Author: Eugene Nalimov
Date: 11:29:55 03/19/03
Go up one level in this thread
I would not expect that going from 12-way associative to 6-way associative cache will slow application so much. There is simpler explanation: I just looked at Crafty's executable, and it contains ~1.6Mb of static data. Of course part of that data are strings and not really used in the search, but vast majority of data is used -- attack tables, bitboard tables, masks, etc. If you'll add to that code and dynamic data, you'll end up with working set that is larger than 1.5Mb. With 3Mb of L3 cache Crafty fits entirely into it (either without hash tables, or with small default hash tables), so latency of each and every non-hash-table access went down from ~150 CPU cycles to ~12 CUP cycles. Thanks, Eugene On March 18, 2003 at 20:33:46, Tom Kerrigan wrote: >Interesting case of seemingly conflicting data... > >Maybe by turning off half the chip's cache, it decreased the associativity >enough to cause contention for some critical data. That's the best explanation I >can think of. > >Nice benchmarking, Aaron. :) > >-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.