Author: Robert Hyatt
Date: 21:00:06 10/04/03
Go up one level in this thread
On October 04, 2003 at 08:57:14, Sune Fischer wrote: >On October 04, 2003 at 07:57:13, Dan Andersson wrote: > >> If you are able to do N probes in parallel without latency penalties that in >>itself means that there is an inefficiency. To be able to do that you will have >>to have local TTs. And those have less global search information. >> >>MvH Dan Andersson > >Are you saying that running two threads will double the latency for each thread, >and running N threads increase latency by a factor N? The issue is this. Each processor has local memory that is faster than accessing remote memory in another processor. With 2 threads, you have a 50-50 chance of probing fast local memory or slower remote memory. And remote probes cause conflicts at remote nodes with their processors trying to access local memory. > >Do they get queued or something? > >That would make it almost impossible to program for, not knowing the latency of >a probe. > >-S. It is nearly impossible to know. You can run a particular program, with a particular position, and compute the average hash access time, but it will vary for a different position, or different numbers of processors, or a different tree splitting, etc.
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.