NUMA remote memory access cost

[Michael Neumann <mneumann@xxxxxxxx>]

Today I had a discussion about SMP and NUMA with my students and I
realized that Opertons really are NUMA as memory accesses are
non-uniform (I was under the impression that SMP not necessarily imply
UMA and that SMP/NUMA are othogonal aspects).

Anyone knows the cost (in case of an Opteron) of a remote memory access
vs. a local memory access?

Is the difference significant? And if it matters, is DragonFly already
aware about which memory locations are local to which CPU, and is this
information used in the memory allocator? (IIRC the memory allocator is
CPU localized, but that does not automatically imply that it knows about
remote/local memory locations). I think that's the only place where this
information is really important. The userland scheduler of course could
take advantage from it as well.

Regards,

  Michael