My nightly run of pybench went up from the usual 7590ms per run to around 8200ms between Monday night and today. Can anyone explain this ? Thanks, -- Marc-Andre Lemburg eGenix.com Professional Python Software directly from the Source (#1, Feb 27 2003)
Python/Zope Products & Consulting ... http://www.egenix.com/ mxODBC, mxDateTime, mxTextTools ... http://python.egenix.com/
Python UK 2003, Oxford: 33 days left EuroPython 2003, Charleroi, Belgium: 117 days left
My nightly run of pybench went up from the usual 7590ms per run to around 8200ms between Monday night and today. Can anyone explain this ?
Did you compare the outputs? There could be importants hint there! :-) Maybe this checkin to ceval.c backfired? ---------------------------- revision 2.351 date: 2003/02/26 18:11:50; author: rhettinger; state: Exp; lines: +18 -4 Micro-optimizations. * List/Tuple checkexact is faster for the common case. * Testing for Py_True and Py_False can be inlined for faster looping. ---------------------------- --Guido van Rossum (home page: http://www.python.org/~guido/)
Guido van Rossum wrote:
My nightly run of pybench went up from the usual 7590ms per run to around 8200ms between Monday night and today. Can anyone explain this ?
Did you compare the outputs? There could be importants hint there! :-)
Unfortunately, the cronjob always uses the same file to store the benchmark data (I've changed that now, to be able to check on a day-per-day basis). Is there a way to check out a Python CVS version dated 2003-02-24 ? I could then give that a try again.
Maybe this checkin to ceval.c backfired?
---------------------------- revision 2.351 date: 2003/02/26 18:11:50; author: rhettinger; state: Exp; lines: +18 -4 Micro-optimizations. * List/Tuple checkexact is faster for the common case. * Testing for Py_True and Py_False can be inlined for faster looping. ----------------------------
Could be. I don't think inlining in ceval will help. The switch is too long already and it's much better to let the compiler/processor decide which parts should really get inlined since they have more knowledge about the platform than we do (sometimes :-). -- Marc-Andre Lemburg eGenix.com Professional Python Software directly from the Source (#1, Feb 27 2003)
Python/Zope Products & Consulting ... http://www.egenix.com/ mxODBC, mxDateTime, mxTextTools ... http://python.egenix.com/
Python UK 2003, Oxford: 33 days left EuroPython 2003, Charleroi, Belgium: 117 days left
[M.-A. Lemburg]
... Is there a way to check out a Python CVS version dated 2003-02-24 ? I could then give that a try again.
Use the -D option to the checkout command: -D date Check out revisions as of date. (implies -P) (is sticky) It's highly creative in what it accepts for "a date" <wink>.
Sjoerd Mullender wrote:
On Thu, Feb 27 2003 "M.-A. Lemburg" wrote:
Is there a way to check out a Python CVS version dated 2003-02-24 ? I could then give that a try again.
cvs update -D '24 Feb 2003'
Thanks ! -- Marc-Andre Lemburg eGenix.com Professional Python Software directly from the Source (#1, Feb 27 2003)
Python/Zope Products & Consulting ... http://www.egenix.com/ mxODBC, mxDateTime, mxTextTools ... http://python.egenix.com/
Python UK 2003, Oxford: 33 days left EuroPython 2003, Charleroi, Belgium: 117 days left
My nightly run of pybench went up from the usual 7590ms per run to around 8200ms between Monday night and today. Can anyone explain this ?
Did you compare the outputs? There could be importants hint there! :-)
Maybe this checkin to ceval.c backfired?
---------------------------- revision 2.351 date: 2003/02/26 18:11:50; author: rhettinger; state: Exp; lines: +18 -4 Micro-optimizations. * List/Tuple checkexact is faster for the common case. * Testing for Py_True and Py_False can be inlined for faster looping. ----------------------------
That would be a bummer. It looked like a pure win. I timed the changes with PyStone with N=200000 and got: Base Code: 12.379 12.2716 12.2716 + CheckExact: 12.2746 12.2278 12.3165 + T/F inlined: 12.108 12.1697 12.1259 Since PyStone does weird things to booleans and uses "while 1", I ran a separate test on "while 1" vs. "while True" showed a speedup of "while True" and no measurable change to "while 1". If your timings show a different result, let me know and I'll back it out. Raymond Hettinger
Raymond Hettinger wrote:
My nightly run of pybench went up from the usual 7590ms per run to around 8200ms between Monday night and today. Can anyone explain this ?
Did you compare the outputs? There could be importants hint there! :-)
Maybe this checkin to ceval.c backfired?
---------------------------- revision 2.351 date: 2003/02/26 18:11:50; author: rhettinger; state: Exp; lines: +18 -4 Micro-optimizations. * List/Tuple checkexact is faster for the common case. * Testing for Py_True and Py_False can be inlined for faster looping. ----------------------------
That would be a bummer. It looked like a pure win.
I timed the changes with PyStone with N=200000 and got: Base Code: 12.379 12.2716 12.2716 + CheckExact: 12.2746 12.2278 12.3165 + T/F inlined: 12.108 12.1697 12.1259
Since PyStone does weird things to booleans and uses "while 1", I ran a separate test on "while 1" vs. "while True" showed a speedup of "while True" and no measurable change to "while 1".
If your timings show a different result, let me know and I'll back it out.
Ok, I reran the tests using the CVS time machine and the results indicate that this was a false alarm. I am not exactly sure what is causing it though: both figures for the Monday checkout and todays checkout hover around 8200ms per run on average, while most historical figures from last month are in the 7590-7950ms range. Oh, BTW, your patch does not result in any significant change in pybench results. Nevermind, -- Marc-Andre Lemburg eGenix.com Professional Python Software directly from the Source (#1, Feb 27 2003)
Python/Zope Products & Consulting ... http://www.egenix.com/ mxODBC, mxDateTime, mxTextTools ... http://python.egenix.com/
Python UK 2003, Oxford: 33 days left EuroPython 2003, Charleroi, Belgium: 117 days left
participants (5)
-
Guido van Rossum -
M.-A. Lemburg -
Raymond Hettinger -
Sjoerd Mullender -
Tim Peters