Author: Vincent Diepeveen
Date: 05:56:01 06/01/01
Go up one level in this thread
On May 31, 2001 at 17:20:08, John Wentworth wrote: >I don't know much about coding for a multiprocessor chess program, but is it >possible to code a multiprocessor program that would harness the processing >power of multiple PC's that are networked together instead of one PC with >multiple processors? Yes it is possible to use the network. I'm busy on a version of diep that is going to run over the network. Indeed it's nice if each node has more as 1 cpu. Also 10mbit is really too little to get a speedup i fear. Further the algorithm i use is probably not going to work that well with a huge number of nodes. In general latency is a big problem, but even bigger is the bandwidth problem. Bandwidth of a 100mbit network is already very small. that's only 12.5MB a second. That's really little. What i also had forgotten a bit in my calculations, but where Bob pointed me on is that each node has to WAKE UP to process all data too which it gets. And if each node is a single cpu then this is a MAJOR problem because it takes hundreds of milliseconds before a process wakes up to check for network traffic. So using a network is far from perfect to get any speedup over a single node which has shared memory. If a single node is also a single processor then getting a speedup is very difficult. I have already something working now, but it definitely doesn't give a speedup. As soon as it's improved that i think it might give a speedup over 1 node, then i'm going to test it. I'm pretty amazed by the huge number of companies and private persons who have a 10mbit network. Only very little people have a 100mbit network, not to mention the companies having a 1Gbit network. Whatever the network now or in the future, you can simply forget to compare a 8 node network of 700Mhz P3s with a quad xeon. In no respect 8 nodes compare to even half of the processors. The more nodes one has the harder it is to get a speedup, where in SMP parallellism this is easier. My concept is NOT completely smp. It involves 1 master and a bunch of slaves. Last but not least of th eproblems is HOW to automatically START such a networked program easily? Of course i can go to each computer and start DIEP there and then type some things till it's connected to the rest, but that's very labour intensive. It's no fun to start at each node a program by hand! Suppose i have a network here and i click the diep icon, then i want it to start at all nodes! Such seemingly easy things are even with a networked version a problem!! 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.