On 2-apr-10, at 14:49, Brice Goglin wrote:
Fawzi Mohamed wrote:
I would take advantage more info about the possible numa node
connectivity (to know where to steal tasks), but I don't have access
to machines that would really take advantage of that, and probably
even then using the HW structure as topology would not bad.
NUMA connectivity is a big problem. Most x86 machines show nothing
interesting in cat /sys/devices/system/node/node*/distance. On AMD
hypertransport, there are some routing information in lspci, but I
think
you have to be root to see it, and I am not even sure it's enough to
discover the actual HT graph.
We have a "measurement-based backend" in the TODO-list. It could be
the
only way to find out the NUMA connectivity.
nice
That said, it's not clear that this will be a big problem. AMD
Magny-Cours machines and Nehalem-EX machines with 2-4 sockets have a
fully-interconnected NUMA graph. No problem if we don't have NUMA
topology information there.. For larger machines, SGI is already
doing a
good job at reporting NUMA distances in sysfs. What remains is other
large machines (several vendors announced 8-socket Nehalem EX
machines).
We'll see.
indeed as I said the max I have access are 2 socket AMD or Nehalem-EX,
so it doesn't matter.
For now I just have:
PU core soket/numa_node machine system
and for the for the foreseeable future that is probably good enough
(all nodes in the same machine are at the same distance).
I don't aim at using this system between different machines...
Fawzi
