(different socket), it becomes a NUMA situation. (Is there a name for
this hybrid architecture?)
yes, it's called "NUMA" ;)
seriously, non-uniform doesn't mean uniformly non-uniform,
if you know what I mean.
otoh, NUMA should always be quantified, since a software net-shared-memory
system like ScaleMP (or other similar ones) is qualitatively different
from a 2-socket k8/i7 system. as a WAG, I'd guess remote memory references
would cost O(2us) in SW NUMA, ~500ns in a scalable cache-line-based NUMA
like SGI or the big-iron systems from IBM/Sun/HP, but ~100 ns in AMD-like
systems. as usual, if your workload is cache-friendly, this doesn't matter.
systems used Itanium processors which aren't binary compatible with x86
processors (a real inconvenience) if you were planning on running
commercial, binary-only x86 software)
IMO, the ia64 environment is going to remain different and obscure,
and there's little reason to go to it. Intel's hope was to push HPC
along the ia64 path, but AMD64 derailed that effort.
but there are newer Altix systems
(XE series) that use x86-based processors.
just IB clusters. I don't think any SW NUMA is included, or do they
support ScaleMP?
_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
