Re: [gnugo-devel] Owl persistent cache

[Trevor Morris]

At 03:44 PM 2/28/2002 -0800, Daniel Bump wrote:
>
>> This makes a lot of sense.  It's right around the time I commented
>> (in hindsight not strongly enough) that I'd noticed a big
>> performance decrease:
>> http://mail.gnu.org/pipermail/gnugo-devel/2001-November/000372.html
>
>This performance improvement will of course show up more in games
>than in the regressions, since it consists of not unncecesarily
>repeating calculations which were made on previous moves. So it
>seems unlikely that this bug was the cause of those slow
>moves. It would be nice to know if those tests are still slow at
>this date and if so we should look at them.

Here are the CPU times reported by www.public32.com on the original
list of slow problems from neurogo.tst.  All except one show
a huge improvement.  Notice that several have big improvements after
the OWL caching correction patch (put in after the run of 2/28).

   #   11/1/01   2/28/02    3/1/02   11/1 to 3/1 % change
   28  68.468 -> 44.781  -> 44.797   34.6%
   12  67.266 -> 38.657  -> 29.125   56.7%
   8   51.641 -> 41.437  -> 31.484   39.0%
   9   40.734 -> 28.532  -> 27.594   32.3%
   11  36.078 -> 64.796  -> 55.735   -54.5
   14  35.969 -> 30.203  -> 17.969   50.0%

Total 300.156 ->248.406  ->206.704   31.1%


For what it's worth, here are the regression run times, too:

3/1/02 times:
Fri 03/01/2002  8:15:59.73
Fri 03/01/2002 11:03:20.34
elapsed: 2:47:21

2/28/02 times:
Thu 02/28/2002  8:46:22.96
Thu 02/28/2002 11:49:41.04
elapsed: 3:03:18

8.7% increase (probably due to gains in game-based problem sets such 
as neurogo.tst where caching actually helps).

-Trevor