Re: [PATCH v3 3/3] Add support for RAPL MSRs in KVM/Qemu

[Anthony Harivel]

Hi Daniel,

Daniel P. Berrangé, Mar 12, 2024 at 16:49:

> The point still stands though. NUMA node ID numbers are not
> guaranteed to be the same as socket ID numbers. Very often
> then will be the same (which makes it annoying to test as it
> is easy to not realize the difference), but we can't rely on
> that.
>
> > I'm using functions of libnuma to populate the maxpkgs of the host. 
> > I tested this on different Intel CPU with multiple packages and this 
> > has always returned the good number of packages. A false positive ?
>
> maxpkgs comes from vmsr_get_max_physical_package() which you're
> reading from sysfs, rather than libnuma.
>
> > So here I'm checking if the thread has run on the package number 'i'. 
> > I populate 'numa_node_id' with numa_node_of_cpu().
> > 
> > I did not wanted to reinvent the wheel and the only lib that was talking 
> > about "node" was libnuma.
>
> I'm not actually convinced we need to use libnuma at all. IIUC, you're
> just trying to track all CPUs within the same physical socket (package).
> I don't think we need to care about NUMA nodes to do that tracking.
>

Alright, having a deeper look I'm actually using NUMA for 2 info:

- How many cpu per Package: this helps me calculate the ratio.

- To whom package the cpu belongs: to calculate the ratio with the right 
  package energy counter.

Without libnuma, I'm bit confused on how to handle this. 

Should I parse /sys/bus/node/devices/node* to know how many packages ?
Should I parse /sys/bus/node/devices/node0/cpu0/topology/core_cpus_list 
to handle which cpu belongs to which package ?

Would that be too cumbusome for the user to enter the detail about how
many packages and how many cpu per pakages ? 

i.e: 
-kvm,rapl=true,maxpkgs=2,cpupkgs=8,rapl-helper-socket=/path/sock.sock

> > Maybe I'm wrong assuming that a "node" (defined as an area where all 
> > memory has the same speed as seen from a particular CPU) could lead me 
> > to the packages number ?
>
> Historically you could have multiple sockets in the same NUMA node
> ie a m:1 mapping.
>
> These days with AMD sockets, you can have 1 socket compromising
> many NUMA nodes, as individual dies within a socket are each their
> own NUMA node. So a 1:m mapping
>
> On Intel I think it is still typical to have 1 socket per numa
> node, but again I don't think we can rely on that 1:1 mapping.
>
> Fortunately I don't think it matters, since it looks like you
> don't really need to track NUMA nodes, only sockets (phnysical
> package IDs)
>

Very informative, thanks !

> With regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

Regards,
Anthony