Author: Robert Hyatt
Date: 20:06:29 01/02/02
Go up one level in this thread
On January 02, 2002 at 16:21:43, David Rasmussen wrote: >On January 02, 2002 at 16:17:31, Robert Hyatt wrote: > >>On January 02, 2002 at 15:41:41, Peter Berger wrote: >> >>>On January 02, 2002 at 15:34:55, David Rasmussen wrote: >>> >>>>I've tried running Crafty at my universitys multiprocessor machines, with poor >>>>results. Using 4 processors was kind of ok, but not great (~750 knps), but using >>>>many more processors (16, 24, etc.) just made it slower. How crummy is that. >>>> >>>>/David >>> >>>Are you sure? Sparc processors usually scale very well. >>> >>>If it really became worse I'd suspect a problem with Crafty. >> >>No.. it is a problem with memory bandwidth. The sparc is too slow >>and the crossbar is even slower... If you don't run out of the cache, >>you get killed on memory bandwidth with large numbers of processors... >> >>> >>>OK, a silly question: did you remember to recompile Crafty so that it really >>>could use all those processors? >> >>That is a good point, with -DCPUS=128 or something similar... >>Otherwise most processors will be spinning and eating bandwidth for >>nothing... >> >> > >I did that already. > >I really don't understand how the Sparcs can be _that_ bad. I mean, they are >used at my university for large mathematical problems using lots of RAM and >processors, and they _do_ scale. Otherwise they wouldn't have bought such an >expensive machine, let alone more than one of them. > >/David It is almost certainly a horrible MUTEX implementation. POSIX threads does terribly on Linux if I have to use the normal MUTEX (pthread_lock()) stuff. The spinlocks I use make everything work. Although since I no longer have to lock the hash stuff, there are not a lot of locks left in Crafty so this might not be true as I haven't tested it in a long while...
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.