Author: Miguel A. Ballicora
Date: 06:52:37 10/07/01
Go up one level in this thread
On October 05, 2001 at 15:12:40, Robert Hyatt wrote: >>>for the addition of a second processor, for this comparison to be reasonable. >>>That is a common flaw in many reported parallel speedups. IE the Waycool guys >>>at CalTech added another processor _and_ another block of memory. And they >>>didn't separate the gain from the additional processor from the gain from the >>>additional hash table size... >> >>Wait just a second! you are not doubling the resources then?? then you >>do not duplicate the worker, you give him another brain but not double arms >>and legs. Ok. >>Anyway, this means that if you double all the resources you admit >>that a super linear speed up is possible (please see level (1) >>in my other post)? yes or no? Let's agree in the question first. >> >>Miguel > >NO one disagrees on that. We are talking about two processors vs one processor, >_everything_ else is the same. In the case of cpus, sure you get 8 registers on >one intel cpu, 16 on another. you get one cache on one cpu, two on a dual. >You get one TLB on one, two on a dual. But that is all. Those don't have any >influence on the _algorithm_. But more memory definitely does. Just take >any program and run it with a 4M hash table, then an 8M hash table. you will >see the difference even though you use only a one-processor machine. This is extremely unfair. You are not doubling everything, you just double part of the hardware, how can you want me to even reach 2.0? The second CPU should be able to use another stack, for instance, so it needs more memory. Anyway, let's go step by step. You admit that if you double ALL the resources it is possible to have a superlinear speedup? yes or no? [snii] >So that is the "context" for any valid super-linear speedup discussion. It is >not about trying to fabricate an algorithm or test case that runs out of >hardware resources on the single but not on the dual. IE there are plenty of >machines with 32 general registers. And they won't suffer like the intel >processors do with only 8. And then there are the crays with hundreds of >registers... We are talking about a superlinear speed up in a given real machine, not a theoretical one with infinite registers, infinite memory etc. You are denying the possibility to have linear speedups even on intel and AMD machines where the resources are limited. Bringing up the Cray is irrelevant. But first, please, answer my above question to we discuss about the same things. Regards, Miguel > > >> >> >>> >>>Here we fix _everything_ except the number of cpus...
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.