Karel Gardas wrote:
Thanks Jan for the nice slides reference. The information is not there
unfortunately. Anyway, as I'm thinking about it more having 4,1,1,1 memory on
boards does not make a difference probably as long as no thread allocates more
than X MB of RAM causing OS to steal a RAM from CPU with 4 GB. So for mu
application which is heavily threaded but consumes just a tiny fraction of
available RAM it should not make any difference.
Thanks,
Karel
Karel - see my other post. You'll still have problems (well, performance
hits) if you use all 4 CPUs for an application, but the relevant memory
sections sit in another CPU group's local RAM.
The 4- and 8-socket systems worked best if you have multiple processes
running, each of which would fit into no more RAM that you would have
attached to a pair of CPUs, and you could pin that process to a CPU pair
group. They'll run in other configs, too, but less efficiently (read:
slower, and scale poorer) then the aforementioned case.
Likely for your test, I would do this:
(a) spend enough money to get a completely balanced RAM config for CPU
0 & 1
(b) buy just enough memory to fit your test into it for CPU 2 & 3,
preferably a multiple of 4 DIMMs
(c) run *2* of your test processes, one pinned to each CPU group.
--
Erik Trimble
Java System Support
Mailstop: usca22-123
Phone: x17195
Santa Clara, CA
_______________________________________________
opensolaris-discuss mailing list
[email protected]
