Athlon MP vs Athlon XP
tlovie at pokey.mine.nu
Thu Nov 8 12:36:08 PST 2001
> On Thu, Nov 08, 2001 at 02:28:40PM -0500, Velocet wrote:
> > From what I understood from the useful articles that were
> posted here,
> > the cache protocol allows sharing data between the CPUs via the
> > northbridge directly.
> Right. What it comes down to is this: Getting data from L2 is
> always fastest if its in your own L2. But if it isn't, some
> machines fetch from main memory faster than they can fetch a
> dirty line from someone else's L2. AMD's scheme has
> reasonably fast main memory fetches, plus even more efficient
> fetches from a remote L2.
> I believe the Sun E10k is one of the few processors where
> main memory is closer than someone else's L2. That makes
> false sharing even worse than usual.
> However, from the beowulf standpoint, most of us are running
> 2 independent mpi proceses on dual cpu boxes, right?
I have a innocent question.... Does the kernel have processor affinity
built in to it yet? The situation may arise that one of the mpi
processes gets bumped from it's processor, by a system task, then it in
turn bumps the other mpi task from the other processor, and in effect,
it's info is cached in the other processor. I see the advantage that
fetching from a remote L2 is better, but does anybody know the status of
assigning a processor affinity mask to processes?
More information about the Beowulf