Author: Robert Hyatt
Date: 19:32:59 09/15/01
Go up one level in this thread
On September 15, 2001 at 16:44:11, Vincent Diepeveen wrote: >On September 15, 2001 at 05:57:28, Dan Andersson wrote: > >>Due to the latencies incurred when communicating in a network. You need >>different work distrubution algorithms when programming a cluster compared to a >>SMP machine. There exists more than a few such algorithms that will get near >>linear speedups until the communication ceiling is reached. > >Network Latencies are of course a big problem for traditional chess programs. >For the approach i use the only 2 problems are there > a) network bandwidth > b) the fact that there is no smart way to let a process/thread > wake up only when there is incoming network traffic Just do a read, recv, recvfrom (udp) or whatever. You will block until the packet arrives, and then you will instantly unblock. I don't quite follow your statement... > >At home i have a 100mbit network which is simply TOO slow to get >a speedup at a big network. Of course between just 2 nodes it's >not easy but possible to get a decent speedup, because each node here >is a dual. > >But still why would i do so much effort more as i already did to get >a bit better speedup? Speedup is not much above 1.0 at 2 nodes right now >and there might be bugs still inside the parallellism. > >Best regards, >Vincent
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.