Author: Robert Hyatt
Date: 12:00:51 09/03/03
Go up one level in this thread
On September 03, 2003 at 14:27:24, Gian-Carlo Pascutto wrote: >On September 03, 2003 at 14:17:56, Robert Hyatt wrote: > >>You are hung up on the X86 SMP platform. Forget X86. Just think "SMP". >> >>Go look at the Cray XMP/YMP/C90/T90. They use a crossbar. _not_ a bus. >>A good C90 for example has 16 processors, 4096 banks of memory, all with a >>big crossbar connecting them. any processor can access any bank in 120ns, >>unless two bang on the same bank. Then there is a 10ns penalty or so >>added in (bank busy delay). >> >>The point is that SMP does _not_ mean "shared bus" except for the world of >>the PC and some other smaller SMP boxes (Sun, alpha, etc.) > >"you give me two boxes that are identical except for SMP vs NUMA, the SMP box >will _always_ have a speed advantage." > >To Cray or not to Cray, I have seen no proof of the above, on the >contrary. I gave a counterexample, which you've not refuted. > >-- >GCP OK. back to basics. A NUMA box has variable latency. SMP does not. Both machines will be built around the _same_ memory, which means raw latency will be the same at the memory bank measuring point. For SMP, that is all the delay there is. IE on a Cray (not a NUMA Cray like the T3 boxes) latency is 120ns for any processor to any word of memory. You take that same memory and put it on your NUMA box, because there is _nothing_ faster. now you don't directly access a memory bank. You go through an extra "hop" to get to a remote bank (one hop if you have two processors, for N processors it depends on how the routers are connected to each other). No matter what you do, your local access is going to be 120ns at best, and remote access will _always_ be less. Q.E.D. The SMP box will _always_ be faster for any application that shares _anything_ at all. Because sharing will _always_ be faster on the SMP box, period. I don't know how it can be explained any simpler than that. NUMA -> cheaper NUMA -> slower NUMA -> scales better to large N
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.