Author: Vincent Diepeveen
Date: 04:11:59 05/19/04
Go up one level in this thread
On May 19, 2004 at 03:02:05, Russell Reagan wrote: >On May 19, 2004 at 00:29:42, Joshua Shriver wrote: > >>Are there any kind of hardware limitations in computer competitions? >> >>If not, I'd imagine people would just bring a small custom cluster. >> >>TSCP would beat Hiarcs or Shredder if tscp was parallelized and Hiarcs was on a >>486. > > >I'm not sure that is a given. Alpha-beta based programs don't scale well on >clusters. A smaller number of processors that scales decently is probably just >as good or better than a cluster. Very good answer. DIEP has a speedup of about 100 out of 460 processors. Latencies same like cluster latencies (nowadays network cards are 2 - 3 times better random access latency but processors also a lot faster than 3 times). This speedup is not lineair. So speedup % at 130 processors for example is significantly better than at 460. Note 1 processor was 500Mhz and nowadays processors are like 2.x Ghz opteron so the speed is 5 times faster now which takes care it is way harder to get that speedup 100 out of 460 at todays clusters. But to give an impression when a quad opteron is not faster than a cluster let's calculate the break even point using lineair extrapolation. 4 * 460 / 100 = 18.4 processors. So for DIEP you will need a 32 processor cluster opteron to be significantly faster than a quad opteron using the same processors. Usually clusters and supercomputers have way slower processors than your quad opteron too, so soon you hit 64 processors as a minimum. In short clusters are great for the only program working at it which is DIEP, but even then the number of processors needed is that huge that it is not easy to get them. > >>Just an idea; perhaps there should be some kind of limitation. >>If not then you're not really testing the strength of the engines, but a >>combination of code and hardware. In that case, whoever has the most money has a >>huge advantage. Especially if clustering is allowed. > > >That is the goal of the WCCC, to have a competition between the best >computerized chess playing entities, not to test which software is the best. How >do you limit computer chess playing entities like Hydra, Deep Blue, etc.? What >about the new chips from AMD that will have multilpe cores? How would you >propose to run an event where hardware isn't a factor? > >Maybe the tournament organizers could provide everyone with equal hardware... 3.0Ghz P4's are available this tournament for those who do not remote connect nor take their own machine. >"But I have spent years working on multiprocessor search and you only provide a >single processor machine. That's not fair!" > >Okay, so provide everyone with dual machines... > >"But I have spent years working on my single processor program, so these guys >who support multiple processors have an unfair advantage!" > >"My program scales up to four processors, but you only provide dual machines. >That's not fair!" > >If we provide Intel machines... > >"But my program runs better on AMD machines! That's not fair!" > >If we provide 32-bit machines... > >"But my program is a bitboard program and runs better on 64-bit machines. That's >not fair!" > >If we provide any hardware at all... > >"But my program runs on specialized dedicated chess processors like Deep Blue!" > >If we provide hardware with Windows installed... > >"But my program only runs in Linux!" > >You get the idea...
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.