Re: [libvirt] [PATCH] nodeinfo: Add workaround if kernel reports bogous numa topology.
On 10/23/12 17:07, Eric Blake wrote:
On 10/23/2012 06:40 AM, Peter Krempa wrote:
> This patch adds a workaround for the cpu topology detection code if the
> kernel reports incorrect count of numa nodes or any other reason that
> might result in duplicate core ID's in one socket.
Do we know what versions of kernels have this bug? If it is still
current, has this been opened as a bug report against the kernel?
The kernel version this happens in is 3.5.3-1.fc17.x86_64. You need a
two socket AMD magny-cours machine. The problem might be that the kernel
doesn't know about the topology due to hardware issues (not yet known)
[...]
If you install the 'hwloc' package on the afflicted machine, can you
then send us the file created by 'lstopo proc.png', so we can get a
better feel for what the machine actually supports? Is it really a case
of duplicate processor ids being mistakenly reported by the kernel, or
more a case of us mis-reading sysfs and seeing duplicates because we
aren't looking in enough other places?
At any rate, this patch looks okay, if we can prove that the kernel
really is giving us bogus information, and not just a case of us
misreading things.
Forwarding the response from George-Cristian Bîrzan who initialy
reported that:
George-Cristian has a bunch of identical machines where some report
having 4 NUMA cells and some just 1:
[...]
I did it on two hosts, one with 1 NUMA cell, one with 4 (as I said before,
they both only report 12 cores though):
http://birzan.org/proc1.png
http://birzan.org/proc4.png
------
I think we should take this patch as it resolves this case. The data
reported by kernel looks OK and the kernel probably trusts that
everything is OK.
Peter
