Le 04/04/2020 à 21:58, Baptiste Jonglez a écrit :
> On 04-04-20, Brice Goglin wrote:
>> Using NUMA nodes for CPU sockets is indeed unreliable these days. Most
>> vendors can expose multiple NUMA nodes per CPU package.
> Ok, thanks.  Does this also apply to x86?


Yes, I was actually mostly referring about x86 here. Both Intel Xeon and
AMD Epyc can be configured with 2 or 4 NUMA nodes per package.

That's also true for POWER, some Sparc from Fujitsu, and likely some ARM
processors too (mostly likely the upcoming A64fx from Fujitsu too).


> Also, do you know if it's possible to obtain all information (number of
> nodes, packages, cores, pu) in a single call to hwloc-*, ideally with an
> easily-parsable output?


Maybe parse the output of "lstopo -s" (summary of the topology). Add
"--filter all:none --filter package:all --filter core:all" so that you
only get what you want (PU/Machine/NUMA are always enabled).

$ lstopo -s --filter all:none --filter package:all --filter core:all
depth 0:           1 Machine (type #0)
 depth 1:          2 Package (type #1)
  depth 2:         40 Core (type #2)
   depth 3:        80 PU (type #3)
Special depth -3:  4 NUMANode (type #13)

Brice


Attachment: signature.asc
Description: OpenPGP digital signature

_______________________________________________
cfarm-users mailing list
[email protected]
https://lists.tetaneutral.net/listinfo/cfarm-users

Reply via email to