Here is a simple benchmark which is NUMA sensitive and simulates a simple
but normal situation in an environment running number crunching jobs. It
starts N independent tasks which access a large array in a random manner.
This is both bandwidth and latency sensitive. The output shows on which
node(s) the tasks have spent their lives. Additionally it shows (on a
NUMA scheduler kernel) the homenode (iSched).

Could you please run it on the virgin kernel and on the
"Both numasched patches, hack node IDs, alloc from current->node" one?

Maybe we see what's wrong...

Thanks,
Erich