Re: [PATCH resend] x86,kvm: Add a kernel parameter to disable PV spinlock

[Oscar Salvador]

On 09/05/2017 08:28 AM, Juergen Gross wrote:
> On 05/09/17 00:21, Davidlohr Bueso wrote:
> > On Mon, 04 Sep 2017, Peter Zijlstra wrote:
> >
> > > For testing its trivial to hack your kernel and I don't feel this is
> > > something an Admin can make reasonable decisions about.
> > >
> > > So why? In general less knobs is better.
> > +1.
> >
> > Also, note how b8fa70b51aa (xen, pvticketlocks: Add xen_nopvspin parameter
> > to disable xen pv ticketlocks) has no justification as to why its wanted
> > in the first place. The only thing I could find was from 15a3eac0784
> > (xen/spinlock: Document the xen_nopvspin parameter):
> >
> > "Useful for diagnosing issues and comparing benchmarks in over-commit
> > CPU scenarios."
> Hmm, I think I should clarify the Xen knob, as I was the one requesting
> it:
>
> In my previous employment we had a configuration where dom0 ran
> exclusively on a dedicated set of physical cpus. We experienced
> scalability problems when doing I/O performance tests: with a decent
> number of dom0 cpus we achieved throughput of 700 MB/s with only 20%
> cpu load in dom0. A higher dom0 cpu count let the throughput drop to
> about 150 MB/s and cpu load was up to 100%. Reason was the additional
> load due to hypervisor interactions on a high frequency lock.
>
> So in special configurations at least for Xen the knob is useful for
> production environment.

It may be that the original patch was just to keep consistency between 
Xen and KVM, and also only for testing purposes.
But we find a case when a customer of ours is running some workloads 
with 1<->1 mapping between physical cores and virtual cores, and we 
realized that with the pv spinlocks disabled there is a 4-5% of 
performance gain.

A perf analysis showed that the application was very lock intensive with 
a lot of time spent in __raw_callee_save___pv_queued_spin_unlock.