Il 26/02/2014 16:14, Christian Borntraeger ha scritto:
Paolo,
here is the reworked yield heuristics series against kvm/queue with
your suggested changes.
Attached is a minimized testcase that reproduces the performance
win (runtime 0:50 instead of 1:00). The constants and the setup
seem a bit artificial but these seem to reproduce the problem on my
test system: as soon as there is contention from other guests the
patch improves the runtime of belows testcase. The original test
had 8 competing guests with 4 cpus, I simplified that a bit to have
one big guest as cpu stealer.
host 20 cpus
GUEST1 (20vCPUS) GUEST2(40vCPUS)
runs the test all 40 CPUs cpu bound
The testcase stresses the IPC system call, since that creates a lot
of diag44 exits on s390 if contended.
Some other spinlock tests like fio on guest tmpfs seem to be unaffected.
Please consider for kvm/queue. If you have some performance tests
a regression run on x86 is welcome.
Yes, will apply as soon as kvm/queue moves to kvm/next (1 hour or so).
Paolo
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
