Author: Robert Hyatt
Date: 10:56:32 09/03/03
Go up one level in this thread
On September 03, 2003 at 13:52:53, Matthew Hull wrote: >On September 03, 2003 at 13:29:02, Robert Hyatt wrote: > >>On September 03, 2003 at 13:12:34, Gian-Carlo Pascutto wrote: >> >>>On September 03, 2003 at 12:05:43, Robert Hyatt wrote: >>> >>>>>As I've said, there's nothing that magically makes SMP *inherently* faster >>>>>than NUMA. Nothing more. >>>> >>>> >>>>That's not what I said. SMP _is_ inherently faster. Because all memory has >>>>the same access latency. You _do_ have to share _something_ in a parallel >>>>algorithm. ANd whatever that is will be slower than doing the same thing on >>>>a SMP box. Even if it is just one word, the SMP box will access that one >>>>word faster all around and the program will run faster. >>>> >>>>Perhaps not a lot faster for 1 word of shared data. But faster nonetheless. >>> >>>I don't understand. >>> >>>Even the slowest access on a NUMA Opeteron is twice as fast as on a SMP >>>Xeon. >>> >>>How can it be slower then? >> >>You keep changing the subject. I am not comparing apples to oranges. I am >>comparing two machines that are _identical_ in every way except one has a pure >>SMP memory interconnection while the other has a pure NUMA interconnection. >> >>No references to X86 vs Opteron. No references to Cray vs Sun. If you give >>me two boxes that are identical except for SMP vs NUMA, the SMP box will >>_always_ have a speed advantage. It might not be much for small numbers of >>processors, but it _will_ be there. >> >>But if you want to compare opteron NUMA to something else, I'll take a Cray T932 >>which is a pure crossbar SMP machine. Want to compare latencies there? The >>Cray is _always_ 120 ns. No matter what part of memory from which processor >>you access. >> >>However, that is just as unfair as opteron to X86. >> >>NUMA is worse, period, when compared to an equivalent non-NUMA machine, in all >>respects _but_ pricing. That is where NUMA shines, and it is why NUMA was >>developed in the first place. NUMA was _not_ a solution to a performance >>problem. It was a solution to a _pricing_ problem. The Crossbar was a solution >>to a performance problem. > >I think I understand exactly what you are saying (this discussion has been very >imformative for me). > >Question: Touching on what GCP seemed to be getting at, if I were to go out and >buy one of these quad Opteron systems and compile a current SMP version of >Crafty on it, could it get a similar n-way speedup percentage like the Xeon >quads, even though it's not SMP, but NUMA with some kind of fast latency (is >that an oxymoron)? I understand the sentiment of "why settle for that when you >can optimise for NUMA and go yet faster". I don't know. Since I don't yet have such a box. I can answer this for a dual in a while, once we get some. However, no matter what the current version does, a newer version that correctly allocates split blocks will be even faster on the opteron (or any NUMA box). Once I get my hands on an opteron box, I'll be more than happy to publish some results. I prefer to avoid speculating (Vincent) when the question can be factually answered in time. > >Thanks, >Matt > > >> >>> >>>-- >>>GCP
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.