[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] Timing data
From: |
Arend Bayer |
Subject: |
Re: [gnugo-devel] Timing data |
Date: |
Mon, 25 Feb 2002 01:37:25 +0100 (CET) |
A few remarks:
> All regressions, current CVS:
>
> alt and exp conn and exp influence: 29108 seconds
> alt and exp conn : 27911 seconds
>
> The penalty for the experimental influence seems to be about 4.2%
I have just finished a different timing test of --experimental-influence vs
--standard-influence via a replay of games. Since the persistent caching
is of no use in the regression test setting, I might have expected a
different result. However, I also got exactly 4.2% time cost for
experimental mode, varying between 1.2% and 6.7% for different games
(details below).
> I have more specific timing information for strategy.tst alone.
> Inge pointed out that I should really use time to get the exact
> CPU time but since comment came after these experiments
> were already underway and I didn't want to switch my methodology
> in the middle. Next time I'll do it Inge's way.
Incidentally, I had used a script that reported both the 'time' output
and measured the time itself by reporting the date before and after the
test. In my setting (abolutely minimalistic Linux - no X, hardly any
drivers, just one bash), I did not get any differences. In theory, one
could use 'time' to do performance tests with GNU Go running in
background. However, I personnaly would not safely assume that this works
very reliably for a program like GNU Go.
Btw, I also did all replays twice, to measure the accuracy of this test
method. The variance I got was of magnitude ~0.2%, very reasonable.
>
> And perhaps we should think about releasing 3.2 soon.
> The question is, if we put out a version which is 45%
> slower than 3.0, is that acceptable?
The speed of PCs still gets doubled every 15-18 months. I'd say
we could allow GNU Go to get slower at the same rate. Then 45% would not
be too far off.
OTOH, I think there are still some speed-ups we can do in the short term.
Arend
Resulst:
stand infl exper. infl diff abs. diff in %
911s 937s 26s 2.8%
651s 689s 37s 5.8%
1038s 1062 24s 2.3%
951s 1007s 56s 5.9%
2363s 2467s 104s 4.4%
1613s 1718s 105s 6.5%
2938s 2975s 37s 1.2%
1572s 1678s 106s 6.7%
752s 796s 44s 5.8%
Sum:
12789s 13329s 4.2%