Author: Robert Hyatt
Date: 10:58:11 08/06/03
Go up one level in this thread
On August 06, 2003 at 00:34:40, Matthew Hull wrote: >On August 06, 2003 at 00:21:09, Robert Hyatt wrote: > >>On August 06, 2003 at 00:09:32, Slater Wold wrote: >> >>>On August 05, 2003 at 23:54:42, Slater Wold wrote: >>> >>>>On August 05, 2003 at 20:00:44, Robert Hyatt wrote: >>>> >>>>>On August 05, 2003 at 19:13:21, Slater Wold wrote: >>>>> >>>>>>On August 05, 2003 at 18:38:51, Robert Hyatt wrote: >>>>>> >>>>>>>On August 05, 2003 at 17:30:51, Matthew Hull wrote: >>>>>>> >>>>>>>>Does Crafty scale above 4 processors? For example, could crafty utilize all the >>>>>>>>CPUs on an IBM pSeries 690 7040-681T 32-way? >>>>>>>> >>>>>>>>http://www-132.ibm.com/content/home/store_IBMPublicUSA/en_US/eServer/pSeries/high_end/690.html >>>>>>>> >>>>>>>>MH >>>>>>> >>>>>>> >>>>>>>I didn't have time to look closely. But it is most likely a NUMA platform, >>>>>>>which means that Crafty as it exists now is not going to work well on it. NUMA >>>>>>>machines require careful attention to what is put where in memory, so that >>>>>>>often-used data is as close to the physical processor (in terms of access >>>>>>>latency) as possible. The current implementation of SMP in Crafty is based >>>>>>>on pure SMP, where memory is simply shared. >>>>>>> >>>>>>>If I ever have time to fiddle with a NUMA machine, I'll probably look at >>>>>>>fixing the major issue, which is to put split blocks close to each processor, >>>>>>>and when giving a specific processor a tree to search, using a split block that >>>>>>>is _close_ to it. >>>>>> >>>>>>I really don't think it is a NUMA machine. >>>>>> >>>>>>"Advanced symmetric multiprocessing (SMP) up to 32-way", is how the exactly >>>>>>describe it. >>>>>> >>>>>>And when you go to their 'Clusters' section, this machine isn't mentioned. >>>>> >>>>> >>>>>NUMA is not necessarily a "cluster". There are several machines that are >>>>>called "SMP" but also use a NUMA approach for memory. The clue is the price. >>>>>A real 32-way shared memory system is _really_ expensive. Cray claimed that >>>>>50+% of the price of the T932 was in the memory interconnect to support full >>>>>memory speed to all processors. >>>>> >>>>>Of course, I'm not wagering anything on this "hip shot" as to the idea behind >>>>>the machine, but from what I have seen in past years, anything beyond 8 CPUS >>>>>is either _way_ expensive, or NUMA. >>>> >>>>Sorry, I mispoke myself....scratch that, let's try again. >>>> >>>>Reading >>>>http://www-1.ibm.com/servers/eserver/pseries/hardware/whitepapers/p690_config.pdf, >>>>it would appear that it is a true SMP system. >>>> >>>>IBM has their own way of doing things, using MCMs and such, with the GX >>>>controllers. >>>> >>>> >>>>32 64-bit 1.7Ghz POWER4+ CPUs, 512GB RAM, and Crafty 19.04. Whew. >>> >>>Oh, and..."Pricing for the p690 starts at US$493,386 with an eight-way >>>configuration featuring Power4+ 1.5 GHz processors, 8 GB of main memory and two >>>36 GB disk drives with AIX preinstalled." >>> >>>Starts...at...half...a...million...dollars... >>> >>>That's about $62k per CPU+1GB RAM. Take off 10% for each CPU+1GB RAM. >>> >>>Add 24 CPUs, then add 24GB RAM. You're at somewhere around $2M. >> >> >>at that price it _could_ be a real SMP platform... >> >>And Crafty _should_ scale just fine on it... > > >It looks like an MCM will use memory interleaving for pure SMP. An MCM can have >up to 8 processors in it. A 32-way has 4 MCMs. But I would assume that a >32-way is just 4 8-ways just like and 8-way Xeon is just two 4-ways. > >MH RIght... but that "smells" like a NUMA approach. IE most likely the 8 processors connected directly to a MCM will have a memory latency of X, while a processor connected to another MCM (say MCM B) trying to access memory connected to MCM A will have a latency of X+Q where Q is the MCM-MCM latency. That is a classic example of NUMA (Non-uniform memory architecture).
This page took 0.01 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.