Author: Ernst A. Heinz
Date: 07:38:29 10/30/98
Go up one level in this thread
On October 30, 1998 at 09:55:15, Robert Hyatt wrote: >On October 30, 1998 at 04:44:36, Ernst A. Heinz wrote: > > [...] >> >>Hmh, AFAIK the Sequent machines are all NUMA types with a shared >>*address space* but non-uniform memory access times. I sincerely doubt >>that your strongly coupled DTS scales well on such machines. >> >>But I am happy to be proven wrong by experimental results ... :-) > >we had an older one with 30 cpus here until about 3 years ago. It was a bus >architecture with a flat shared memory design, large cache for each CPU, and >so forth. I did most of the DTS program development (Cray Blitz parallel >search) on that machine and it was pretty good... This was probably a "Sequent Symmetry" which was Sequent's last symmetric multiprocessor AFAIK. Thereafter, Sequent dived deep into the data warehousing market and switched to cc-NUMA architectures. >Haven't looked at them specifically in a while... but there are plenty of >machines around with a symmetrical shared memory... Yes, and they are getting better and more scalable all the time ... :-) >when you hit NUMA you >suddenly have to start thinking about what is allocated where to optimize >performance... yet another "issue"... Sure -- unfortunately, NUMA is especially ill-suited for algorithms that rely on centralized and shared work-queues. =Ernst=
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.