Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Some opteron results for Crafty

Author: Tom Likens

Date: 14:48:34 11/26/03

Go up one level in this thread


On November 26, 2003 at 16:08:50, Robert Hyatt wrote:

>On November 26, 2003 at 15:36:58, Dann Corbit wrote:
>
>>On November 26, 2003 at 15:25:10, Robert Hyatt wrote:
>>
>>>I have been working both with Eugene and AMD.  The following bench run is
>>>on a quad 1.8ghz opteron, 8 gigs of ram.  The only "option" I have set is
>>>"mt=4".  There is _no_ assembly code in this version, pure C only.  I am
>>>looking at updating the asm to 64 bit but that will take some time and
>>>studying.
>>>
>>>Meanwhile:
>>>
>>>Crafty v19.6 (1 cpus)
>>>
>>>White(1): mt=4
>>>max threads set to 4
>>>White(1): bench
>>>Running benchmark. . .
>>>......
>>>Total nodes: 105863114
>>>Raw nodes per second: 5881284
>>>Total elapsed time: 18
>>>SMP time-to-ply measurement: 35.555556
>>>
>>>This is using gcc, although I am not sure whether it is producing 64 bit
>>>or 32 bit code at the moment.  However, 5.8M nps is not bad.  About 1M less
>>>than Eugene's MSVC numbers.  I will look into the 64 bit stuff more to see if
>>>gcc is producing real opteron assembly or not...  And I will study the
>>>PGO options although the list time I tried them on GCC the compiler promptly
>>>crashed. :)
>>>
>>>Note that the above is with default hash and everything, no endgame tables,
>>>no opening book, etc...
>>
>>Could we see the numbers for 1,2,3 threads active also?
>>I would be interested to see how it scales.
>
>
>Sure.
>
>one processor:
>
>White(1): bench
>Running benchmark. . .
>......
>Total nodes: 100409437
>Raw nodes per second: 1498648
>Total elapsed time: 67
>SMP time-to-ply measurement: 9.552239
>
>
>two processors:
>
>max threads set to 2
>White(1): bench
>Running benchmark. . .
>......
>Total nodes: 99562452
>Raw nodes per second: 3017044
>Total elapsed time: 33
>SMP time-to-ply measurement: 19.393939
>
>three processors:
>
>max threads set to 3
>White(1): bench
>Running benchmark. . .
>......
>Total nodes: 102543114
>Raw nodes per second: 4458396
>Total elapsed time: 23
>SMP time-to-ply measurement: 27.826087
>
>four processors:
>
>max threads set to 4
>White(1): bench
>Running benchmark. . .
>......
>Total nodes: 102606915
>Raw nodes per second: 5700384
>Total elapsed time: 18
>SMP time-to-ply measurement: 35.555556
>
>
>Let me note here that this is not a very NUMA-aware implementation,
>nowhere near as good as what we did (Eugene and I) for windows.  I
>am going to look at the Linux NUMA library tonight and work on getting
>some of those features in, which should further push performance up.
>
>This is way better than 19.4, but it is not "all there" yet.  Note also
>that there is no assembly language of any kind in this version, it is pure
>C.  I plan on rectifying that _soon_.  :)

Sometimes Bob, you not only lift the bar for the rest of us, but you put it
in orbit ;-)

I was thinking about entering CCT6 but I'm not sure there's much point!

BTW, I finally received the AMD FX-51 and my preliminary tests under Windows
XP Pro (I'm loading Linux as I type) gives Djinn a (roughly) 4x speedup.
Unlike your test, I am including 32-bit inline assembly, but no 64-bit
assembly which should boost things nicely.  I also intend to use profile
guided optimizations after I get Linux set up to see how that improves
things (hopefully, quite a bit since the Windows version was compiled
specifically for a P4 system).

One caveat, so far I haven't been able to get the 64-bit version of SUSE 9.0
to recognize my SATA hard-drives or my Promise controller.  The 32-bit
version *does* recognize the components so that's what I'm loading to get
Linux on it initially.  It's not too bad since I have my home directories
mounted on another machine via NFS, and intend to load the 64-bit version
of the OS when it works.

More info as it becomes available.

regards,
--tom




This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.