Number of virtual cores equals number of virtual sockets instead of using hardware topology for QEMU version 2 and up

[martijnkruiten]

In QEMU version 2 and higher each core on the VM is assigned to its own virtual socket instead of using the hardware topology:

pcocc/lib/pcocc/Hypervisor.py

Lines 954 to 955 in f9a7944

	cmdline += ['-smp', 'threads=1,cores=1,sockets=%d' %
	(num_cores)]

On older QEMU versions the sockets are specified instead:

pcocc/lib/pcocc/Hypervisor.py

Lines 957 to 958 in f9a7944

	cmdline += ['-smp', '%d,sockets=%d' %
	(num_cores, len(cores_on_numa))]

Is there a reason to avoid this on version 2 and higher, or has this just not been tested? I wrote a small patch that mimics the version 1 behaviour for version 2 and up, and I don't see ill effects. The intra and inter socket communication behaves as expected. I was thinking of enabling this only when autobind_cpumem equals True, and possibly with a check if the topology is symmetrical. Just checking if you already have some experience with this. It boils down to something like this:

cmdline += ['-smp', 'threads=1,cores=%d,sockets=%d' %
            (num_cores/len(cores_on_numa), len(cores_on_numa))]

[fdiakh]

Hi,

I think the reason was that this only works for homogeneous sockets which is not necessarily the case when you only allocate part of a node. You could indeed add a check for symmetrical sockets and only respect the host socket characteristics in that case. However, looking at recent Qemu documentation it looks like it should now be possible to properly handle heterogeneous sockets using the -numa cpu syntax. Would you be interested in trying this method ?