[gentoo-amd64] Re: Re: Re: experiences with Tyan S4881?

[Duncan]

"Tamas Karpati" <tkarpati@×××××.com> posted
bd453d7f0606220440n65ae201ftc7a5ebf17416eb8e@××××××××××.com, excerpted
below, on  Thu, 22 Jun 2006 13:40:48 +0200:

> Anyways, I hope memory can be shared on S4881 (or -D, S4882) as well
> as on S2882, meaning that any thread can utilize any part of (or all) the
> physical memory mounted on the mob. They want to use all the 64G memory
> (only "all" because (i) no mob supports more and (ii) resources are limited).

That should be no problem.  It's part of the AMD64 architecture, and
anything /not/ supporting it would therefore be considered crippled.

I'm not sure about 4-way, tho I would assume it would be similar to 2-way,
but with my 2-way (s2885), the BIOS has memory per-node and across-node
interleaving options.  For highest bandwidth access by a single thread,
one turns on the across-node interleaving.  For running multiple
threads, where each thread has faster access to the memory local to the
processor/node it is on, but /can/ access memory cross-nodes if necessary,
one turns off across-node interleaving, and enabled NUMA mode memory
management in the kernel.  The kernel process and memory control then
attempts to keep the processes balanced both by CPU use and memory use,
also accounting for threads sharing memory and trying to keep them on the
same node and therefore both accessing local memory, if possible. 
However, any thread or process can still access memory on any other node
if need be.  (There are separate kernel options available to wall off
particular applications, limiting them to specific CPUs and memory, if
desired, but I know little about these beyond the fact that they are still
somewhat experimental, from what I've read.)



-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

-- 
gentoo-amd64@g.o mailing list