On Wed, Oct 17, 2012 at 12:54 AM, Andrey Korolyov <andrey(a)xdel.ru> wrote:
Hi,
At least on 0.10.2 setting a cpuset doesn`t match a real process
placement - VM still consumes all available cores.
VM config:
.snip.
<vcpu placement='static' cpuset='0-5,12-17'>12</vcpu>
.snip.
for cpuset in $(find /cgroup/cpuset/libvirt/qemu/vmid/ -name
cpuset.cpus) ; do grep 0-5 $cpuset ; done
got: empty responce, e.g. 0-23 in my setup
expected: at least bounded vcpuX threads
In neighbor 'cpu' group libvirt sets cfq weights just fine, so this is
clearly a bug.
Oh, I was wrong in suggestion that libvirt using cgroups for task
pinning, nevermind please. There must be some issue with taskset which
does not prevent vm execution on all cores even if a specific set was
passed, I have wrongly thought it was because cpuset value remains
default in libvirt sub-cgroup.