Author: Vincent Diepeveen
Date: 07:54:12 09/03/03
Go up one level in this thread
On September 03, 2003 at 10:18:38, Matthew Hull wrote: >On September 03, 2003 at 10:13:30, Gian-Carlo Pascutto wrote: > >>On September 03, 2003 at 10:10:53, Sune Fischer wrote: >> >>>What I mean is, that since communication between threads is expensive it is >>>better to keep it to a minimum, obviously. >>> >>>Hence it is more efficient for the tread that discoveres something new to >>>'message' the other threads when that (rare) event happens, then for the other >>>threads to check for new 'messages' at *every* node. >>> >>>Of course the message should be delivered in the child threads local mailbox, >>>with low latency. >>> >>>Or am I missing something? >> >>Yes. To discover whether it has happened, you need score updates from >>the other processors anyway. >> >>You end up doing remote memory access whatever you solution you try. > >Are you saying that such accesses effectively stall the search, or is this a >relatively small portion of total time spent? For crafty when after a while you manage to compile it for a cc-NUMA machine, you will find out very soon that it is wasting 99.9999% of its system time to those accesses. It takes forever for it at 500 processors to get a decent nps. >MH > >> >>-- >>GCP
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.