Re: [PATCH RFC 0/4] Paravirtual spinlocks

[Rusty Russell <rusty@rustcorp.com.au>]

On Tuesday 08 July 2008 10:37:54 Jeremy Fitzhardinge wrote:
> Rusty Russell wrote:
> > On Tuesday 08 July 2008 05:07:49 Jeremy Fitzhardinge wrote:
> >> At the most recent Xen Summit, Thomas Friebel presented a paper
> >> ("Preventing Guests from Spinning Around",
> >> http://xen.org/files/xensummitboston08/LHP.pdf) investigating the
> >> interactions between spinlocks and virtual machines.  Specifically, he
> >> looked at what happens when a lock-holding VCPU gets involuntarily
> >> preempted.
> >
> > I find it interesting that gang scheduling the guest was not suggested as
> > an obvious solution.
>
> It's an obvious answer, but not an obvious solution.  You trade off
> wasting time spinning vs wasting time waiting for N vcpus to be free for
> scheduling.

Perhaps, but with huge numbers of cores (as The Future seems to promise) and 
significant overcommit not sure how bad this would be.

> Or something; seems much more complex, particularly if you
> can do a small guest tweak to solve the problem.

But AFAICT it's one of a related set of problems where all VCPUs are required 
for a task.  Hackbench comes to mind.  There's going to be a lot of 
ping-ponging and you'll approach gang scheduling to get decent performance.

> > A little disappointing that you can't patch your version inline.
>
> Spinlock code isn't inlined currently, so I hadn't considered it.  The
> fast path code for both lock and unlock is nearly small enough to
> consider it, but it seems a bit fiddly.

Yeah, OK.

Thanks,
Rusty.