On Wed, 12 Aug 2009 01:29:46 -0400, Alan Altmark <alan_altm...@us.ibm.com > wrote:
>On Sunday, 08/09/2009 at 11:01 EDT, Alan Ackerman ><alan.acker...@earthlink.net> wrote: > >> The MP effect is determined by the operating system and by the workloa d. >> It is not simply a hardware effect at all. If two or more processors >> access the same memory location -- one will have to wait. > >True enough, but understand that the Non-Uniform Memory Access (NUMA) >designs of modern machines changes the equation. It is no longer a bunc h >of CPUs equally competing for memory access. The hierarchy of non-share d >and shared cache brings an exquisite complexity to dispatching that >affects memory access and throughput. > >So when you add a CPU, its contribution to The MP Effect depends on wher e >in the cluster it is physically added and on the operating system's >ability to recognize the topology and to adjust its dispatching algorith ms >accordingly. But TANSTAAFL. It is an axiom that adding a CPU adds >overhead. > snip This reads a bit like a Lynn Wheeler post. :-) In support of Alan's comments on the general benefit of NUMA, I reached into my dusty memory. Back in 2000 and 2001 while I was still at IBM, we did some measurements of the same VM CMS workloads that were used to measure VM performance on regular IBM mainframes on the NUMA-Q Enabled For S/390 product in the lab . I recall that we went as far as measuring a 7-way (two NUMA quads of 700M hz procs lashed together) VM/ESA 2.4 system in increments of 1 CPU added at a time until we reached seven. The MP effect was much less with FLEX-ES on the NUMA-Q system than any conventional S/390 machines of that era where the workloads had been measured. I seem to recall a chart that was produced t hat showed ITR increase to be dang near linear as processors were added. We h ad not hit a knee in the line through seven processors. It (VM MP-effect) wa s there, just much smaller than usual. Sadly, I couldn't get my hands on another quad to continue the runs to 8-way and beyond. It would have been interesting to try and locate the knee. -- Gary Eheman
