Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Bigger hash really better?

Author: Robert Hyatt

Date: 09:15:37 12/19/03

Go up one level in this thread


On December 19, 2003 at 10:52:57, Anthony Cozzie wrote:

>On December 19, 2003 at 10:41:27, Robert Hyatt wrote:
>
>>On December 19, 2003 at 01:21:45, Jouni Uski wrote:
>>
>>>After installing more RAM to my Pentium 2,4 GHz I tested Fritz8 in some test
>>>suites with 128MB and 384MB hash (time limit was 10 minutes and positions quite
>>>hard = average solution times around 3-5 minutes): to my surprise
>>>average solution time was shorter with 128MB! Why? Absolutely no hard disk
>>>swapping with 512MB total RAM!
>>>
>>>Jouni
>>
>>
>>Hashing is based on random numbers.  Which means the result of using them
>>is going to have a bit of randomness as well.  Sometimes bigger hash slows
>>a program down, because it makes the search more accurate, which might make
>>it a bit slower.  But accuracy and speed don't necessarily match up so it is
>>not easy to say "this is slower, so it is worse."
>>
>>Also, there are hardware considerations.  The size of the TLB for example.
>>If you blow that out, you make your memory access time go from maybe 150ns
>>to 3X that.  Since hash tables are addressed randomly, this is a real
>>possibility.  IE the opteron I was using earlier this week has just over
>>1000 TLB entries.  That lets me address 1000 * 4K very quickly.  Anything
>>beyond that sees slower memory access times.
>
>there is a patch for 2.6 that will automatically use the large pages.  that plus
>prefetching should give quite a speedup to crafty . . .
>
>anthony

I know.  I have mentioned that in the past already.  However, on my xeon, the
usual memory benchmark program only reports something like 60 TLB entries,
which seems low (lmbench).  I have not tried 2.6 on my machine, as it isn't
just a matter of grabbing and building the kernel, there are a bunch of other
things including binutils, and the like, that also have to be upgraded.  I hope
someone will do a 2.6 distribution before long and I'll dive in.

the 4M page size sounds pretty good, from a memory access point of view (not
thrashing the TLB so badly).




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.