[gentoo-amd64] Re: SMP

[Duncan]

"Mihir Sevak" <mihir.sevak@×××××.com> posted
48fcf8860605260703m2264be8hcd5a5fc9d43c86dd@××××××××××.com, excerpted
below, on  Fri, 26 May 2006 10:03:24 -0400:

> i am using C as a programming language. I have an AMD 64 dual core
> system. Is there any way i can take charge of those two processors and
> command them which thread should run on which processor??? I am prepared
> with scheduling mechanism but only thing i am lacking is how to take
> charge and decide that which processor should run which thread.

I'm rather far from an expert in the field, but AFAIK, you use kernel
scheduling domains for that, and configure your app to use a specific
scheduling domain.  AFAIK, the functionality is fairly new, and you'll
need 2.6.16 or even the 2.6.17-rcs, plus some special user side utilities
to configure and manage the domains, to get it working.

Automatic kernel SMP scheduling normally does a pretty good job, noting of
course that on some hardware, some chipset elements may be connected to
only one of the CPUs (tho that shouldn't be a problem for dual-core), and
the various costs of switching between cores on the same CPU, as opposed
to different CPU/sockets, as opposed to different hosts (in a clustering
environment), what with local vs remote memory and cache coherency and the
like.  Specifying a CPU/core is now possible using the special tools 
mentioned above, but it's seldom worth the trouble, and can actually be
worse.  It may be better to tweak the various tweakables exposed by the
kernel for its automated scheduling, changing how it chooses to switch
CPUs if one is busier than the other.

FWIW, I have a dual Opteron system here, single-cores.  I expect to be
upgrading later this year to dual-cores, for a 4-core system on dual
socket-940.  I've generally been fairly happy with the kernel's scheduling
as-is, allowing for when I'm running single-threaded apps that /can't/
divide the work.  If the app is multithreaded, or I'm running several apps
in parallel, the kernel generally does a pretty good job of keeping things
balanced.  Back when I had a single-disk system, that was often the
bottleneck.  However, with a 4 disk RAID system (kernel-RAID, RAID-6 for
most of it, RAID-0 for the performance critical redundancy unneeded stuff,
RAID-1 for /boot) and 8 gig of memory, the bottleneck is again usually the
CPUs, thus the planned upgrade to 2 by dual-cores later this year.

Of course, that's for personal desktop/workstation use.  If I were running
something less general purpose, say a big database server or the like,
there would indeed be reason to think about specific tuning including
possibly confining some tasks to specific CPUs.

-- 
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