Changeset 1371 for docs/HPCA2012


Ignore:
Timestamp:
Aug 24, 2011, 5:52:24 PM (8 years ago)
Author:
lindanl
Message:

modifications

Location:
docs/HPCA2012
Files:
2 edited

Legend:

Unmodified
Added
Removed
  • docs/HPCA2012/05-corei3.tex

    r1370 r1371  
    1919mispredictions. Parabix's data reorganization significantly improves
    2020the overall cache miss rate. We experience 7$\times$ less misses than
    21 Expat and 25 $\times$ less misses than Xerces at the L1 and 104$\times$ less misses than
    22 Expat and 15 $\times$ less misses than Xerces at the L2 level. The improved cache
     21Expat and 25$\times$ less misses than Xerces at the L1 and 104$\times$ less misses than
     22Expat and 15$\times$ less misses than Xerces at the L2 level. The improved cache
    2323utilization keeps the SIMD units busy and prevent memory related
    2424stalls. Note that cache misses also cause increased application energy
     
    6464eliminates many branches. Further optimizations take advantage of
    6565Parabix's data organization and replace condition branches with {\em
    66   bit scan} operations that can process up to 64 characters worth of
     66  bit scan} operations that can process up to 128 characters worth of
    6767branches with one operation. In many cases, we also replace the
    6868branches with logical predicate operations. Our predicate are cheaper
     
    7171 As shown in Figure \ref{corei3_BR},
    7272Parabix processing is almost branch free. Parabix exhibits minimal
    73 dependence on source XML markup density; it experiences a constant
    74 number of branch mispredictions irrespective of the input. The cost of
     73dependence on source XML markup density; it experiences between 19.5 and
     7430.7 branch mispredictions per thousand of XML byte. The cost of
    7575branch mispredictions for the Expat parser can be over 7 cycles per
    7676XML byte (see Figure \ref{corei3_BM}) ---this cost alone is higher
Note: See TracChangeset for help on using the changeset viewer.