Re: [Qemu-devel] [PATCH v3 0/7] aio: Support epoll by introducing qemu_poll abstraction

[Fam Zheng <famz@redhat.com>]

On Thu, 04/16 14:03, Stefan Hajnoczi wrote:
> On Thu, Apr 16, 2015 at 12:57:29PM +0800, Fam Zheng wrote:
> > v3: Rebase to master for 2.4.
> >     Although epoll_pwait1 syscall is still being worked on [1], the QEMU part
> >     (if any) will base on this, so let's merge it first.
> >     
> >     That part is not included in this version because I'm still evaluating by
> >     comparing epoll_pwait1 with epoll+timerfd as with current master they seem
> >     to be really close.
> > 
> >     [1]: http://www.spinics.net/lists/linux-api/msg08216.html
> > 
> > v2: Emulate nanoseconds precison of timeout with ppoll and timerfd.
> >     Their performance is on par with each other, but both much better than
> >     qemu.git:
> > 
> >     syscall         high # of fd      low # of fd
> >     -------------------------------------------------
> >     qemu.git(ppoll) 44                96
> >     ppoll+epoll     85                101
> >     timerfd+epoll   87                109
> > 
> > (In high # of fd case, 3 activated but idle virtio-console devices are
> > attached, which will add us hundereds of fds to poll)
> 
> Have you rerun benchmarks with this patch series?

Yes, here:

    syscall         high # of fd      low # of fd
                    (Unit MB/s)       (Unit MB/s)
    -------------------------------------------------
    qemu.git(ppoll) 24                73
    ppoll+epoll     49                77
    timerfd+epoll   49                82

> 
> I wonder how the ppoll-only performance changes.  It seems like there
> are now additional copies of <fd, events, revents> information and
> corresponding malloc/realloc/frees.
> 
> Stefan