www/performance.html

   1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
   2           "http://www.w3.org/TR/html4/strict.dtd">
   3 <!-- Material used from: HTML 4.01 specs: http://www.w3.org/TR/html401/ -->
   4 <html>
   5 <head> <META http-equiv="Content-Type" content="text/html; charset=ISO-8859-1">
   6   <title>Polly - Performance</title>
   7   <link type="text/css" rel="stylesheet" href="menu.css">
   8   <link type="text/css" rel="stylesheet" href="content.css">
   9 </head>
  10 <body>
  11 <div id="box">
  12 <!--#include virtual="menu.html.incl"-->
  13 <div id="content">
  14 <h1>Performance</h1>
  15
  16 <p>To evaluate the performance benefits Polly currently provides we compiled the
  17 <a href="http://www.cse.ohio-state.edu/~pouchet/software/polybench/">Polybench
  18 2.0</a> benchmark suite.  Each benchmark was run with double precision floating
  19 point values on an Intel Core Xeon X5670 CPU @ 2.93GHz (12 cores, 24 thread)
  20 system. We used <a href="http://pocc.sf.net">PoCC</a> and the included <a
  21 href="http://pluto-compiler.sf.net">Pluto</a> transformations to optimize the
  22 code. The source code of Polly and LLVM/clang was checked out on
  23 25/03/2011.</p>
  24
  25 <p>The results shown were created fully automatically without manual
  26 interaction. We did not yet spend any time to tune the results. Hence
  27 further improvments may be achieved by tuning the code generated by Polly, the
  28 heuristics used by Pluto or by investigating if more code could be optimized.
  29 As Pluto was never used at such a low level, its heuristics are probably
  30 far from perfect. Another area where we expect larger performance improvements
  31 is the SIMD vector code generation. At the moment, it rarely yields to
  32 performance improvements, as we did not yet include vectorization in our
  33 heuristics. By changing this we should be able to significantly increase the
  34 number of test cases that show improvements.</p>
  35
  36 <p>The polybench test suite contains computation kernels from linear algebra
  37 routines, stencil computations, image processing and data mining. Polly
  38 recognices the majority of them and is able to show good speedup. However,
  39 to show similar speedup on larger examples like the SPEC CPU benchmarks Polly
  40 still misses support for integer casts, variable-sized multi-dimensional arrays
  41 and probably several other construts. This support is necessary as such
  42 constructs appear in larger programs, but not in our limited test suite.
  43
  44 <h2> Sequential runs</h2>
  45
  46 For the sequential runs we used Polly to create a program structure that is
  47 optimized for data-locality. One of the major optimizations performed is tiling.
  48 The speedups shown are without the use of any multi-core parallelism. No
  49 additional hardware is used, but the single available core is used more
  50 efficiently.
  51 <h3> Small data size</h3>
  52 <img src="images/performance/sequential-small.png" /><br />
  53 <h3> Large data size</h3>
  54 <img src="images/performance/sequential-large.png" />
  55 <h2> Parallel runs</h2>
  56 For the parallel runs we used Polly to expose parallelism and to add calls to an
  57 OpenMP runtime library. With OpenMP we can use all 12 hardware cores
  58 instead of the single core that was used before. We can see that in several
  59 cases we obtain more than linear speedup. This additional speedup is due to
  60 improved data-locality.
  61 <h3> Small data size</h3>
  62 <img src="images/performance/parallel-small.png" /><br />
  63 <h3> Large data size</h3>
  64 <img src="images/performance/parallel-large.png" />
  65 </div>
  66 </div>
  67 </body>
  68 </html>