Author: Robert Hyatt
Date: 21:07:54 04/25/02
Go up one level in this thread
On April 25, 2002 at 23:01:37, Tom Kerrigan wrote: >On April 25, 2002 at 13:29:34, Robert Hyatt wrote: > >>On April 25, 2002 at 04:12:26, Dan Andersson wrote: >> >>>You are correct. I read that first order NUMA optimizations where unneccessary. >>>But a memory access might make the other CPU wait for a while. And the overhead >>>for running a SMP driver is anyones guess. >>> >>>MvH Dan Andersson >> >> >>As you said, it _must_ make a difference. NUMA _always_ does... >> >>programming for NUMA machines requires additional analysis and planning to >>stay off the low-performance memory accesses whenever possible. > >It requires nothing of the sort. Sure, you'll get better performance if you do. > >-Tom Ever try to think first and _then_ write? Why does someone develop a parallel chess program? Anything to do with performance? When designing a _reasonable_ parallel algorithm the underlying hardware _must_ be considered. At least for those of us that don't just throw an algorithm onto a machine with no regard for reasonable performance. And _performance_ is what a parallel algorithm is developed for in the first place... otherwise nobody would take the time and effort to do so...
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.