cancel
Showing results for 
Search instead for 
Did you mean: 

Server Gurus Discussions

iamdady
Journeyman III

The EPYC 7451 NUMA config is weird in Linux, why?

Hi all,

   We have test a AMD EPYC 7451 server, and some thine weird comes up:

The topology of 7451 is created by lstopo Linux command in the attachment file, what's confusing me is why PU P40, PU P88, PU P2, PU P50 share a same L3 Cache(see bellow) but in two different NUMANode? Different NUMANode means different die, that's impossible for one L3 cache cross two die.

The L3 share information is obtained from bellow command:

$ cat cpu40/cache/index3/shared_cpu_list

2,40,50,88

Thanks a lot!

0 Likes
1 Solution

Your topology is now correct. As abucodonosor​ pointed out, the fix for this particular issue (applies to all parts with < 32 cores) went into the upstream 4.13 kernel and has been back-ported to RHEL 7.4 as of the late December update (just before the Spectre / Meltdown mitigation releases).

-Monkey

View solution in original post

0 Likes
8 Replies
Anonymous
Not applicable

Hi iamdady,

That is strange. Would you do me a favor and provide the following:

  • Vendor
  • BIOS
  • OS and version
  • Kernel version if not default
  • Dump of lscpu command to a file as well
0 Likes

Hi,

Sorry for the account name, my name is Heron.

Vendor: Dell

BIOS: version 1.1.3

OS: No LSB modules are available.

Distributor ID: Ubuntu

Description: Ubuntu 16.04.4 LTS

Release: 16.04

Codename: xenial

Kernel: Linux 4.4.0-119-generic #143-Ubuntu SMP Mon Apr 2 16:08:24 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux

Lscpu: see attached file

Other info may help:

System Model: Name

PowerEdge R7425

UEFI Compliance Version

2.5

Thanks a lot!

0 Likes

This indeed looks weird .. However I don't think 4.4.* kernels have all AMD topology fixes.

Is possble for you to try an newer kernel ?

Hi,

I upgrade the Kernel to 4.16.2 and the topology changed as you think, see the attached file, so is this right now?

0 Likes

Can you please just post output of lscpu ?

0 Likes

Thank you very much.

Attached file is the lscpu output for Linux Kernel-4.16.2.

0 Likes

Your topology is now correct. As abucodonosor​ pointed out, the fix for this particular issue (applies to all parts with < 32 cores) went into the upstream 4.13 kernel and has been back-ported to RHEL 7.4 as of the late December update (just before the Spectre / Meltdown mitigation releases).

-Monkey

0 Likes

Thanks a lot!

0 Likes