Author: Vincent Diepeveen
Date: 14:26:05 01/20/06
Go up one level in this thread
On January 20, 2006 at 13:52:30, Dann Corbit wrote: >On January 19, 2006 at 22:10:05, Mridul Muralidharan wrote: >[snip] >>Now my average speedup is 1.89x for time to depth for the positions that Vincent >>posted (Thanks Daniel !) with worst case being 1.3 and hopefully improving (on >>my dual core athlon box). > >That's really outstanding! >I think 1.89 average might be the best of any engine Not really, Mridul just started improving his both chessengines based upon diep's parallel source code. Diep's by far best scaling program on planet in shared memory machines. Shared memory doesn't mean it is a non-numa machine. It just means shared memory works, though it can be very slow NUMA :) What matters is this in scaling, that's the amount of effort you plan to spend on parallellizing your software and the scaling. Bob is doing something simple with a central lock for crafty. Works fine for him at PC hardware. If that works for you, then go for it. If you want even less effort to spend in parallellism, just share a hashtable, and do 4 independant searches! Speedup comes from hashtable in that case. In case of tree splitting we can categorize software of course. We can distinguish software that's global locking and software that's not global locking. The last category is what diep falls in. I assume Zappa soon will too. Obviously that last is also going to scale best. If you can't get a near to 2.0 scaling out of 2 cpu's at a dual core machine, at most losing system time to the more busy memory controller, then obviously you won't run very well at a 512 processor machine where memory latencies are a factor 50-100 worse than at an A64 :)
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.