#631 TIP #285 perf changes

TIP Implementation
open
5
2011-01-30
2011-01-30
No

These changes are designed to optimize (and modularize) the TIP #285 features.

Discussion

  • Joe Mistachkin

    Joe Mistachkin - 2011-01-30

    TIP 285 rev2-v2

     
  • Alexandre Ferrieux

    pstack sampling

     
  • Alexandre Ferrieux

    Attaching pstack sampling results.

    pid5971 is 8.6HEAD (Jan25)
    pid12410 has your TIP285rev2 patch

    both are from a 10-mn run over idiotloop2.tcl

     
  • Alexandre Ferrieux

    Timings on NRE:

    23:50] ferrieux on one specific test, it is impressive:
    head285.res:647968 microseconds per iteration
    head285.res:646554 microseconds per iteration
    head285.res:626336 microseconds per iteration
    head285.res:642746 microseconds per iteration

    head.res:733830 microseconds per iteration
    head.res:711075 microseconds per iteration
    head.res:718664 microseconds per iteration
    head.res:702698 microseconds per iteration
    [23:50] ferrieux first with patch, second without. 4 runs.

     
  • Alexandre Ferrieux

    Timings on non-NRE variant.

    01:28] ferrieux timing results for 285 patch on 86-without-nre:
    sansnre285.res:565905 microseconds per iteration
    sansnre285.res:572837 microseconds per iteration
    sansnre285.res:569185 microseconds per iteration
    sansnre285.res:568799 microseconds per iteration
    sansnre.res:618944 microseconds per iteration
    sansnre.res:619809 microseconds per iteration
    sansnre.res:615293 microseconds per iteration
    sansnre.res:622320 microseconds per iteration
    [01:28] mistachkin ferrieux : now, it's "real fast"
    [01:29] ferrieux have not computed the ratios, but yeah, that sounds orthogonal to nre

     
  • Alexandre Ferrieux

    Bottom line: should be confirmed with more extensive tclbench runs, but it superficially looks like Joe's patch are an absolute Good Thing, and are orthogonal to the NRE slowdown.

     

Get latest updates about Open Source Projects, Conferences and News.

Sign up for the SourceForge newsletter:





No, thanks