"> CtCD_start </td>
    681                      <td align="left"> Marks the position of the start of any comment or CDATA
    682                         section at the '!' character position.</td>
    683                   </tr>
    684                   <tr>
    685                      <td align="left"> EndTag_start </td>
    686                      <td align="left"> Marks the position of any end tag at the '/' character
    687                         position.</td>
    688                   </tr>
    689                   <tr>
    690                      <td align="left"> CD_end </td>
    691                      <td align="left"> Marks the position of the end of any CDATA section at the '>'
    692                         character position. </td>
    693                   </tr>
    694                   <tr>
    695                      <td align="left"> DoubleHyphen </td>
    696                      <td align="left"> Marks the position of any double hyphen character.</td>
    697                   </tr>
    698                   <tr>
    699                      <td align="left"> RefStart </td>
    700                      <td align="left"> Marks the position of any ampersand character.</td>
    701                   </tr>
    702                   <tr>
    703                      <td align="left"> Hash </td>
    704                      <td align="left"> Marks the position of any hash character.</td>
    705                   </tr>
    706                   <tr>
    707                      <td align="left"> x </td>
    708                      <td align="left"> Marks the position of any 'x' character.</td>
    709                   </tr>
    710                   <tr>
    711                      <td align="left"> Digit </td>
    712                      <td align="left"> Marks the position of any digit.</td>
    713                   </tr>
    714                   <tr>
    715                      <td align="left"> Hex </td>
    716                      <td align="left"> Marks the position of any hexidecimal character.</td>
    717                   </tr>
    718                   <tr>
    719                      <td align="left"> Semicolon </td>
    720                      <td align="left"> Marks the position of any semicolon character.</td>
    721                   </tr>
    722                </tbody>
    723             </table>
    724          </para>
    725          <para> The following illustrates a number of the lexical item streams. </para>
    726          <para>
    727             <table>
    728                <caption>
    729                   <para>Lexical Item Streams</para>
    730                </caption>
    731                <colgroup>
    732                   <col align="left" valign="top"/>
    733                   <col align="left" valign="top"/>
    734                </colgroup>
    735                <tbody>
    736                   <tr valign="top">
    737                      <td>Input Data</td>
    738                      <td>
    739                         <code>&lt;tag&gt;&lt;tag&gt; text &amp;lt;
    740                            &amp;#x3e; &lt;/tag&gt;&lt;/tag&gt;</code>
    741                      </td>
    742                   </tr>
    744                   <tr valign="top">
    745                      <td>LAngle</td>
    746                      <td>
    747                         <code>1____1______________________1_____1_____</code>
    748                      </td>
    749                   </tr>
    750                   <tr valign="top">
    751                      <td>RAngle</td>
    752                      <td>
    753                         <code>____1____1_______________________1_____1</code>
    754                      </td>
    755                   </tr>
    756                   <tr valign="top">
    757                      <td>WS</td>
    758                      <td>
    759                         <code>__________1____1____1______1____________</code>
    760                      </td>
    761                   </tr>
    762                   <tr valign="top">
    763                      <td>RefStart</td>
    764                      <td>
    765                         <code>________________1____1__________________</code>
    766                      </td>
    767                   </tr>
    768                   <tr valign="top">
    769                      <td>Hex</td>
    770                      <td>
    771                         <code>__1____1____1___________11_____1_____1__</code>
    772                      </td>
    773                   </tr>
    775                   <tr valign="top">
    776                      <td>Semicolon</td>
    777                      <td>
    778                         <code>___________________1______1_____________</code>
    779                      </td>
    780                   </tr>
    781                   <tr valign="top">
    782                      <td>Slash</td>
    783                      <td>
    784                         <code>_____________________________1_____1____</code>
    785                      </td>
    786                   </tr>
    787                </tbody>
    788             </table>
    789          </para>
    791       </section>
    793       <section>
    794          <title>UTF-8 Byte Classification, Scope and Validation Streams</title>
    795          <para> An XML parser must accept the UTF-8 encoding of Unicode [<xref linkend="XML10"/>].
    796             It is a fatal error if an XML document determined to be in UTF-8 contains byte sequences
    797             that are not legal in that encoding. UTF-8 byte classification, scope, XML character
    798             validation and error flag bit streams are defined to validate UTF-8 byte sequences and
    799             support transcoding to UTF-16.</para>
    801          <section>
    802             <title>UTF-8 Byte Classification Streams</title>
    803             <para>UTF-8 byte classification bit streams classify UTF-8 bytes based on their role in
    804                forming single and multibyte sequences. The u8Prefix and u8Suffix bit streams
    805                identify bytes that represent, respectively, prefix or suffix bytes of multibyte
    806                UTF-8 sequences. The u8UniByte bit stream identifies those bytes that may be
    807                considered single-byte sequences. The u8Prefix2, u8Prefix3, and u8Prefix4 refine the
    808                u8Prefix respectively indicating prefixes of two, three or four byte
    809             sequences respectively.</para>
    810          </section>
    812          <section>
    813             <title>UTF-8 Scope Streams</title>
    814             <para> Scope streams represent expectations established by UTF-8 prefix bytes. For
    815                example, the u8Scope22 bit stream represents the positions at which the second byte of a
    816                two-byte sequence is expected based on the occurrence of a two-byte prefix in the
    817                immediately preceding position. The u8scope32, u8Scope33, u8Scope42, u8scope43, and
    818                u8Scope44 complete the set of UTF-8 scope streams.</para>
    819             <para> The following example demonstrates the UTF-8 character encoding validation
    820                process using parallel bit stream techniques. The result of this validation process
    821                is an error flag stream identifying the positions at which errors occur.</para>
    822             <para>
    823                <table>
    824                   <caption>
    825                      <para>UTF-8 Scope Streams</para>
    826                   </caption>
    827                   <colgroup>
    828                      <col align="left" valign="top"/>
    829                      <col align="left" valign="top"/>
    830                   </colgroup>
    831                   <tbody> <tr valign="top"><td>Input Data</td><td><code>A Text in Farsi:&#160;ى&#160;ك&#160;&#160;Ù
    832 &#160;ت&#160;ن&#160;&#160;ف&#160;ا&#160;ر&#160;س&#160;ى</code></td></tr>
    834                      <tr valign="top">
    835                         <td>High Nybbles</td>
    836                         <td>
    837                            <code>42567726624677632D8DBDBDAD82D8DAD82D8D8</code>
    838                         </td>
    839                      </tr>
    840                      <tr valign="top">
    841                         <td>Low Nybbles</td>
    843                         <td>
    844                            <code>10458409E061239A099838187910968A9509399</code>
    845                         </td>
    846                      </tr>
    847                      <tr valign="top">
    848                         <td>u8Unibyte</td>
    849                         <td>
    850                            <code>11111111111111111__________1______1____</code>
    852                         </td>
    853                      </tr>
    854                      <tr valign="top">
    855                         <td>u8Prefix</td>
    856                         <td>
    857                            <code>_________________1_1_1_1_1__1_1_1__1_1_</code>
    858                         </td>
    859                      </tr>
    861                      <tr valign="top">
    862                         <td>u8Suffix</td>
    863                         <td>
    864                            <code>__________________1_1_1_1_1__1_1_1__1_1</code>
    865                         </td>
    866                      </tr>
    867                      <tr valign="top">
    868                         <td>u8Prefix2</td>
    870                         <td>
    871                            <code>_________________1_1_1_1_1__1_1_1__1_1_</code>
    872                         </td>
    873                      </tr>
    874                      <tr valign="top">
    875                         <td>u8Scope22</td>
    876                         <td>
    877                            <code>__________________1_1_1_1_1__1_1_1__1_1</code>
    879                         </td>
    880                      </tr>
    881                      <tr valign="top">
    882                         <td>ErrorFlag</td>
    883                         <td>
    884                            <code>_______________________________________</code>
    885                         </td>
    886                      </tr>
    888                   </tbody>
    889                </table>
    892             </para>
    893          </section>
    895          <section>
    896             <title>UTF-8 Validation Streams</title>
    897             <para> Proper formation of UTF-8 byte sequences requires that the correct number of
    898                suffix bytes always follow a UTF-8 prefix byte, and that certain illegal byte
    899                combinations are ruled out. For example, sequences beginning with the prefix bytes
    900                0xF5 through 0xFF are illegal as they would represent code point values above 10FFFF.
    901                In addition, there are constraints on the first suffix byte following certain special
    902                prefixes, namely that a suffix following the prefix 0xE0 must fall in the range
    903                0xA0–0xBF, a suffix following the prefix 0xED must fall in the range 0x80–0x9F, a
    904                suffix following the prefix 0xF0 must fall in the range 0x90–0xBF and a suffix
    905                following the prefix 0xF4 must fall in the range 0x80–0x8F. The task of ensuring that
    906                each of these constraints hold is known as UTF-8 validation. The bit streams xE0,
    907                xED, xF0, xF4, xA0_xBF, x80_x9F, x90_xBF, and x80_x8F are constructed to flag the
    908                aforementioned UTF-8 validation errors. The result of UTF-8 validation is a UTF-8
    909                error flag bit stream contructed as the ORing of a series of UTF-8 validation tests.
    910             </para>
    911          </section>
    913          <section>
    914             <title>XML Character Validation Streams</title>
    915             <para>The UTF-8 character sequences <emphasis>0xEF 0xBF 0xBF</emphasis> and
    916                   <emphasis>0xEF 0xBF 0xBE</emphasis> correspond to the Unicode code points 0xFFFE
    917                and 0xFFFF respectively. In XML 1.0, 0xFFFE and 0xFFFF represent characters outside
    918                the legal XML character ranges. As such, bit streams which mark 0xEF, 0xBF, and 0xBE
    919                character are constructed to flag illegal UTF-8 character sequences. </para>
    920          </section>
    922          <section>
    923             <title>UTF-8 to UTF-16 Transcoding</title>
    924             <para>UTF-8 is often preferred for storage and data exchange, it is suitable for
    925                processing, but it is significantly more complex to process than UTF-16 [<xref
    926                   linkend="Unicode"/>]. As such, XML documents are typically encoded in UTF-8 for
    927                serialization and transport, and subsequently transcoded to UTF-16 for processing
    928                with programming languages such as Java and C#. Following the parallel bit stream
    929                methods developed for the u8u16 transcoder, a high-performance standalone UTF-8 to
    930                UTF-16 transcoder [<xref linkend="u8u16"/>], transcoding to UTF-16 may be achieved by
    931                computing a series of 16 bit streams. One stream for each of the individual bits of a
    932                UTF-16 code unit. </para>
    933             <para>The bit streams for UTF-16 are conveniently divided into groups: the eight streams
    934                u16Hi0, u16Hi1, ..., u16Hi7 for the high byte of each UTF-16 code unit and the eight
    935                streams u16Lo1, ..., u16Lo7 for the low byte. Upon conversion of the parallel bit
    936                stream data back to byte streams, eight sequential byte streams U16h0, U16h1, ...,
    937                U16Hi7 are used for the high byte of each UTF-16 code unit, while U16Lo0, U16Lo1,...,
    938                U16Lo7 are used for the corresponding low byte. Interleaving these streams then
    939                produces the full UTF-16 doublebyte stream.</para>
    940          </section>
    942          <section>
    943             <title>UTF-8 Indexed UTF-16 Streams</title>
    944             <para>UTF-16 bit streams are initially defined in UTF-8 indexed form. That is, with sets
    945                of bits in one-to-one correspondence with UTF-8 bytes. However, only one set of
    946                UTF-16 bits is required for encoding two or three-byte UTF-8 sequences and only two
    947                sets are required for surrogate pairs corresponding to four-byte UTF-8 sequences. The
    948                u8LastByte (u8UniByte , u8Scope22 , u8Scope33 , and u8Scope44 ) and u8Scope42 streams
    949                mark the positions at which the correct UTF-16 bits are computed. The bit sets at
    950                other positions must be deleted to compress the streams to the UTF-16 indexed form.
    951             </para>
    952          </section>
    953       </section>
    955       <section>
    956          <title>Control Character Streams</title>
    957          <para>The control character bit streams marks ASCII control characters in the range
    958             0x00-0x1F. Additional control character bit streams mark the tab, carriage return, line
    959             feed, and space character. In addition, a bit stream to mark carriage return line
    960             combinations is also constructed. Presently, control character bit streams support the
    961             operations of XML 1.0 character validation and XML end-of-line handling.</para>
    963          <section>
    964             <title>XML Character Validation</title>
    965             <para>Legal characters in XML are the tab, carriage return, and line feed characters,
    966                together with all Unicode characters and excluding the surrogate blocks, as well as hexadecimal OxFFFE and
    967                OxFFFF [<xref linkend="XML10"/>]. The x00_x1F bit stream is constructed and used in
    968                combination with the additional control character bit streams to flags the positions
    969                of illegal control characters.</para>
    970          </section>
    972          <section>
    973             <title>XML 1.0 End-of-line Handling</title>
    974             <para>In XML 1.0 the two-character sequence CR LF (carriage return, line feed) as well as
    975                any CR character not followed by a LF character must be converted to a single LF
    976                character [<xref linkend="XML10"/>].</para>
    977             <para>By defining carriage return, line feed, and carriage return line feed bit streams,
    978                dentoted CR, LF and CRLF respectively, end-of-line normalization processing can be
    979                performed in parallel using only a small number of logical and shift operations.</para>
    980             <para/>
    981             <para>The following example demonstrates the generation of the CRLF deletion mask. In
    982                this example, the position of all CR characters followed by LF characters are marked
    983                for deletion. Isolated carriage returns are then replaced with LF characters.
    984                Completion of this process satisfies the XML 1.0 end-of-line handling requirements.
    985                For clarity, this example encodes input data carriage returns as
    986                <emphasis>C</emphasis> characters, whereas line feed characters are shown as
    987                   <emphasis>L</emphasis> characters.</para>
    988             <para>
    989                <table>
    990                   <caption>
    991                      <para>XML 1.0 End-of-line Handling</para>
    992                   </caption>
    993                   <colgroup>
    994                      <col align="left" valign="top"/>
    995                   </colgroup>
    996                   <tbody>
    997                      <tr valign="top">
    998                         <td>Input Data</td>
    999                         <td>
    1000                            <code>first line C second line CL third line L one more C nothing
    1001                            left</code>
    1002                         </td>
    1003                      </tr>
    1004                      <tr valign="top">
    1005                         <td>CR</td>
    1006                         <td>
    1007                            <code>-----------1-------------1------------------------1-------------</code>
    1008                         </td>
    1009                      </tr>
    1010                      <tr valign="top">
    1011                         <td>LF</td>
    1012                         <td>
    1013                            <code>--------------------------1------------1------------------------</code>
    1014                         </td>
    1015                      </tr>
    1016                      <tr valign="top">
    1017                         <td>DelMask</td>
    1018                         <td>
    1019                            <code>--------------------------1-------------------------------------</code>
    1020                         </td>
    1021                      </tr>
    1022                   </tbody>
    1023                </table>
    1025             </para>
    1026          </section>
    1028       </section>
    1030       <section>
    1031          <title>Call Out Streams</title>
    1032          <para> Call out bit streams mark the extents of XML markup structures such as comments,
    1033             processing instruction and CDATA sections as well as physical structures such as character and
    1034             entity references and general references.  Call out streams are also formed for logical markup structures such
    1035             start tags, end tags and empty element tags. </para>
    1036          <section>
    1037             <title>Comment, Processing Instruction and CDATA Section Call Out Streams</title>
    1038             <para>Comments, processing instructions and CDATA sections call out streams, Ct_Span,
    1039                PI_Span and CD_Span respectively, define sections of an XML document which
    1040                contain markup that is not interpreted by an XML processor. As such, the union of
    1041                Ct_Span, PI_Span and CD_Span streams defines the regions of non-interpreteable markup.
    1042                The stream formed by this union is termed the CtCDPI_Mask.</para>
    1043             <para>The following tables provides an example of constructing the CtCDPI_Mask. </para>
    1044             <table>
    1045                <caption>
    1046                   <para>CtCDPI Mask Generation</para>
    1047                </caption>
    1048                <colgroup><col align="left" valign="top" /><col align="left" valign="top" /></colgroup>
    1049                <tbody> <tr valign="top"><td>Input Data</td><td><code>&lt;?php?&gt; &lt;!-- example --&gt; &lt;![CDATA[ shift: a&lt;&lt;1 ]]&gt;</code></td></tr>
    1051                   <tr valign="top"><td>CD_Span</td><td><code>___________________________1111111111111111111111_</code></td></tr>
    1052                   <tr valign="top"><td>Ct_Span</td><td><code>___________111111111111___________________________</code></td></tr>
    1053                   <tr valign="top"><td>PI_Span</td><td><code>_11111____________________________________________</code></td></tr>
    1054                   <tr valign="top"><td>CtCDPI_Mask</td><td><code>_111111__111111111111111__111111111111111111111111</code></td></tr>
    1055                   <tr valign="top"><td>ErrorFlag</td><td><code>__________________________________________________</code></td></tr>
    1057                </tbody>
    1058             </table>
    1063             <para> With the removal of all non-interpreteable markup, several phases of parallel bit
    1064                stream based SIMD operations may follow operating on up to 128 byte positions on
    1065                current commondity processors and assured of XML markup relevancy. For
    1066                example, with the extents identification of comments, processing instructions and
    1067                CDATA secions, XML names may be identified and length sorted for efficient symbol
    1068                table construction. </para>
    1069             <para> As an aside, comments and CDATA sections must first be validated to ensure
    1070                that comments do not contain "--" sequences and that CDATA sections do not contain illegal
    1071                "]]&gt;" sequences prior to ignorable markup stream generation.</para>
    1072          </section>
    1074          <section>
    1075             <title>Reference Call Out Streams</title>
    1076             <para>The reference call out streams are the GenRefs, DecRefs, and HexRefs streams. This
    1077                subset of the call out streams marks the extents of all but the closing semicolon of
    1078                general and character references.</para>
    1079             <para>Predefined character
    1080                (<![CDATA[&lt;,&gt;,&amp;,&apos;,&quot;]]>) and numeric character
    1081                references (&amp;#nnnn;, &amp;#xhhhh;) must be replaced by a single character
    1082                   [<xref linkend="XML10"/>]. As previously shown, this subset of call out streams enables the construction of a DelMask for
    1083                references.</para>
    1084          </section>
    1086          <section>
    1087             <title>Tag Call Out Streams</title>
    1088             <para>Whereas sequential bit scans over lexical item streams form the basis of XML
    1089                parsing, in the current Parabix parser a new method of parallel parsing has been
    1090                developed and prototyped using the concept of bitstream addition. Fundamental to this
    1091                method is the concept of a <emphasis>cursor</emphasis> stream, a bit stream marking
    1092                the positions of multiple parallel parses currently in process. </para>
    1093             <para>The results of parallel parsing using the bit stream addition technique produces a
    1094                set of tag call out bit streams. These streams mark the extents of each start tag,
    1095                end tag and empty element tag. Within tags, additional streams mark start
    1096                and end positions for tag names, as well as attribute names and values. An error flag
    1097                stream marks the positions of any syntactic errors encountered during parsing.</para>
    1098             <para> The set of tag call out streams consists of the ElemNames, AttNames, AttVals, Tags,
    1099                EmptyTagEnds and EndTags bit streams. The following example demonstrates the bit
    1100                stream output produced which from parallel parsing using bit stream addition. </para>
    1101             <table>
    1102                <caption>
    1103                   <para>Tag Call Out Streams</para>
    1104                </caption>
    1105                <colgroup>
    1106                   <col align="left" valign="top"/>
    1107                   <col align="left" valign="top"/>
    1108                </colgroup>
    1109                <tbody>
    1110                   <tr valign="top">
    1111                      <td>Input Data</td>
    1112                      <td>
    1113                         <code>&lt;root&gt;&lt;t1&gt;text&lt;/t1&gt;&lt;t2
    1114                            a1=&apos;foo&apos; a2 =
    1115                            &apos;fie&apos;&gt;more&lt;/t2&gt;&lt;tag3
    1116                            att3=&apos;b&apos;/&gt;&lt;/root&gt;</code>
    1117                      </td>
    1118                   </tr>
    1120                   <tr valign="top">
    1121                      <td>ElemNames</td>
    1122                      <td>
    1123                         <code>_1111__11___________11_______________________________1111__________________</code>
    1124                      </td>
    1125                   </tr>
    1126                   <tr valign="top">
    1127                      <td>AttNames</td>
    1128                      <td>
    1129                         <code>_______________________11_______11________________________1111_____________</code>
    1130                      </td>
    1131                   </tr>
    1132                   <tr valign="top">
    1133                      <td>AttrVals</td>
    1134                      <td>
    1135                         <code>__________________________11111______11111_____________________111_________</code>
    1136                      </td>
    1137                   </tr>
    1138                   <tr valign="top">
    1139                      <td>EmptyTagEnds</td>
    1140                      <td>
    1141                         <code>___________________________________________________________________1_______</code>
    1142                      </td>
    1143                   </tr>
    1144                   <tr valign="top">
    1145                      <td>EndTags</td>
    1146                      <td>
    1147                         <code>_______________111______________________________111__________________11111_</code>
    1148                      </td>
    1149                   </tr>
    1151                   <tr valign="top">
    1152                      <td>Start/EmptyTags</td>
    1153                      <td>
    1154                         <code>_1111__11___________1111111111111111111111___________11111111111111________</code>
    1155                      </td>
    1156                   </tr>
    1157                   <tr valign="top">
    1158                      <td>ErrorFlag</td>
    1159                      <td>
    1160                         <code>___________________________________________________________________________</code>
    1161                      </td>
    1162                   </tr>
    1163                </tbody>
    1164             </table>
    1166          </section>
    1167       </section>
    1168    </section>
    1169    <section>
    1170       <title>SIMD Beyond Bitstreams: Names and Numbers</title>
    1172       <para>Whereas the fundamental innovation of our work is the use of SIMD technology in
    1173          implementing parallel bit streams for XML, there are also important ways in which more
    1174          traditional byte-oriented SIMD operations can be useful in accelerating other aspects of
    1175          XML processing.</para>
    1177       <section>
    1178          <title>Name Lookup</title>
    1179          <para>Efficient symbol table mechanisms for looking up element and attribute names is
    1180             important for almost all XML processing applications. It is also an important technique
    1181             merely for assessing well-formedness of an XML document; rather than validating the
    1182             character-by-character composition of each occurrence of an XML name as it is
    1183             encountered, it is more efficient to validate all but the first occurrence by first
    1184             determining whether the name already exists in a table of prevalidated names.</para>
    1186          <para>The first symbol table mechanism deployed in the Parabix parser simply used the
    1187             hashmaps of the C++ standard template library, without deploying any SIMD technology.
    1188             However, with the overhead of character validation, transcoding and parsing dramatically
    1189             reduced by parallel bit stream technology, we found that symbol lookups then accounted
    1190             for about half of the remaining execution time in a statistics gathering application
    1191                [<xref linkend="CASCON08"/>]. Thus, symbol table processing was identified as a major
    1192             target for further performance improvement. </para>
    1193          <para> Our first effort to improve symbol table performance was to employ the splash tables
    1194             with cuckoo hashing as described by Ross [<xref linkend="Ross06"/>], using SIMD
    1195             technology for parallel bucket processing. Although this technique did turn out to have
    1196             the advantage of virtually constant-time performance even for very large vocabularies,
    1197             it was not particularly helpful for the relatively small vocabularies typically found in
    1198             XML document processing. </para>
    1199          <para> However, a second approach has been found to be quite useful, taking advantage of
    1200             parallel bit streams for cheap determination of symbol length. In essence, the length of
    1201             a name can be determined very cheaply using a single bit scan operation. This then makes
    1202             it possible to use length-sorted symbol table processing, as follows. First, the
    1203             occurrences of all names are stored in arrays indexed by length. Then the length-sorted
    1204             arrays may each be inserted into the symbol table in turn. The advantage of this is that
    1205             a separate loop may be written for each length. Length sorting makes for very efficient
    1206             name processing. For example hash value computations and name comparisons can be made by
    1207             loading multibyte values and performing appropriate shifting and masking operations,
    1208             without the need for a byte-at-a-time loop. In initial experiments, this length-sorting
    1209             approach was found to reduce symbol lookup cost by a factor of two. </para>
    1210          <para> Current research includes the application of SIMD technology to further enhance the
    1211             performance of length-sorted lookup. We have identified a promising technique for
    1212             parallel processing of multiple name occurrences using a parallel trie lookup technique.
    1213             Given an array of occurrences of names of a particular length, the first one, two or
    1214             four bytes of each name are gathered and stored in a linear array. SIMD techniques are
    1215             then used to compare these prefixes with the possible prefixes for the current position
    1216             within the trie. In general, a very small number of possibilities exist for each trie
    1217             node, allowing for fast linear search through all possibilities. Typically, the
    1218             parallelism is expected to exceed the number of possibilities to search through at each
    1219             node. With length-sorting to separate the top-level trie into many small subtries, we
    1220             expect only a single step of symbol lookup to be needed in most practical instances. </para>
    1222          <para>The gather step of this algorithm is actually a common technique in SIMD processing.
    1223             Instruction set support for gather operations is a likely future direction for SIMD
    1224             technology.</para>
    1225       </section>
    1227       <section>
    1228          <title>Numeric Processing</title>
    1229          <para> Many XML applications involve numeric data fields as attribute values or element
    1230             content. Although most current XML APIs uniformly return information to applications in
    1231             the form of character strings, it is reasonable to consider direct API support for
    1232             numeric conversions within a high-performance XML engine. With string to numeric
    1233             conversion such a common need, why leave it to application programmers? </para>
    1234          <para> High-performance string to numeric conversion using SIMD operations also can
    1235             considerably outperform the byte-at-a-time loops that most application programmers or
    1236             libraries might employ. A first step is reduction of ASCII bytes to corresponding
    1237             decimal nybbles using a SIMD packing operation. Then an inductive doubling algorithm
    1238             using SIMD operations may be employed. First, 16 sets of adjacent nybble values in the
    1239             range 0-9 can be combined in just a few SIMD operations to 16 byte values in the range
    1240             0-99. Then 8 sets of byte values may similarly be combined with further SIMD processing
    1241             to produce doublebyte values in the range 0-9999. Further combination of doublebyte
    1242             values into 32-bit integers and so on can also be performed using SIMD operations. </para>
    1243          <para> Using appropriate gather operations to bring numeric strings into appropriate array
    1244             structures, an XML engine could offer high-performance numeric conversion services to
    1245             XML application programmers. We expect this to be an important direction for our future
    1246             work, particularly in support of APIs that focus on direct conversion of XML data into
    1247             business objects. </para>
    1249       </section>
    1250    </section>
    1252    <section>
    1253       <title>APIs and Parallel Bit Streams</title>
    1255       <section>
    1256          <title>The ILAX Streaming API</title>
    1257          <para>The In-Line API for XML (ILAX) is the base API provided with the Parabix parser. It
    1258             is intended for low-level extensions compiled right into the engine, with minimum
    1259             possible overhead. It is similar to streaming event-based APIs such as SAX, but
    1260             implemented by inline substitution rather than using callbacks. In essence, an extension
    1261             programmer provides method bodies for event-processing methods declared internal to the
    1262             Parabix parsing engine, compiling the event processing code directly with the core code
    1263             of the engine. </para>
    1264          <para> Although ILAX can be used directly for application programming, its primary use is
    1265             for implementing engine extensions that support higher-level APIs. For example, the
    1266             implementation of C or C++ based streaming APIs based on the Expat [<xref
    1267                linkend="Expat"/>] or general SAX models can be quite directly implemented. C/C++ DOM
    1268             or other tree-based APIs can also be fairly directly implemented. However, delivering
    1269             Parabix performance to Java-based XML applications is challenging due to the
    1270             considerable overhead of crossing the Java Native Interface (JNI) boundary. This issue
    1271             is addressed with the Array Set Model (ASM) concept discussed in the following section. </para>
    1272          <para> With the recent development of parallel parsing using bitstream addition, it is
    1273             likely that the underlying ILAX interface of Parabix will change. In essence, ILAX
    1274             suffers the drawback of all event-based interfaces: they are fundamentally sequential in
    1275             number. As research continues, we expect efficient parallel methods building on parallel
    1276             bit stream foundations to move up the stack of XML processing requirements. Artificially
    1277             imposing sequential processing is thus expected to constrain further advances in XML
    1278             performance. </para>
    1279       </section>
    1281       <section>
    1282          <title>Efficient XML in Java Using Array Set Models</title>
    1283          <para> In our GML-to-SVG case study, we identified the lack of high-performance XML
    1284             processing solutions for Java to be of particular interest. Java byte code does not
    1285             provide access to the SIMD capabilities of the underlying machine architecture. Java
    1286             just-in-time compilers might be capable of using some SIMD facilities, but there is no
    1287             real prospect of conventional compiler technology translating byte-at-a-time algorithms
    1288             into parallel bit stream code. So the primary vehicle for delivering high-performance
    1289             XML processing is to call native parallel bit stream code written in C through JNI
    1290             capabilities. </para>
    1291          <para>However, each JNI call is expensive, so it is desirable to minimize the number of
    1292             calls and get as much work done during each call as possible. This mitigates against
    1293             direct implementation of streaming APIs in Java through one-to-one mappings to an
    1294             underlying streaming API in C. Instead, we have concentrated on gathering information on
    1295             the C side into data structures that can then be passed to the Java side. However, using
    1296             either C pointer-based structures or C++ objects is problematic because these are
    1297             difficult to interpret on the Java side and are not amenable to Java's automatic storage
    1298             management system. Similarly, Java objects cannot be conveniently created on the C side.
    1299             However, it is possible to transfer arrays of simple data values (bytes or integers)
    1300             between C and Java, so that makes a reasonable focus for bulk data communication between
    1301             C and Java. </para>
    1302          <para><emphasis>Array Set Models</emphasis> are array-based representations of information
    1303             representing an XML document in accord with XML InfoSet [<xref linkend="InfoSet"/>] or
    1304             other XML data models relevant to particular APIs. As well as providing a mechanism for
    1305             efficient bulk data communication across the JNI boundary, ASMs potentially have a
    1306             number of other benefits in high-performance XML processing. <itemizedlist>
    1307                <listitem>
    1308                   <para>Prefetching. Commodity processors commonly support hardware and/or software
    1309                      prefetching to ensure that data is available in a processor cache when it is
    1310                      needed. In general, prefetching is most effective in conjunction with the
    1311                      continuous sequential memory access patterns associated with array
    1312                   processing.</para>
    1313                </listitem>
    1314                <listitem>
    1315                   <para>DMA. Some processing environments provide Direct Memory Access (DMA)
    1316                      controllers for block data movement in parallel with computation. For example,
    1317                      the Cell Broadband Engine uses DMA controllers to move the data to and from the
    1318                      local stores of the synergistic processing units. Arrays of contiguous data
    1319                      elements are well suited to bulk data movement using DMA.</para>
    1320                </listitem>
    1321                <listitem>
    1322                   <para>SIMD. Single Instruction Multiple Data (SIMD) capabilities of modern
    1323                      processor instruction sets allow simultaneous application of particular
    1324                      instructions to sets of elements from parallel arrays. For effective use of
    1325                      SIMD capabilities, an SoA (Structure of Arrays) model is preferrable to an AoS
    1326                      (Array of Structures) model. </para>
    1327                </listitem>
    1328                <listitem>
    1329                   <para>Multicore processors. Array-oriented processing can enable the effective
    1330                      distribution of work to the individual cores of a multicore system in two
    1331                      distinct ways. First, provided that sequential dependencies can be minimized or
    1332                      eliminated, large arrays can be divided into separate segments to be processed
    1333                      in parallel on each core. Second, pipeline parallelism can be used to implement
    1334                      efficient multipass processing with each pass consisting of a processing kernel
    1335                      with array-based input and array-based output. </para>
    1336                </listitem>
    1337                <listitem>
    1338                   <para>Streaming buffers for large XML documents. In the event that an XML document
    1339                      is larger than can be reasonably represented entirely within processor memory,
    1340                      a buffer-based streaming model can be applied to work through a document using
    1341                      sliding windows over arrays of elements stored in document order. </para>
    1342                </listitem>
    1344             </itemizedlist>
    1345          </para>
    1347          <section>
    1348             <title>Saxon-B TinyTree Example</title>
    1349             <para>As a first example of the ASM concept, current work includes a proof-of-concept to
    1350                deliver a high-performance replacement for building the TinyTree data structure used
    1351                in Saxon-B 6.5.5, an open-source XSLT 2.0 processor written in Java [<xref
    1352                   linkend="Saxon"/>]. Although XSLT stylesheets may be cached for performance, the
    1353                caching of source XML documents is typically not possible. A new TinyTree object to
    1354                represent the XML source document is thus commonly constructed with each new query so
    1355                that the overall performance of simple queries on large source XML documents is
    1356                highly dependent on TinyTree build time. Indeed, in a study of Saxon-SA, the
    1357                commercial version of Saxon, query time was shown to be dominated by TinyTree build
    1358                time [<xref linkend="Kay08"/>]. Similar performance results are demonstrable for the
    1359                Saxon-B XSLT processor as well. </para>
    1360             <para> The Saxon-B processor studied is a pure Java solution, converting a SAX (Simple
    1361                API for XML) event stream into the TinyTree Java object using the efficient Aelfred
    1362                XML parser [<xref linkend="AElfred"/>]. The TinyTree structure is itself an
    1363                array-based structure mapping well suited to the ASM concept. It consists of six
    1364                parallel arrays of integers indexed on node number and containing one entry for each
    1365                node in the source document, with the exception of attribute and namespace nodes
    1366                   [<xref linkend="Saxon"/>]. Four of the arrays respectively provide node kind, name
    1367                code, depth, and next sibling information for each node, while the two others are
    1368                overloaded for different purposes based on node kind value. For example, in the
    1369                context of a text node , one of the overloaded arrays holds the text buffer offset
    1370                value whereas the other holds the text buffer length value. Attributes and namespaces
    1371                are represented using similiar parallel array of values. The stored TinyTree values
    1372                are primarily primitive Java types, however, object types such as Java Strings and
    1373                Java StringBuffers are also used to hold attribute values and comment values
    1374                respectively. </para>
    1375             <para> In addition to the TinyTree object, Saxon-B maintains a NamePool object which
    1376                represents a collection of XML name triplets. Each triplet is composed of a Namespace
    1377                URI, a Namespace prefix and a local name and encoded as an integer value known as a
    1378                namecode. Namecodes permit efficient name search and look-up using integer
    1379                comparison. Namecodes may also be subsequently decoded to recover namespace and local
    1380                name information. </para>
    1381             <para> Using the Parabix ILAX interface, a high-performance reimplementation of TinyTree
    1382                and NamePool data structures was built to compare with the Saxon-B implementation. In
    1383                fact, two functionally equivalent versions of the ASM java class were constructed. An
    1384                initial version was constructed based on a set of primitive Java arrays constructed
    1385                and allocated in the Java heap space via JNI New&lt;PrimitiveType&gt;Array
    1386                method call. In this version, the JVM garbage collector is aware of all memory
    1387                allocated in the native code. However, in this approach, large array copy operations
    1388                limited overall performance to approximately a 2X gain over the Saxon-B build time. </para>
    1389             <para>To further address the performance penalty imposed by copying large array values,
    1390                a second version of the ASM Java object was constructed based on natively backed
    1391                Direct Memory Byte Buffers [<xref linkend="JNI"/>]. In this version the JVM garbage
    1392                collector is unaware any native memory resources backing the Direct Memory Byte
    1393                Buffers. Large JNI-based copy operations are avoided; however, system memory must be
    1394                explicitly deallocated via a Java native method call. Using this approach, our
    1395                preliminary results show an approximate total 2.5X gain over Saxon-B build time.
    1396             </para>
    1397          </section>
    1398       </section>
    1399    </section>
    1402    <section>
    1403       <title>Compiler Technology</title>
    1405       <para> An important focus of our recent work is on the development of compiler technology to
    1406          automatically generate the low-level SIMD code necessary to implement bit stream processing
    1407          given suitable high-level specifications. This has several potential benefits. First, it
    1408          can eliminate the tedious and error-prone programming of bit stream operations in terms of
    1409          register-at-a-time SIMD operations. Second, compilation technology can automatically employ
    1410          a variety of performance improvement techniques that are difficult to apply manually. These
    1411          include algorithms for instruction scheduling and register allocation as well as
    1412          optimization techniques for common subexpression expression elimination and register
    1413          rematerialization among others. Third, compiler technology makes it easier to make changes
    1414          to the low-level code for reasons of perfective or adaptive maintenance.</para>
    1416       <para>Beyond these reasons, compiler technology also offers the opportunity for retargetting
    1417          the generation of code to accommodate different processor architectures and API
    1418          requirements. Strategies for efficient parallel bit stream code can vary considerably
    1419          depending on processor resources such as the number of registers available, the particular
    1420          instruction set architecture supported, the size of L1 and L2 data caches, the number of
    1421          available cores and so on. Separate implementation of custom code for each processor
    1422          architecture would thus be likely to be prohibitively expensive, prone to errors and
    1423          inconsistencies and difficult to maintain. Using compilation technology, however, the idea
    1424          would be to implement a variety of processor-specific back-ends all using a common front
    1425          end based on parallel bit streams. </para>
    1427       <section>
    1428          <title>Character Class Compiler</title>
    1430          <para>The first compiler component that we have implemented is a character class compiler,
    1431             capable of generation all the bit stream logic necessary to produce a set of lexical
    1432             item streams each corresponding to some particular set of characters to be recognized.
    1433             By taking advantage of common patterns between characters within classes, and special
    1434             optimization logic for recognizing character-class ranges, our existing compiler is able
    1435             to generate well-optimized code for complex sets of character classes involving numbers
    1436             of special characters as well as characters within specific sets of ranges. </para>
     142         <title>Xerces C++ Structure</title>
     144The Xerces C++ parser
     145<!-- is a widely-used standards-conformant -->
     146<!-- XML parser produced as open-source software -->
     147<!-- by the Apache Software Foundation. -->
     148<!-- It -->
     149features comprehensive support for a variety of character encodings
     150both commonplace (e.g., UTF-8, UTF-16) and rarely used (e.g., EBCDIC), support for
     151multiple XML vocabularies through the XML namespace
     152mechanism, as well as complete implementations
     153of structure and data validation through multiple grammars
     154declared using either legacy DTDs (document type
     155definitions) or modern XML Schema facilities.
     156Xerces also supports several APIs for accessing
     157parser services, including event-based parsing
     158using either pull parsing or SAX/SAX2 push-style
     159parsing as well as a DOM tree-based parsing interface.
     162<!--What is the story behind the xerces-profile picture? should it contain one single file or several from our test suite?-->
     163<!--Our test suite does not have any grammars in it; ergo, processing those files will give a poor indication of the cost of using grammars-->
     164<!--Should we show a val-grind summary of a few files in a linechart form?-->
     165Xerces, like all traditional parsers, processes XML documents sequentially a byte-at-a-time from the
     166first to the last byte of input data. Each byte passes through several processing layers and is
     167classified and eventually validated within the context of the document state.
     168This introduces implicit dependencies between the various tasks within the application that make it
     169difficult to optimize for performance.
     170As a complex software system, no one feature dominates the overall parsing performance.
     171Figure \ref{fig:xerces-profile} shows the execution time profile of the top ten functions in a typical run.
     172Even if it were possible, Amdahl's Law dictates that tackling any one of these functions for
     173parallelization in isolation would only produce a minute improvement in performance.
     174Unfortunately, early investigation into these functions found
     175that incorporating speculation-free thread-level parallelization was impossible
     176and they were already performing well in their given tasks;
     177thus only trivial enhancements were attainable.
     178In order to obtain a systematic acceleration of Xerces,
     179it should be expected that a comprehensive restructuring
     180is required, involving all aspects of the parser.
     183<!-- In order to obtain systematic acceleration of the Xerces parser,-->
     184<!-- it should be expected that a comprehensive restructuring-->
     185<!-- is required, involving all aspects of the parser.-->
     186<!-- FIGURE
     189Time (\%) & Function Name \\
     19113.29   &       XMLUTF8Transcoder::transcodeFrom \\
     1927.45    &       IGXMLScanner::scanCharData \\
     1936.83    &       memcpy \\
     1945.83    &       XMLReader::getNCName \\
     1954.67    &       IGXMLScanner::buildAttList \\
     1964.54    &       RefHashTableOf\verb|<>|::findBucketElem \\
     1974.20    &       IGXMLScanner::scanStartTagNS \\
     1983.75    &       ElemStack::mapPrefixToURI \\
     1993.58    &       ReaderMgr::getNextChar \\
     2003.20    &       IGXMLScanner::basicAttrValueScan \\
     202\caption{Execution Time of Top 10 Xerces Functions}
     203\label {fig:xerces-profile}
    1438207      </section>
    1439208      <section>
    1440          <title>Regular Expression Compilation</title>
     209         <title>The Parabix Framework</title>
     211The Parabix (parallel bit stream) framework is a transformative approach to XML parsing
     212(and other forms of text processing.) The key idea is to exploit the availability of wide
     213SIMD registers (e.g., 128-bit) in commodity processors to represent data from long blocks
     214of input data by using one register bit per single input byte.
     215To facilitate this, the input data is first transposed into a set of basis bit streams.
     216In <!--FIGURE REF Figure~\ref{fig:BitStreamsExample}, the ASCII string ``{\ttfamily b7\verb|<|A}''
     217is represented as 8 basis bit streams, $\tt b_{0 \ldots 7}$.
     219<!-- The bits used to construct $\tt b_7$ have been highlighted in this example. -->
     220Boolean-logic operations\footnote{&#8743;, \&#8744; and &#172; denote the boolean AND, OR and NOT operators.}
     221are used to classify the input bits into a set of {\it character-class bit streams}, which identify key
     222characters (or groups of characters) with a $1$.
     223For example, one of the fundamental characters in XML is a left-angle bracket.
     224A character is a<code>&lt; if and only if $\lnot(b_0 \lor b_1) \land (b_2 \land
     225b_3) \land (b_4 \land b_5) \land \lnot (b_6 \lor b_7) = 1</code>.
     226Similarly, a character is numeric
     227<code>[0-9] if and only if $\lnot(b_0 \lor b_1) \land (b_2 \land b_3) \land \lnot(b_4 \land (b_5 \lor b_6))</code>.
     228An important observation here is that ranges of characters may
     229require fewer operations than individual characters and
     230<!-- the classification cost could be amortized over many character classes.-->
     231multiple classes can share the classification cost.
     234<!-- FIGURE
     237\begin{tabular}{r c c c c }
     238String & \ttfamily{b} & \ttfamily{7} & \ttfamily{\verb`<`} & \ttfamily{A} \\
     239ASCII & \ttfamily{\footnotesize 0110001{\bfseries 0}} & \ttfamily{\footnotesize 0011011{\bfseries 1}} & \ttfamily{\footnotesize 0011110{\bfseries 0}} & \ttfamily{\footnotesize 0100000{\bfseries 1}} \\
     244\begin{tabular}{r |c |c |c |c |c |c |c |c |}
     245 & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{0}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{1}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{2}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{3}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{4}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{5}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{6}$}$ & $\mbox{\fontsize{11}{11}\selectfont $\tt b_{7}$}$ \\
     246 & \ttfamily{0} & \ttfamily{1} & \ttfamily{1} & \ttfamily{0} & \ttfamily{0} & \ttfamily{0} & \ttfamily{1} & \bfseries\ttfamily{0} \\
     247 & \ttfamily{0} & \ttfamily{0} & \ttfamily{1} & \ttfamily{1} & \ttfamily{0} & \ttfamily{1} & \ttfamily{1} & \bfseries\ttfamily{1} \\
     248 & \ttfamily{0} & \ttfamily{0} & \ttfamily{1} & \ttfamily{1} & \ttfamily{1} & \ttfamily{1} & \ttfamily{0} & \bfseries\ttfamily{0} \\
     249 & \ttfamily{0} & \ttfamily{1} & \ttfamily{0} & \ttfamily{0} & \ttfamily{0} & \ttfamily{0} & \ttfamily{0} & \bfseries\ttfamily{1} \\
     252\caption{8-bit ASCII Basis Bit Streams}
     257<!-- Using a mixture of boolean-logic and arithmetic operations, character-class -->
     258<!-- bit streams can be transformed into lexical bit streams, where the presense of -->
     259<!-- a 1 bit identifies a key position in the input data. As an artifact of this -->
     260<!-- process, intra-element well-formedness validation is performed on each block -->
     261<!-- of text. -->
     263Consider, for example, the XML source data stream shown in the first line of <!-- FIGURE REF Figure \ref{fig:parabix1} -->.
     264The remaining lines of this figure show several parallel bit streams that are computed in Parabix-style
     265parsing, with each bit of each stream in one-to-one correspondence to the source character code units
     266of the input stream.
     267For clarity, 1 bits are denoted with 1 in each stream and 0 bits are represented as underscores.
     268The first bit stream shown is that for the opening
     269angle brackets that represent tag openers in XML.
     270The second and third streams show a partition of the
     271tag openers into start tag marks and end tag marks
     272depending on the character immediately following the
     273opener (i.e., ``\verb:/:'') or not.  The remaining three
     274lines show streams that can be computed in subsequent
     275parsing (using the technique
     276of \bitstream{} addition \cite{cameron-EuroPar2011}), namely streams marking the element names,
     277attribute names and attribute values of tags. 
     280Two intuitions may help explain how the Parabix approach can lead
     281to improved XML parsing performance. The first is that
     282the use of the full register width offers a considerable
     283information advantage over sequential byte-at-a-time
     284parsing.  That is, sequential processing of bytes
     285uses just 8 bits of each register, greatly limiting the
     286processor resources that are effectively being used at any one time.
     287The second is that byte-at-a-time loop scanning loops are actually
     288often just computing a single bit of information per iteration:
     289is the scan complete yet?
     290Rather than computing these individual decision-bits, an approach that computes
     291many of them in parallel (e.g., 128 bytes at a time using 128-bit registers)
     292should provide substantial benefit.
     295Previous studies have shown that the Parabix approach improves many aspects of XML processing,
     296including transcoding \cite{Cameron2008}, character classification and validation,
     297tag parsing and well-formedness checking. 
     298The first Parabix parser used processor bit scan instructions to considerably accelerate
     299sequential scanning loops for individual characters \cite{CameronHerdyLin2008}.
     300Recent work has incorporated a method of parallel
     301scanning using \bitstream{} addition \cite{cameron-EuroPar2011}, as
     302well as combining SIMD methods with 4-stage pipeline parallelism to further improve
     303throughput \cite{HPCA2012}.
     304Although these research prototypes handled the full syntax of schema-less XML documents,
     305they lacked the functionality required by full XML parsers.
     308Commercial XML processors support transcoding of multiple character sets and can parse and
     309validate against multiple document vocabularies.
     310Additionally, they provide API facilities beyond those found in research prototypes,
     311including the widely used SAX, SAX2 and DOM interfaces.
     313      </section>
     314      <section>
     315         <title>Sequential vs. Parallel Paradigm</title>
     317Xerces&#8212;like all traditional XML parsers&#8212;processes XML documents sequentially.
     318Each character is examined to distinguish between the
     319XML-specific markup, such as a left angle bracket <code>&lt;</code>, and the
     320content held within the document. 
     321As the parser progresses through the document, it alternates between markup scanning,
     322validation and content processing modes.
     325In other words, Xerces belongs to an equivalent class applications termed FSM applications\footnote{
     326  Herein FSM applications are considered software systems whose behaviour is defined by the inputs,
     327  current state and the events associated with transitions of states.}.
     328Each state transition indicates the processing context of subsequent characters.
     329Unfortunately, textual data tends to be unpredictable and any character could induce a state transition.
     332Parabix-style XML parsers utilize a concept of layered processing.
     333A block of source text is transformed into a set of lexical \bitstream{}s,
     334which undergo a series of operations that can be grouped into logical layers,
     335e.g., transposition, character classification, and lexical analysis.
     336Each layer is pipeline parallel and require neither speculation nor pre-parsing stages\cite{HPCA2012}.
     337To meet the API requirements of the document-ordered Xerces output,
     338the results of the Parabix processing layers must be interleaved to produce the equivalent behaviour.
     340      </section>                       
     341     </section>                 
     342                <section>
     343                        <title>Architecture</title>             
     344                        <section>
     345                       <title>Overview</title>
     346<!--\def \CSG{Content Stream Generator}-->
     348\icXML{} is more than an optimized version of Xerces. Many components were grouped, restructured and
     349rearchitected with pipeline parallelism in mind.
     350In this section, we highlight the core differences between the two systems.
     351As shown in Figure \ref{fig:xerces-arch}, Xerces
     352is comprised of five main modules: the transcoder, reader, scanner, namespace binder, and validator.
     353The {\it Transcoder} converts source data into UTF-16 before Xerces parses it as XML;
     354the majority of the character set encoding validation is performed as a byproduct of this process.
     355The {\it Reader} is responsible for the streaming and buffering of all raw and transcoded (UTF-16) text.
     356It tracks the current line/column position,
     357<!--(which is reported in the unlikely event that the input contains an error), -->
     358performs line-break normalization and validates context-specific character set issues,
     359such as tokenization of qualified-names.
     360The {\it Scanner} pulls data through the reader and constructs the intermediate representation (IR)
     361of the document; it deals with all issues related to entity expansion, validates
     362the XML well-formedness constraints and any character set encoding issues that cannot
     363be completely handled by the reader or transcoder (e.g., surrogate characters, validation
     364and normalization of character references, etc.)
     365The {\it Namespace Binder} is a core piece of the element stack.
     366It handles namespace scoping issues between different XML vocabularies.
     367This allows the scanner to properly select the correct schema grammar structures.
     368The {\it Validator} takes the IR produced by the Scanner (and
     369potentially annotated by the Namespace Binder) and assesses whether the final output matches
     370the user-defined DTD and schema grammar(s) before passing it to the end-user.
     373<!-- FIGURE
     377\caption{Xerces Architecture}
     384In \icXML{} functions are grouped into logical components.
     385As shown in Figure \ref{fig:icxml-arch}, two major categories exist: (1) the \PS{} and (2) the \MP{}.
     386All tasks in (1) use the Parabix Framework \cite{HPCA2012}, which represents data as a set of parallel \bitstream{}s.
     387The {\it Character Set Adapter}, discussed in Section \ref{arch:character-set-adapter},
     388mirrors Xerces's Transcoder duties; however instead of producing UTF-16 it produces a
     389set of lexical \bitstream{}s, similar to those shown in Figure \ref{fig:parabix1}.
     390These lexical \bitstream{}s are later transformed into UTF-16 in the \CSG{},
     391after additional processing is performed.
     392The first precursor to producing UTF-16 is the {\it Parallel Markup Parser} phase.
     393It takes the lexical streams and produces a set of marker \bitstream{}s in which a 1-bit identifies
     394significant positions within the input data. One \bitstream{} for each of the critical piece of information is created, such as
     395the beginning and ending of start tags, end tags, element names, attribute names, attribute values and content.
     396Intra-element well-formedness validation is performed as an artifact of this process.
     397Like Xerces, \icXML{} must provide the Line and Column position of each error.
     398The {\it Line-Column Tracker} uses the lexical information to keep track of the document position(s) through the use of an
     399optimized population count algorithm, described in Section \ref{section:arch:errorhandling}.
     400From here, two data-independent branches exist: the Symbol Resolver and Content Preparation Unit.
     403A typical XML file contains few unique element and attribute names&#8212;but each of them will occur frequently.
     404\icXML{} stores these as distinct data structures, called symbols, each with their own global identifier (GID).
     405Using the symbol marker streams produced by the Parallel Markup Parser, the {\it Symbol Resolver} scans through
     406the raw data to produce a sequence of GIDs, called the {\it symbol stream}.
     409The final components of the \PS{} are the {\it Content Preparation Unit} and {\it \CSG{}}.
     410The former takes the (transposed) basis \bitstream{}s and selectively filters them, according to the
     411information provided by the Parallel Markup Parser, and the latter transforms the
     412filtered streams into the tagged UTF-16 {\it content stream}, discussed in Section \ref{section:arch:contentstream}.
     415Combined, the symbol and content stream form \icXML{}'s compressed IR of the XML document.
     416The {\it \MP{}}~parses the IR to validate and produce the sequential output for the end user.
     417The {\it Final WF checker} performs inter-element well-formedness validation that would be too costly
     418to perform in bit space, such as ensuring every start tag has a matching end tag.
     419Xerces's namespace binding functionality is replaced by the {\it Namespace Processor}. Unlike Xerces,
     420it is a discrete phase that produces a series of URI identifiers (URI IDs), the {\it URI stream}, which are
     421associated with each symbol occurrence.
     422This is discussed in Section \ref{section:arch:namespacehandling}.
     423Finally, the {\it Validation} layer implements the Xerces's validator.
     424However, preprocessing associated with each symbol greatly reduces the work of this stage.
     427<!-- FIGURE
     432\caption{\icXML{} Architecture}
     437            </section>
     438                        <section>
     439                       <title>Character Set Adapters</title>
     441In Xerces, all input is transcoded into UTF-16 to simplify the parsing costs of Xerces itself and
     442provide the end-consumer with a single encoding format.
     443In the important case of UTF-8 to UTF-16 transcoding, the transcoding costs can be significant,
     444because of the need to decode and classify each byte of input, mapping variable-length UTF-8
     445byte sequences into 16-bit UTF-16 code units with bit manipulation operations.   
     446In other cases, transcoding may involve table look-up operations for each byte of input.  In any case,
     447transcoding imposes at least a cost of buffer copying.
     450In \icXML{}, however,  the concept of Character Set Adapters (CSAs) is used to minimize transcoding costs.
     451Given a specified input encoding, a CSA is responsible for checking that
     452input code units represent valid characters, mapping the characters of the encoding into
     453the appropriate \bitstream{}s for XML parsing actions (i.e., producing the lexical item
     454streams), as well as supporting ultimate transcoding requirements.   All of this work
     455is performed using the parallel \bitstream{} representation of the source input.
     458An important observation is that many character sets are an
     459extension to the legacy 7-bit ASCII character set.  This includes the
     460various ISO Latin character sets, UTF-8, UTF-16 and many others.
     461Furthermore, all significant characters for parsing XML are confined to the
     462ASCII repertoire.   Thus, a single common set of lexical item calculations
     463serves to compute lexical item streams for all such ASCII-based character sets.
     466A second observation is that&#8212;regardless of which character set is used&#8212;quite
     467often all of the characters in a particular block of input will be within the ASCII range.
     468This is a very simple test to perform using the \bitstream{} representation, simply confirming that the
     469bit 0 stream is zero for the entire block.   For blocks satisfying this test,
     470all logic dealing with non-ASCII characters can simply be skipped.
     471Transcoding to UTF-16 becomes trivial as the high eight \bitstream{}s of the
     472UTF-16 form are each set to zero in this case.
     475A third observation is that repeated transcoding of the names of XML
     476elements, attributes and so on can be avoided by using a look-up mechanism.
     477That is, the first occurrence of each symbol is stored in a look-up
     478table mapping the input encoding to a numeric symbol ID.   Transcoding
     479of the symbol is applied at this time.  Subsequent look-up operations
     480can avoid transcoding by simply retrieving the stored representation.
     481As symbol look up is required to apply various XML validation rules,
     482there is achieves the effect of transcoding each occurrence without
     483additional cost.
     486The cost of individual character transcoding is avoided whenever a block of input is
     487confined to the ASCII subset and for all but the first occurrence of any XML element or attribute name.
     488Furthermore, when transcoding is required, the parallel \bitstream{} representation
     489supports efficient transcoding operations.   
     490In the important case of UTF-8 to UTF-16 transcoding, the corresponding UTF-16 \bitstream{}s
     491can be calculated in bit parallel fashion based on UTF-8 streams \cite{Cameron2008},
     492and all but the final bytes of multi-byte sequences can be marked for deletion as
     493discussed in the following subsection.
     494In other cases, transcoding within a block only need be applied for non-ASCII
     495bytes, which are conveniently identified by iterating through the bit 0 stream
     496using bit scan operations.
     498            </section>
     499                        <section>
     500                       <title>Combined Parallel Filtering</title>
     502As just mentioned, UTF-8 to UTF-16 transcoding involves marking
     503all but the last bytes of multi-byte UTF-8 sequences as
     504positions for deletion.   For example, the two
     505Chinese characters \begin{CJK*}{UTF8}{gbsn}䜠奜\end{CJK*}
     506are represented as two three-byte UTF-8 sequences \verb'E4 BD A0'
     507and \verb'E5 A5 BD' while the UTF-16 representation must be
     508compressed down to the two code units \verb'4F60' and \verb'597D'.
     509In the bit parallel representation, this corresponds to a reduction
     510from six bit positions representing UTF-8 code units (bytes)
     511down to just two bit positions representing UTF-16 code units
     512(double bytes).   This compression may be achieved by
     513arranging to calculate the correct UTF-16 bits at the
     514final position of each sequence and creating a deletion
     515mask to mark the first two bytes of each 3-byte sequence
     516for deletion.   In this case, the portion of the mask
     517corresponding to these input bytes is the bit sequence
     518\verb'110110'.  Using this approach, transcoding may then be
     519completed by applying parallel deletion and inverse transposition of the
     520UTF-16 \bitstream{}s\cite{Cameron2008}.
     523<!-- FIGURE
     527Source Data & \verb`<document>fee<element a1='fie' a2 = 'foe'></element>fum</document>`\\
     529<!-- Tag Openers & \verb`1____________1____________________________1____________1__________`\\-->
     530<!-- Start Tag Marks & \verb`_1____________1___________________________________________________`\\-->
     531<!-- End Tag Marks & \verb`___________________________________________1____________1_________`\\-->
     532<!-- Empty Tag Marks & \verb`__________________________________________________________________`\\-->
     533<!-- Element Names & \verb`_11111111_____1111111_____________________________________________`\\-->
     534<!-- Attribute Names & \verb`______________________11_______11_________________________________`\\-->
     535<!-- Attribute Values & \verb`__________________________111________111__________________________`\\-->
     536<!-- FIGURE
     537String Ends & \verb`1____________1_______________1__________1_1____________1__________`\\
     538Markup Identifiers & \verb`_________1______________1_________1______1_1____________1_________`\\
     539Deletion Mask & \verb`_11111111_____1111111111_1____1111_11_______11111111_____111111111`\\
     540Undeleted Data & \verb``{\tt\it 0}\verb`________>fee`{\tt\it 0}\verb`__________=_fie`{\tt\it 0}\verb`____=__foe`{\tt\it 0}\verb`>`{\tt\it 0}\verb`/________fum`{\tt\it 0}\verb`/_________`
     543\caption{XML Source Data and Derived Parallel Bit Streams}
     549Rather than immediately paying the
     550costs of deletion and transposition just for transcoding,
     551however, \icXML{} defers these steps so that the deletion
     552masks for several stages of processing may be combined.
     553In particular, this includes core XML requirements
     554to normalize line breaks and to replace character
     555reference and entity references by their corresponding
     556text.   In the case of line break normalization,
     557all forms of line breaks, including bare carriage
     558returns (CR), line feeds (LF) and CR-LF combinations
     559must be normalized to a single LF character in
     560each case.   In \icXML{}, this is achieved by
     561first marking CR positions, performing two
     562bit parallel operations to transform the marked
     563CRs into LFs, and then marking for deletion any
     564LF that is found immediately after the marked CR
     565as shown by the Pablo source code in Figure \ref{fig:LBnormalization}.
     566<!-- FIGURE
     569# XML 1.0 line-break normalization rules.
     570if lex.CR:
     571# Modify CR (#x0D) to LF (#x0A)
     572  u16lo.bit_5 ^= lex.CR
     573  u16lo.bit_6 ^= lex.CR
     574  u16lo.bit_7 ^= lex.CR
     575  CRLF = pablo.Advance(lex.CR) & lex.LF
     576  callouts.delmask |= CRLF
     577# Adjust LF streams for line/column tracker
     578  lex.LF |= lex.CR
     579  lex.LF ^= CRLF
     581\caption{Line Break Normalization Logic}\label{fig:LBnormalization}
     586In essence, the deletion masks for transcoding and
     587for line break normalization each represent a bitwise
     588filter; these filters can be combined using bitwise-or
     589so that the parallel deletion algorithm need only be
     590applied once.
     593A further application of combined filtering
     594is the processing of XML character and entity
     595references.   Consider, for example, the references <code>&amp;</code> or <code>&#x3C;</code>.
     596which must be replaced in XML processing with 
     597the single <code>&amp;</code> and <code>&lt;</code> characters, respectively.
     598The approach in \icXML{} is to mark all but the first character
     599positions of each reference for deletion, leaving a
     600single character position unmodified.  Thus, for the
     601references <code>&amp;</code> or <code>&#x3C;</code> the
     602masks <code>01111</code> and <code>011111</code> are formed and
     603combined into the overall deletion mask.   After the
     604deletion and inverse transposition operations are finally
     605applied, a post-processing step inserts the proper character
     606at these positions.   One note about this process is
     607that it is speculative; references are assumed to generally
     608be replaced by a single UTF-16 code unit.   In the case,
     609that this is not true, it is addressed in post-processing.
     612The final step of combined filtering occurs during
     613the process of reducing markup data to tag bytes
     614preceding each significant XML transition as described
     615in section~\ref{section:arch:contentstream}.  Overall, \icXML{}
     616avoids separate buffer copying operations for each of the
     617these filtering steps, paying the cost of parallel
     618deletion and inverse transposition only once. 
     619Currently, \icXML{} employs the parallel-prefix compress algorithm
     620of Steele~\cite{HackersDelight}  Performance
     621is independent of the number of positions deleted.
     622Future versions of \icXML{} are expected to
     623take advantage of the parallel extract operation~\cite{HilewitzLee2006}
     624that Intel is now providing in its Haswell architecture.
     626            </section>
     627                        <section>
     628                       <title>Content Stream</title>
     630A relatively-unique concept for \icXML{} is the use of a filtered content stream.
     631Rather that parsing an XML document in its original format, the input is transformed
     632into one that is easier for the parser to iterate through and produce the sequential
     634In <!-- FIGURE REF Figure~\ref{fig:parabix2} -->, the source data
     635<!-- \verb|<root><t1>text</t1><t2 a1=’foo’ a2 = ’fie’>more</t2><tag3 att3=’b’/></root>| -->
     636is transformed into
     637<!-- CODE -->
     638<!--``{\tt\it 0}\verb`>fee`{\tt\it 0}\verb`=fie`{\tt\it 0}\verb`=foe`{\tt\it 0}\verb`>`{\tt\it 0}\verb`/fum`{\tt\it 0}\verb`/`''-->
     639through the parallel filtering algorithm, described in section \ref{sec:parfilter}.
     642Combined with the symbol stream, the parser traverses the content stream to effectively
     643reconstructs the input document in its output form.
     644The initial {\tt\it 0} indicates an empty content string. The following \verb|>|
     645indicates that a start tag without any attributes is the first element in this text and
     646the first unused symbol, <code>document</code>, is the element name.
     647Succeeding that is the content string <code>fee</code>, which is null-terminated in accordance
     648with the Xerces API specification. Unlike Xerces, no memory-copy operations
     649are required to produce these strings, which as Figure~\ref{fig:xerces-profile} shows
     650accounts for 6.83% of Xerces's execution time.
     651Additionally, it is cheap to locate the terminal character of each string:
     652using the String End \bitstream{}, the \PS{} can effectively calculate the offset of each
     653null character in the content stream in parallel, which in turn means the parser can
     654directly jump to the end of every string without scanning for it.
     657Following ``\verb`fee`'' is a \verb`=`, which marks the existence of an attribute.
     658Because all of the intra-element was performed in the \PS{}, this must be a legal attribute.
     659Since attributes can only occur within start tags and must be accompanied by a textual value,
     660the next symbol in the symbol stream must be the element name of a start tag,
     661and the following one must be the name of the attribute and the string that follows the \verb`=` must be its value.
     662However, the subsequent \verb`=` is not treated as an independent attribute because the parser has yet to
     663read a \verb`>`, which marks the end of a start tag. Thus only one symbol is taken from the symbol stream and
     664it (along with the string value) is added to the element.
     665Eventually the parser reaches a \verb`/`, which marks the existence of an end tag. Every end tag requires an
     666element name, which means they require a symbol. Inter-element validation whenever an empty tag is detected to
     667ensure that the appropriate scope-nesting rules have been applied.
     669            </section>
     670                        <section>
     671                       <title>Namespace Handling</title>
     672<!-- Should we mention canonical bindings or speculation? it seems like more of an optimization than anything. -->
     674In XML, namespaces prevents naming conflicts when multiple vocabularies are used together.
     675It is especially important when a vocabulary application-dependant meaning, such as when
     676XML or SVG documents are embedded within XHTML files.
     677Namespaces are bound to uniform resource identifiers (URIs), which are strings used to identify
     678specific names or resources.
     679On line 1 of Figure \ref{fig:namespace1}, the \verb|xmlns| attribute instructs the XML
     680processor to bind the prefix <code>p</code> to the URI &apos;<code>pub.net</code>&apos; and the default (empty)
     681prefix to <code>book.org</code>. Thus to the XML processor, the \verb|title| on line 2 and
     682\verb|price| on line 4 both read as \verb|"book.org":title| and \verb|"book.org":price|
     683respectively, whereas on line 3 and 5, \verb|p:name| and \verb|price| are seen as
     684\verb|"pub.net":name| and \verb|"pub.net":price|. Even though the actual element name
     685\verb|price|, due to namespace scoping rules they are viewed as two uniquely-named items
     686because the current vocabulary is determined by the namespace(s) that are in-scope.
     689<!-- FIGURE
     6921. & \verb|<book xmlns:p="pub.net" xmlns="book.org">| \\
     6932. & \verb|  <title>BOOK NAME</title>| \\
     6943. & \verb|  <p:name>PUBLISHER NAME</p:name>| \\
     6954. & \verb|  <price>X</price>| \\
     6965. & \verb|  <price xmlns="publisher.net">Y</price>| \\
     6976. & \verb|</book>| \\
     699\caption{XML Namespace Example}
     700\label {fig:namespace1}
     705In both Xerces and \icXML{}, every URI has a one-to-one mapping to a URI ID.
     706These persist for the lifetime of the application through the use of a global URI pool.
     707Xerces maintains a stack of namespace scopes that is pushed (popped) every time a start tag (end tag) occurs
     708in the document. Because a namespace declaration affects the entire element, it must be processed prior to
     709grammar validation. This is a costly process considering that a typical namespaced XML document only comes
     710in one of two forms:
     711(1) those that declare a set of namespaces upfront and never change them, and
     712(2) those that repeatedly modify the namespaces in predictable patterns.
     715For that reason, \icXML{} contains an independent namespace stack and utilizes bit vectors to cheaply perform
     716<!-- speculation and scope resolution options with a single XOR operation &#8212; even if many alterations are performed. -->
     717<!-- performance advantage figure?? average cycles/byte cost? -->
     718When a prefix is declared (e.g., \verb|xmlns:p="pub.net"|), a namespace binding is created that maps
     719the prefix (which are assigned Prefix IDs in the symbol resolution process) to the URI.
     720Each unique namespace binding has a unique namespace id (NSID) and every prefix contains a bit vector marking every
     721NSID that has ever been associated with it within the document. For example, in Table \ref{tbl:namespace1}, the
     722prefix binding set of \verb|p| and \verb|xmlns| would be \verb|01| and \verb|11| respectively.
     723To resolve the in-scope namespace binding for each prefix, a bit vector of the currently visible namespaces is
     724maintained by the system. By ANDing the prefix bit vector with the currently visible namespaces, the in-scope
     725NSID can be found using a bit-scan intrinsic.
     726A namespace binding table, similar to Table \ref{tbl:namespace1}, provides the actual URI ID.
     729<!-- FIGURE
     733NSID & Prefix & URI & Prefix ID & URI ID \\ \hline\hline
     7340 & {\tt p} & {\tt pub.net} & 0 & 0 \\ \hline
     7351 & {\tt xmlns} & {\tt books.org} & 1 & 1 \\ \hline
     7362 & {\tt xmlns} & {\tt pub.net} & 1 & 0 \\ \hline
     738\caption{Namespace Binding Table Example}
     745<!-- PrefixBindings = PrefixBindingTable[prefixID]; -->
     746<!-- VisiblePrefixBinding = PrefixBindings & CurrentlyVisibleNamespaces; -->
     747<!-- NSid = bitscan(VisiblePrefixBinding); -->
     748<!-- URIid = NameSpaceBindingTable[NSid].URIid; -->
     751To ensure that scoping rules are adhered to,
     752whenever a start tag is encountered, any modification to the currently visible namespaces is calculated and stored
     753within a stack of bit vectors denoting the locally modified namespace bindings. When an end tag is found, the
     754currently visible namespaces is XORed with the vector at the top of the stack.
     755This allows any number of changes to be performed at each scope-level with a constant time.
     756<!-- Speculation can be handled by probing the historical information within the stack but that goes beyond the scope of this paper.-->
     758            </section>
     759                        <section>
     760                       <title>Error Handling</title>
     762<!-- XML errors are rare but they do happen, especially with untrustworthy data sources.-->
     763Xerces outputs error messages in two ways: through the programmer API and as thrown objects for fatal errors.
     764As Xerces parses a file, it uses context-dependant logic to assess whether the next character is legal;
     765if not, the current state determines the type and severity of the error.
     766\icXML{} emits errors in the similar manner&#8212;but how it discovers them is substantially different.
     767Recall that in Figure \ref{fig:icxml-arch}, \icXML{} is divided into two sections: the \PS{} and \MP{},
     768each with its own system for detecting and producing error messages.
     771Within the \PS{}, all computations are performed in parallel, a block at a time.
     772Errors are derived as artifacts of \bitstream{} calculations, with a 1-bit marking the byte-position of an error within a block,
     773and the type of error is determined by the equation that discovered it.
     774The difficulty of error processing in this section is that in Xerces the line and column number must be given
     775with every error production. Two major issues exist because of this:
     776(1) line position adheres to XML white-normalization rules; as such, some sequences of characters, e.g., a carriage return
     777followed by a line feed, are counted as a single new line character.
     778(2) column position is counted in characters, not bytes or code units;
     779thus multi-code-unit code-points and surrogate character pairs are all counted as a single column position.
     780Note that typical XML documents are error-free but the calculation of the
     781line/column position is a constant overhead in Xerces. <!-- that must be maintained in the case that one occurs. -->
     782To reduce this, \icXML{} pushes the bulk cost of the line/column calculation to the occurrence of the error and
     783performs the minimal amount of book-keeping necessary to facilitate it.
     784\icXML{} leverages the byproducts of the Character Set Adapter (CSA) module and amalgamates the information
     785within the Line Column Tracker (LCT).
     786One of the CSA's major responsibilities is transcoding an input text. <!-- from some encoding format to near-output-ready UTF-16. -->
     787During this process, white-space normalization rules are applied and multi-code-unit and surrogate characters are detected
     788and validated.
     789A {\it line-feed \bitstream{}}, which marks the positions of the normalized new lines characters, is a natural derivative of
     790this process.
     791Using an optimized population count algorithm, the line count can be summarized cheaply for each valid block of text.
     792<!-- The optimization delays the counting process .... -->
     793Column position is more difficult to calculate.
     794It is possible to scan backwards through the \bitstream{} of new line characters to determine the distance (in code-units)
     795between the position between which an error was detected and the last line feed. However, this distance may exceed
     796than the actual character position for the reasons discussed in (2).
     797To handle this, the CSA generates a {\it skip mask} \bitstream{} by ORing together many relevant \bitstream{}s,
     798such as all trailing multi-code-unit and surrogate characters, and any characters that were removed during the
     799normalization process.
     800When an error is detected, the sum of those skipped positions is subtracted from the distance to determine the actual
     801column number.
     804The \MP{} is a state-driven machine. As such, error detection within it is very similar to Xerces.
     805However, reporting the correct line/column is a much more difficult problem.
     806The \MP{} parses the content stream, which is a series of tagged UTF-16 strings.
     807Each string is normalized in accordance with the XML specification.
     808All symbol data and unnecessary whitespace is eliminated from the stream;
     809thus its impossible to derive the current location using only the content stream.
     810To calculate the location, the \MP{} borrows three additional pieces of information from the \PS{}:
     811the line-feed, skip mask, and a {\it deletion mask stream}, which is a \bitstream{} denoting the (code-unit) position of every
     812datum that was suppressed from the source during the production of the content stream.
     813Armed with these, it is possible to calculate the actual line/column using
     814the same system as the \PS{} until the sum of the negated deletion mask stream is equal to the current position.
     816            </section>
     817                </section>
    1442          <para>Based on the character class compiler, we are currently investigating the
    1443             construction of a regular expression compiler that can implement bit-stream based
    1444             parallel regular-expression matching similar to that describe previously for parallel
    1445             parsing by bistream addition. This compiler works with the assumption that bitstream
    1446             regular-expression definitions are deterministic; no backtracking is permitted with the
    1447             parallel bit stream representation. In XML applications, this compiler is primarily
    1448             intended to enforce regular-expression constraints on string datatype specifications
    1449             found in XML schema. </para>
     819                <section>
     820                        <title>Multithreading with Pipeline Parallelism</title>         
     822As discussed in section \ref{background:xerces}, Xerces can be considered a FSM application.
     823These are ``embarrassingly sequential.''\cite{Asanovic:EECS-2006-183} and notoriously difficult to parallelize.
     824However, \icXML{} is designed to organize processing into logical layers.   
     825In particular, layers within the \PS{} are designed to operate
     826over significant segments of input data before passing their outputs on for
     827subsequent processing.  This fits well into the general model of pipeline
     828parallelism, in which each thread is in charge of a single module or group
     829of modules.
     832The most straightforward division of work in \icXML{} is to separate
     833the \PS{} and the \MP{} into distinct logical layers into two separate stages.
     834The resultant application, {\it\icXMLp{}}, is a course-grained software-pipeline application.
     835In this case, the \PS{} thread $T_1$ reads 16k of XML input $I$ at a time and produces the
     836content, symbol and URI streams, then stores them in a pre-allocated shared data structure $S$.
     837The \MP{} thread $T_2$ consumes $S$, performs well-formedness and grammar-based validation,
     838and the provides parsed XML data to the application through the Xerces API. 
     839The shared data structure is implemented using a ring buffer,
     840where every entry contains an independent set of data streams.
     841In the examples of Figure \ref{threads_timeline1} and \ref{threads_timeline2}, the ring buffer has four entries.
     842A lock-free mechanism is applied to ensure that each entry can only be read or written by one thread at the same time.
     843In Figure \ref{threads_timeline1} the processing time of $T_1$ is longer than $T_2$;
     844thus $T_2$ always waits for $T_1$ to write to the shared memory.
     845Figure \ref{threads_timeline2} illustrates the scenario in which $T_1$ is faster
     846and must wait for $T_2$ to finish reading the shared data before it can reuse the memory space.
     849<!-- FIGURE
     860\caption{Thread Balance in Two-Stage Pipelines}
    1451       </section>
     866Overall, our design is intended to benefit a range of applications.
     867Conceptually, we consider two design points.
     868The first, the parsing performed by the \PS{} dominates at 67% of the overall cost,
     869with the cost of application processing (including the driver logic within the \MP{}) at 33%.   
     870The second is almost the opposite scenario, the cost of application processing dominates at 60%,
     871while the cost of XML parsing represents an overhead of 40%.
     874Our design is predicated on a goal of using the Parabix
     875framework to achieve a 50% to 100% improvement in the parsing engine itself.   
     876In a best case scenario,
     877a 100% improvement of the \PS{} for the design point in which
     878XML parsing dominates at 67% of the total application cost.
     879In this case, the single-threaded \icXML{} should achieve a 1.5x speedup over Xerces
     880so that the total application cost reduces to 67% of the original. 
     881However, in \icXMLp{}, our ideal scenario gives us two well-balanced threads
     882each performing about 33% of the original work.   
     883In this case, Amdahl's law predicts that we could expect up to a 3x speedup at best.
     886At the other extreme of our design range, we consider an application
     887in which core parsing cost is 40%.   Assuming the 2x speedup of
     888the \PS{} over the corresponding Xerces core, single-threaded
     889\icXML{} delivers a 25% speedup.   However, the most significant
     890aspect of our two-stage multi-threaded design then becomes the
     891ability to hide the entire latency of parsing within the serial time
     892required by the application.   In this case, we achieve
     893an overall speedup in processing time by 1.67x.
     896Although the structure of the \PS{} allows division of the work into
     897several pipeline stages and has been demonstrated to be effective
     898for four pipeline stages in a research prototype \cite{HPCA2012},
     899our analysis here suggests that the further pipelining of work within
     900the \PS{} is not worthwhile if the cost of application logic is little as
     90133% of the end-to-end cost using Xerces.  To achieve benefits of
     902further parallelization with multi-core technology, there would
     903need to be reductions in the cost of application logic that
     904could match reductions in core parsing cost.
     906                </section>
    1453       <section>
    1454          <title>Unbounded Bit Stream Compilation</title>
     908                <section>
     909                        <title>Performance</title>             
     911We evaluate \xerces{}, \icXML{}, \icXMLp{} against two benchmarking applications:
     912the Xerces C++ SAXCount sample application,
     913and a real world GML to SVG transformation application.
     914We investigated XML parser performance using an Intel Core i7 quad-core
     915(Sandy Bridge) processor (3.40GHz, 4 physical cores, 8 threads (2 per core),
     91632+32 kB (per core) L1 cache,
     917256 kB (per core) L2 cache,
     9188 MB L3 cache) running the 64-bit version of Ubuntu 12.04 (Linux).
     921We analyzed the execution profiles of each XML parser
     922using the performance counters found in the processor.
     923We chose several key hardware events that provide insight into the profile of each
     924application and indicate if the processor is doing useful work. 
     925The set of events included in our study are:
     926processor cycles, branch instructions, branch mispredictions,
     927and cache misses. The Performance Application Programming Interface
     928(PAPI) Version 5.5.0 \cite{papi} toolkit
     929was installed on the test system to facilitate the
     930collection of hardware performance monitoring
     931statistics. In addition, we used the Linux perf \cite{perf} utility
     932to collect per core hardware events.
     934                        <section>
     935                       <title>Xerces C++ SAXCount</title>
     937Xerces comes with sample applications that demonstrate salient features of the parser.
     938SAXCount is the simplest such application:
     939it counts the elements, attributes and characters of a given XML file using the (event based) SAX API
     940and prints out the totals.
     943<!-- TABLE
     950File Name               & jaw.xml               & road.gml      & po.xml        & soap.xml \\ \hline   
     951File Type               & document              & data          & data          & data   \\ \hline     
     952File Size (kB)          & 7343                  & 11584         & 76450         & 2717 \\ \hline
     953Markup Item Count       & 74882                 & 280724        & 4634110       & 18004 \\ \hline
     954Markup Density          & 0.13                  & 0.57          & 0.76          & 0.87  \\ \hline
     958\caption{XML Document Characteristics}
     964Table \ref{XMLDocChars} shows the document characteristics of the XML input
     965files selected for the Xerces C++ SAXCount benchmark. The jaw.xml
     966represents document-oriented XML inputs and contains the three-byte and four-byte UTF-8 sequence
     967required for the UTF-8 encoding of Japanese characters. The remaining data files are data-oriented
     968XML documents and consist entirely of single byte encoded ASCII characters.
     971A key predictor of the overall parsing performance of an XML file is markup density\footnote{
     972  Markup Density: the ratio of markup bytes used to define the structure of the document vs. its file size.}.
     973This metric has substantial influence on the performance of traditional recursive descent XML parsers
     974because it directly corresponds to the number of state transitions that occur when parsing a document.
     975We use a mixture of document-oriented and
     976data-oriented XML files to analyze performance over a spectrum
     977of markup densities.
     980Figure \ref{perf_SAX} compares the performance of Xerces, \icXML{} and pipelined \icXML{} in terms of
     981CPU cycles per byte for the SAXCount application.
     982The speedup for \icXML{} over Xerces is 1.3x to 1.8x.
     983With two threads on the multicore machine, \icXMLp{} can achieve speedup up to 2.7x.
     984Xerces is substantially slowed by dense markup
     985but \icXML{} is less affected through a reduction in branches and the use of parallel-processing techniques.
     986\icXMLp{} performs better as markup-density increases because the work performed by each stage is
     987well balanced in this application.
     990<!-- FIGURE
     993\caption{SAXCount Performance Comparison}
     998            </section>
     999                        <section>
     1000                       <title>GML2SVG</title>
     1001                       <para></para>
     1002            </section>
     1003                </section>
    1456          <para>The Catalog of XML Bit Streams presented earlier consist of a set of abstract,
    1457             unbounded bit streams, each in one-to-one correspondence with input bytes of a text
    1458             file. Determining how these bit streams are implemented using fixed-width SIMD
    1459             registers, and possibly processed in fixed-length buffers that represent some multiple
    1460             of the register width is a source of considerable programming complexity. The general
    1461             goal of our compilation strategy in this case is to allow operations to be programmed in
    1462             terms of unbounded bit streams and then automatically reduced to efficient low-level
    1463             code with the application of a systematic code generation strategy for handling block
    1464             and buffer boundary crossing. This is work currently in progress. </para>
     1005                <section>
     1006                        <title>Conclusion and Future Work</title>               
     1008This paper is the first case study documenting the significant
     1009performance benefits that may be realized through the integration
     1010of parallel \bitstream{} technology into existing widely-used software libraries.
     1011In the case of the Xerces-C++ XML parser, the
     1012combined integration of SIMD and multicore parallelism was
     1013shown capable of dramatic producing dramatic increases in
     1014throughput and reductions in branch mispredictions and cache misses.
     1015The modified parser, going under the name \icXML{} is designed
     1016to provide the full functionality of the original Xerces library
     1017with complete compatibility of APIs.  Although substantial
     1018re-engineering was required to realize the
     1019performance potential of parallel technologies, this
     1020is an important case study demonstrating the general
     1021feasibility of these techniques.
     1024The further development of \icXML{} to move beyond 2-stage
     1025pipeline parallelism is ongoing, with realistic prospects for
     1026four reasonably balanced stages within the library.  For
     1027applications such as GML2SVG which are dominated by time
     1028spent on XML parsing, such a multistage pipelined parsing
     1029library should offer substantial benefits. 
     1032The example of XML parsing may be considered prototypical
     1033of finite-state machines applications which have sometimes
     1034been considered ``embarassingly sequential'' and so
     1035difficult to parallelize that ``nothing works.''  So the
     1036case study presented here should be considered an important
     1037data point in making the case that parallelization can
     1038indeed be helpful across a broad array of application types.
     1041To overcome the software engineering challenges in applying
     1042parallel \bitstream{} technology to existing software systems,
     1043it is clear that better library and tool support is needed.
     1044The techniques used in the implementation of \icXML{} and
     1045documented in this paper could well be generalized for
     1046applications in other contexts and automated through
     1047the creation of compiler technology specifically supporting
     1048parallel \bitstream{} programming.
     1050                </section>
    1466       </section>
    1467    </section>
    1469    <section>
    1470       <title>Conclusion</title>
    1471       <para>Parallel bit stream technology offers the opportunity to dramatically speed up the core
    1472          XML processing components used to implement virtually any XML API. Character validation and
    1473          transcoding, whitespace processing, and parsing up to including the full validation of tag
    1474          syntax can be handled fully in parallel using bit stream methods. Bit streams to mark the
    1475          positions of all element names, attribute names and attribute values can also be produced,
    1476          followed by fast bit scan operations to generate position and length values. Beyond bit
    1477          streams, byte-oriented SIMD processing of names and numerals can also accelerate
    1478          performance beyond sequential byte-at-a-time methods. </para>
    1479       <para>Advances in processor architecture are likely to further amplify the performance of
    1480          parallel bit stream technology over traditional byte-at-a-time processing over the next
    1481          decade. Improvements to SIMD register width, register complement and operation format can
    1482          all result in further gains. New SIMD instruction set features such as inductive doubling
    1483          support, parallel extract and deposit instructions, bit interleaving and scatter/gather
    1484          capabilities should also result in significant speed-ups. Leveraging the intraregister
    1485          parallelism of parallel bit stream technology within SIMD registers to take of intrachip
    1486          parallelism on multicore processors should accelerate processing further. </para>
    1487       <para>Technology transfer using a patent-based open-source business model is a further goal of
    1488          our work with a view to widespread deployment of parallel bit stream technology in XML
    1489          processing stacks implementing a variety of APIs. The feasibility of substantial
    1490          performance improvement in replacement of technology implementing existing APIs has been
    1491          demonstrated even in complex software architectures involving delivery of performance
    1492          benefits across the JNI boundary. We are seeking to accelerate these deployment efforts
    1493          both through the development of compiler technology to reliably apply these methods to a
    1494          variety of architectures as well as to identify interested collaborators using open-source
    1495          or commercial models. </para>
    1496    </section>
    14981053   <section>
    14991054      <title>Acknowledgments</title>
    1500       <para>This work is supported in part by research grants and scholarships from the Natural
    1501          Sciences and Engineering Research Council of Canada, the Mathematics of Information
    1502          Technology and Complex Systems Network and the British Columbia Innovation Council. </para>
    1503       <para>We thank our colleague Dan Lin (Linda) for her work in high-performance symbol table
    1504          processing. </para>
     1055      <para></para>
    15051056   </section>
    15071058   <bibliography>
    15081059      <title>Bibliography</title>
Note: See TracChangeset for help on using the changeset viewer.