Bill Davidsen wrote:
People are still doing graduate thesis on NUMA, it's not clear that
more people will mean shorter time to solution. The first cut may
settle for being stable and running well where Ntask = Ncpu. Getting
it wrong means bigtime bad throughput, which is probably an issue,
since there are more mature Linux and Solaris (I'm told) models as
competition.

I think the rest of the steps are pretty well understood and can be
solved in reasonable time. Or course things like keeping the CPU
number in a byte will need work ;-)


I just remembered that pretty soon there will be dual-core Opterons too, so
that in itself will add another level of NUMA to take into account. Internal
chip-connect vs. Hypertransport connect vs. customized board-to-board
connect.

Yousuf Khan