source: docs/Working/icGrep/evaluation.tex @ 4475

Last change on this file since 4475 was 4475, checked in by cameron, 4 years ago

Discussion of performance results

File size: 7.8 KB
Line 
1\section{Evaluation}\label{sec:evaluation}
2
3In this section, we report on the evaluation of ICgrep performance, looking
4at three aspects.   First we consider a performance studies in a series
5of Unicode regular expression search problems in comparison to the
6contemporary competitors, including pcre2grep released in January 2015
7and ugrep of the ICU 54.1 software distribution.  Then we move on to
8investigate some performance aspects of ICgrep internal methods, looking
9at the impact of optimizations and multithreading.
10
11\subsection{ICgrep vs. Contemporary Competitors}
12
13A key feature of Unicode level 1 support in regular expression engines
14is how the support that they provide for property expressions and combinations of property expressions
15using set union, intersection and difference operators.   Both {\tt ugrep}
16and {\tt icgrep} provide systematic support for all property expressions
17at Unicode Level 1 as well as set union, intersection and difference.
18On the other hand, {\tt pcre2grep} does not support the set intersection and difference operators directly.
19However, these operators can instead be expressed using a regular expression
20feature known as a lookbehind assertion.   Set intersection involves a
21regular expression formed with a one of the property expressions and a
22positive lookbehind assertion on the other, while set difference uses
23a negative lookbehind assertion.  As all three programs support lookbehind
24assertions in this way, we systematically generated set intersection and
25difference in this way.
26
27We generated a set of regular expressions involving all Unicode values of
28the Unicode general
29category property ({\tt gc}) and all values of the Unicode script property ({\tt sc}).  We then generated
30expressions involving random pairs of {\tt gc} and {\tt sc}
31values combined with a random set operator chosen from union, intersection and difference.
32All property values are represented at least once.   A small number of
33expressions were removed because they involved properties not supported by pcre2grep.
34In the end 246 test expressions were constructed in this process.
35
36We selected a set of Wikimedia XML files in several major languages representing
37most of the world's major language families as a test corpus.   For each program
38under test, we perform searches for each regular expression against each XML document.
39Results are presented in \Figure \ref{fig:property_test}.  Performance is reported
40in CPU cycles per byte on an Intel Core i7 machine.   The results were grouped
41by the percentage of matching lines found in the XML document, grouped in
425\% increments.  ICgrep shows dramatically better performance, particularly
43when searching for rare items.
44As shown in the figure, pcre2grep and ugrep both show
45increased performance (reduced CPU cycles per byte) with increasing percentage
46of matches found.  In essence, each match found allows these programs
47to skip the full processing of the rest of the line.   On the other
48hand, icGrep shows a slight drop-off in performance with the number
49of matches found.   This is primarily due to property classes that
50include large numbers of codepoints.   These classes require more
51bitstream equations for calculation and also have a greater probability
52of matching.   Nevertheless, the performance of icGrep in matching
53the defined property expressions is stable and well ahead of the competitors
54in all cases.
55
56\begin{figure}
57\begin{center}
58\pgfplotstableread[col sep = comma]{data/icgrep-scatter.csv}\icgrep
59\pgfplotstableread[col sep = comma]{data/ugrep541-scatter.csv}\ugrep
60\pgfplotstableread[col sep = comma]{data/pcre2grep-scatter.csv}\pcre
61
62\begin{tikzpicture}
63\begin{axis}[
64grid=both,
65x tick label style={ /pgf/number format/1000 sep=},
66% x buffer=sort,
67ylabel={CPU Cycles Per Byte},
68xlabel={Percentage of Matching Lines},
69minor y tick num={1},
70xmax=100
71%ymax=30000000
72]
73\addplot+[no markers,line width=2pt,color=blue!60,solid] table {\icgrep};
74\addplot+[no markers,line width=2pt,color=red!60,solid] table {\ugrep};
75\addplot+[no markers,line width=2pt,color=brown!60,solid] table {\pcre};
76\legend{icGrep,ugrep541,pcre2grep}
77\end{axis}
78
79
80\end{tikzpicture}
81\end{center}
82\caption{Comparative Matching Performance}\label{fig:property_test}
83\end{figure}
84
85% \begin{figure}
86% \pgfplotstableread[col sep = comma]{data/icgrep-cp-scatter.csv}\icgrep
87% \pgfplotstableread[col sep = comma]{data/icgrep-flat-cp-scatter.csv}\icgrepf
88%
89%
90% \begin{tikzpicture}
91% \begin{semilogxaxis}[
92% grid=y,
93% x tick label style={ /pgf/number format/1000 sep=},
94% % x buffer=sort,
95% ylabel={Cycles Per Byte},
96% xlabel={Match Percentage},
97% minor y tick num={1},
98% %xmax=100
99% %ymax=30000000
100% ]
101% \addplot+[no markers,line width=2pt,color=blue!60,solid] table {\icgrep};
102% \addplot+[no markers,line width=2pt,color=red!60,solid] table {\icgrepf};
103%
104% \end{semilogxaxis}
105% \end{tikzpicture}
106%
107% \end{figure}
108
109
110\subsection{Optimizations of Bitwise Methods}
111
112In order to support evaluation of bitwise methods, as well as to support
113the teaching of those methods and ongoing research, \icGrep{} has an array
114of command-line options.   This makes it relatively straightforward
115to report on certain performance aspects of ICgrep, while others require
116special builds. 
117
118For example, the command-line switch {\tt -disable-matchstar} can be used
119to eliminate the use of the MatchStar operation for handling
120Kleene-* repetition of character classes.   In this case, \icGrep{} substitutes
121a while loop that iteratively extends match results.   
122Surprisingly, this
123does not change performance much in many practical cases.   
124In each block,
125the maximum iteration count is the maximum length run encountered; the
126overall performance is based on the average of these maximums throughout the
127file.   But when search for XML tags using the regular expression
128\verb:<[^!?][^>]*>:, a slowdown of more than 2X may be found in files
129with many long tags. 
130
131The {\tt -disable-log2-bounded-repetition} flag allows these
132effectiveness of the special techniques for bounded repetition of
133byte classes to be assessed.   A slowdown of 30\% was observed
134with the searches using the regular expression
135\verb:(^|[ ])[a-zA-Z]{11,33}([.!? ]|$):, for example.
136
137To assess the effectiveness of inserting if-statements, the
138number of non-nullable pattern elements between the if-tests
139can be set with the {\tt -if-insertion-gap=} option.   The
140default value in \icGrep{} is 3, setting the gap to 100 effectively
141turns of if-insertion.   Eliminating if-insertion sometimes improves
142performance by avoiding the extra if tests and branch mispredications.
143For patterns with long strings, however, there can be a substantial
144slowdown; searching for a pattern of length 40 slows down by more
145than 50\% without the if-statement short-circuiting.
146
147ICgrep also provides options that allow
148various internal representations to be printed out.   These
149can aid in understanding and/or debugging performance issues.
150For example, the option
151{\tt -print-REs} show the parsed regular expression as it goes
152through various transformations.   The internal Pablo code generated
153may be displayed with {\tt -print-pablo}.  This can be quite useful in
154helping understand the match process.   It also possible to print out the
155generated LLVM IR code ({\tt -dump-generated-IR}), but this may be
156less useful as it includes many
157details of low-level carry-handling that obscures the core logic.
158
159The precompiled calculations of the various Unicode properties
160are each placed in if-hierarchies as described previously.   To assess the
161impact of this strategy, we built a version of icGrep without such
162if-hierarchies.  In this case, when a Unicode property class is defined,
163bitwise logic equations are applied for all members of the class independent
164of the Unicode blocks represented in the input document.   For the classes
165covering the largest numbers of codepoints, we observed slowdowns of up to 5X.
166
167
168
169\subsection{Single vs. Multithreaded Performance}
Note: See TracBrowser for help on using the repository browser.