Re: [Xen-users] Xen 4.6 and Intel MKL Linpack behaviour

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

Adding Wei and Dario who did some work on topology before.

On Sun, May 08, 2016 at 01:44:48PM +0200, Marko Đukić wrote:
> On 7 May 2016 at 16:41, Marko Đukić <marko.djukic@xxxxxxxxx> wrote:
> >
> > Hello!
> >
> > I am doing some testing with Xen by running some benchmarking software 
> > inside a virtual machine. I have set up a VM with 4 vCPUs, the host 
> > hardware has 1 CPU with 4 cores. Both host and guest are running ubuntu 
> > server 16.04.
> >
> > When running Intel MKL Linpack 
> > (https://software.intel.com/en-us/articles/intel-mkl-benchmarks-suite) 
> > inside a PVH or PV guest Linpack only detects 1 CPU with 1 core (output at 
> > the start of the program). If I change the guest configuration to HVM, 
> > Linpack detects 1 CPU with 4 cores.
> >
> > Is this a bug in PV/PVH mode or am I missing a configuration setting?
> >
> > The configuration file for PV guest is as follows:
> >
> > name = "ubuntu64"
> > bootloader = "/usr/lib/bin/xen-4.6/bin/pygrub"
> > memory = 8192
> > vcpus = 4
> >
> > vif = ['script=vif-bridge,bridge=br0']
> > disk = ['tap:aio:/home/vm/xen_ubuntu16.04_amd64.raw,xvda,w' ]
> >
> > For PVH guest just pvh = 1 is added.
> >
> > Regards
> >
> > Marko
> 
> 
> I checked the /proc/cpuinfo of PVH and HVM guest and saw a difference
> in cores - in HVM guest cpuinfo shows 4 cores, in PVH cpuinfo shows
> only one core.
> Google then pointed me out that a similar question has apparently
> already been answered (http://markmail.org/message/vpj3ajlg6h7fkzro):
> quote (Wei Liu):
> "I guess LINPACK relies heavily on using topology to make decisions on
> how it should run?
> 
> Give guest host-cpu toplogy -- no, it's not possible to do this at the
> moment. That's a feature under development."
> 
> Can anyone explain why vcpus are presented differently to the guest if
> we look at PV/PVH and HVM?

There are several differences between PV(H) and HVM. For once, the cpuid 
information returned to PV and HVM guests is different. Also, PV(H) guests 
don't have ACPI tables, which is were some of this topology is reported. HVM 
guests OTOH have ACPI tables that report a sensible topology.

You should try to figure out how Linpack finds out about the CPU topology, 
and then we can maybe try to fix it.

Roger.

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users