Are my fuego parameters ok?
> opponent_settings2='uct_param_player ignore_clock 1\nuct_param_player
> max_games '+str(playouts)+'\nuct_param_player resign_min_games 5000
> \nuct_param_search number_threads 8\nuct_max_memory 8000000000
> \nuct_param_player reuse_subtree 1'
Hi Detlef,
looks OK generally. Some comments:
In my normal tests I set
uct_param_search number_threads 1
instead of 8, then I run several test games in parallel with the —threads
option in gogui-twogtp.
The reason for single-threaded is that I want to separate “baseline” strength
comparison from parallel scaling performance. So I want to keep the baseline as
simple as possible.
Of course then you need to reduce the memory per player as well. I do not know
how much weaker the current Fuego is on 8 threads, as compared to running
single-threaded 8 times longer. It totally depends on your hardware too. You
can easily measure the raw parallel speedup by using the tool
fuego/tools/misc/fuego-speed-test.pl
E.g. 6.5 for 8 threads is not bad in terms of nodes/sec. speedup. But strength
speedup is less because the parallel algorithm is less informed than the
sequential baseline.
Talking of memory, using 8000000000 on a 8GB machine may be a bit too much,
depending on what else runs on your computer and how large the memory footprint
of the rest of Fuego is. You don’t want it to start swapping. We have used
something like 7.3GB for the tree which is probably conservative. There is a
patch coming up which will reduce the memory footprint of Greenpeep style
patterns a bit.
Then of course, once in a while we need to check parallel scaling as well, as
Markus did for the paper we wrote a long time ago.
The other thing I do is I set
go_rules cgos
This keeps things simpler for scoring in the end. Other rules support in Fuego
is a bit flaky. (I am not sure what is the default if you specify nothing)
And I still use GnuGo as a referee, there are always some games where the
programs disagree in the end. Those end up in the “future work” bucket :)
Martin_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
