Le 27/11/2018 à 02:05, Segher Boessenkool a écrit :
> Cool stuff :-)
> gcc110 (a power7, 2 packages, 16 cores, 64 threads):
> $ hwloc-ls -s --no-io
> depth 0: 1 Machine (type #1)
> depth 1: 2 NUMANode (type #2)
> depth 2: 16 Package (type #3)
> depth 3: 16 L3Cache (type #4)
> depth 4: 16 L2Cache (type #4)
> depth 5: 16 L1dCache (type #4)
> depth 6: 16 L1iCache (type #4)
> depth 7: 16 Core (type #5)
> depth 8: 64 PU (type #6)
>
> gcc112 (a power8, 2 DCMs (i.e. 2 packages, 4 dies), 20 cores, 160 threads):
> $ hwloc-ls -s --no-io
> depth 0: 1 Machine (type #1)
> depth 1: 2 Group0 (type #7)
> depth 2: 4 NUMANode (type #2)
> depth 3: 4 Package (type #3)
> depth 4: 20 L3Cache (type #4)
> depth 5: 20 L2Cache (type #4)
> depth 6: 20 L1dCache (type #4)
> depth 7: 20 L1iCache (type #4)
> depth 8: 20 Core (type #5)
> depth 9: 160 PU (type #6)
>
> gcc135 (a power9, 2 packages, 32 cores, 128 threads):
> $ hwloc-ls -s --no-io
> depth 0: 1 Machine (type #1)
> depth 1: 2 NUMANode (type #2)
> depth 2: 2 Package (type #3)
> depth 3: 16 L3Cache (type #4)
> depth 4: 16 L2Cache (type #4)
> depth 5: 32 L1dCache (type #4)
> depth 6: 32 L1iCache (type #4)
> depth 7: 32 Core (type #5)
> depth 8: 128 PU (type #6)
>
> so it gets p7 packages wrong, and it doesn't understand p8 DCMs. The rest
> is fine, and pretty etc. :-)
According to IBM, the POWER7 info is what they want to report :) I tried
to convince them to fix their firmware/kernel but they considered it was
better to report this strange stuff because of an obscure POWER feature
("LPAR" if I remember correctly).
User-space tools such as hwloc or lscpu often require recent kernels to
report correct information on some non-x86 hardware. On x86, things
usually work fine (but hwloc can still directly use the CPUID
instruction whenever the kernel reports something wrong).
Brice
_______________________________________________
cfarm-users mailing list
cfarm-users@lists.tetaneutral.net
https://lists.tetaneutral.net/listinfo/cfarm-users
