HelloIndeed the L1d and L1i affinity of some cores is wrong. L1 are private to each core, hence shared by two threads in Pcores and single thread in Ecores. But what Linux report is 3 threads for some L1:
$ cat fedora/sys/devices/system/cpu/cpu{?,??}/cache/index0/shared_cpu_list 0,5 1-2,16 /* Pcore 8 + Ecore 0, wrong */ 1-2,16 3-4,18 /* Pcore 12 + Ecore 6, wrong */ 3-4,18 0,5 6-7 6-7 8-9 8-9 10-11 10-11 12 13 14 15 1-2,16 17 3-4,18 19 20 21Since we parse L1 first, we create the corresponding (buggy) L1. Later we fail to create L2 and clusters because those have correct localities that conflicts with the incorrect L1s.
The CPUID information looks OK but Linux fails to expose it correctly, maybe because the CPU numbering is very strange :
Lcores are 0-7 and 32-33 Pcore 8 has hyperthreads 1,2 /* doesn't start with 0 ?? */ Pcore 12 has 3,4 Pcore 16 has 0,5 /* what? */ Pcore 20 has 6,7 Pcore 24 has 8,9 Pcore 28 has 10,11 Usually it's rather 0,1 then 2,3 ... 10,11Core APIC ID are similarly misordered (grep apicid /proc/cpuinfo | grep -v initial)
These recent Intel CPUs are supposedly well supported by Linux, but the CPU numbering is so strange here that maybe Linux doesn't support that well. Maybe try upgrading the BIOS in case the numbering is fixed. Anyway there's a kernel bug here, although they might be unhappy to support such a strange case.
You should be able to workaround the issue in hwloc by exporting HWLOC_COMPONENTS=x86 in the environment so that hwloc reads CPU informations from CPUID directly instead of from Linux.
Brice Le 25/04/2025 à 12:48, Suyash a écrit :
hwloc prints the following error: * Failed with error: intersection without inclusion* while inserting Group0 (P#8 subtype Cluster groupkind 222-0 cpuset 0x000f0000)* at L1d (P#8 cpuset 0x00010006) * coming from: linux:sysfs:cluster Attached the resulting files from running `hwlock-gather-topology`.
To unsubscribe from this group and stop receiving emails from it, send an email to users+unsubscr...@lists.open-mpi.org.
OpenPGP_signature.asc
Description: OpenPGP digital signature