Re: [Bug-gnubg] Performance Gains from 20090915 to 20110525 Codebases

[Ian Shaw]

Cool! A nice comparison, and congratulations, Phillipe.

-----Original Message-----
From: address@hidden [mailto:address@hidden On Behalf Of Michael Petch
Sent: 27 May 2011 06:09
To: address@hidden
Subject: [Bug-gnubg] Performance Gains from 20090915 to 20110525 Codebases

Hi All,

I had been compiling a bunch of statistics lately, and happened to have
built two 32Bit version of GNUBG under Debian Squeeze (6.0) Stable. I
used 32 Bit since the older one didn't compile cleanly for the older
version. I pretty much used full optimizations and SSE2, Threading on.
GTK was built in but the test were run in terminal mode.

Hardware: 2xXenon 5405 (4 cores each, 12MB L2 Cache, No HT), 2.00 GHZ,
10 GB DDR2-6400 RAM - 8 Cores total.

Debian 6.0 Kernel 2.6.32-5-amd64 x86_64
GCC version 4.4.5-8 .

Configure Options:
CFLAGS="-O3 -funsafe-loop-optimizations -funsafe-math-optimizations
-ffast-math -freciprocal-math -ftree-vectorize -mfpmath=sse  -mssse3
-msse3 -msse -msse2 -fomit-frame-pointer -msahf" ./configure
--enable-threads --enable-sse=sse2

The attached graph (also found at
http://www.capp-sysware.com/analysis/studies/20090915to20110525PerfGain.png
) is basically a plot of the Cache and Threads vs the Performance gain
from the Old 20090915 code to the 20090525 code. I haven't run it
against a rollout (or batch of them yet).

I believe Philippe Michel mentioned a 30-40% gain. I'm seeing 36%-40%
depending on the variables but 38% would be about the average. I believe
most of Philippe's changes provide the bulk of the performance increases.

-- 

Michael Petch
CApp::Sysware Consulting Ltd.
OpenPGP FingerPrint=D81C 6A0D 987E 7DA5 3219 6715 466A 2ACE 5CAE 3304