Re: [gnugo-devel] Default reading cache size / tiny benchmark

[alain Baeckeroot]

Le Lundi 16 Janvier 2006 21:43, Gunnar Farneb�ck a écrit :
> I don't think there has been any systematic investigations of the
> effects of cache size after the reading cache was completely
> reimplemented around 3.5.5. It is possible that a significantly
> smaller cache than before would be adequate as default. Presumably a
> large cache is more useful in games which require much reading. You
> could for example try to replay regression/games/FSF-neurogo.sgf and
> increase the level to 13 or so, and see whether that behaves
> differently.
> 

i did some runs with different cache size, on this game at level 13 , 
on the "terrible ticket 10 game" and on other sample pro games from gobase.org 
for comparison

Probably nothing new for you, or trivial things but here are the results:
-M1 seems to give same speed as default -M8 

The worst case is at level 13, 10% slower with -M1 , for the "ticket 10" game.
So default to 8M seems largely enough.

my 2 cents :)
Alain

===================================================================
gnugo 3.7.7 compiled with gcc-3.2.3 CFLAGS="-O2 -march=athlon-xp"
        (athlon-xp 2200+, =1600 real MHz, 512 kB cache)
        gcc-4.0.2 seems 5% to 10% _slower_   (google confirm that)
===================================================================
Reference with default params, replay gnugo
 - 40 min "terrible" ticket 10 game             (opponent _much_ stronger than 
gnugo)
 - 22 min regression/games/FSF-neurogo.sgf (opponent weaker, but chaotic)

replay one player (games from gobase.org)
 - 17 min game166 Kato /Takemiya (kato kill huge group)
 - 15 min Takagawa Kaku / Sakata Eio (huge furikawari) 
 - 7 min game147  Ishida / Takemiya     (cool game, nearly understandable)

===================================================================
Reference Neurogo game, level 13 , ordered by decreasing cache size:
------------------------------------------------------------------------------------------

time gnugo --replay white -M128 --level 13 regression/games/FSF-neurogo.sgf
    
Global score: 1905.30 / 5571.52
real    68m54.705s
user    66m46.482s
sys     0m4.824s
---------------------------------------
 time gnugo --replay white -M8 --level 13 regression/games/FSF-neurogo.sgf  

Global score: 1913.73 / 5579.94
real    68m1.381s
user    66m3.980s
sys     0m4.964s
---------------------------------------
time gnugo --replay white -M1 --level 13 regression/games/FSF-neurogo.sgf    

Global score: 1913.46 / 5580.64
real    69m29.165s
user    67m20.877s
sys     0m4.588s
---------------------------------------
time gnugo --replay white -M0.5 --level 13 regression/games/FSF-neurogo.sgf    

Global score: 1920.08 / 5579.41
real    70m22.111s
user    68m51.198s
sys     0m4.592s
---------------------------------------
time gnugo --replay white -M0.1 --level 13 regression/games/FSF-neurogo.sgf   

Global score: 1943.98 / 5578.18
real    80m3.740s
user    77m18.018s
sys     0m6.316s
---------------------------------------
time gnugo --replay white -M0.01 --level 13 regression/games/FSF-neurogo.sgf   

Global score: 1934.58 / 5565.24
real    111m24.070s
user    106m19.267s
sys     0m11.113s

===================================================================
Reference "ticket 10 terrible game" level 13:
----------------------------------------------------------------

time gnugo --replay black -M10 --level 
13 /Big/GO/new-tests/terrible-18584.0.9handicap.sgf

Global score: 2939.84 / 13866.69
real    104m55.099s
user    101m18.448s
sys     0m8.637s
---------------------------------------
time gnugo --replay black -M1 --level 
13 /Big/GO/new-tests/terrible-18584.0.9handicap.sgf
Global score: 2956.42 / 13810.38
real    113m7.724s
user    108m27.259s
sys     0m8.501s
===================================================================