Re: [Qemu-devel] [question] virtio-blk performance degradationhappened with virito-serial

[Fam Zheng]

On Tue, 09/02 12:06, Amit Shah wrote:
> On (Mon) 01 Sep 2014 [20:52:46], Zhang Haoyu wrote:
> > >>> Hi, all
> > >>> 
> > >>> I start a VM with virtio-serial (default ports number: 31), and found 
> > >>> that virtio-blk performance degradation happened, about 25%, this 
> > >>> problem can be reproduced 100%.
> > >>> without virtio-serial:
> > >>> 4k-read-random 1186 IOPS
> > >>> with virtio-serial:
> > >>> 4k-read-random 871 IOPS
> > >>> 
> > >>> but if use max_ports=2 option to limit the max number of virio-serial 
> > >>> ports, then the IO performance degradation is not so serious, about 5%.
> > >>> 
> > >>> And, ide performance degradation does not happen with virtio-serial.
> > >>
> > >>Pretty sure it's related to MSI vectors in use.  It's possible that
> > >>the virtio-serial device takes up all the avl vectors in the guests,
> > >>leaving old-style irqs for the virtio-blk device.
> > >>
> > >I don't think so,
> > >I use iometer to test 64k-read(or write)-sequence case, if I disable the 
> > >virtio-serial dynamically via device manager->virtio-serial => disable,
> > >then the performance get promotion about 25% immediately, then I re-enable 
> > >the virtio-serial via device manager->virtio-serial => enable,
> > >the performance got back again, very obvious.
> > add comments:
> > Although the virtio-serial is enabled, I don't use it at all, the 
> > degradation still happened.
> 
> Using the vectors= option as mentioned below, you can restrict the
> number of MSI vectors the virtio-serial device gets.  You can then
> confirm whether it's MSI that's related to these issues.

Amit,

It's related to the big number of ioeventfds used in virtio-serial-pci. With
virtio-serial-pci's ioeventfd=off, the performance is not affected no matter if
guest initializes it or not.

In my test, there are 12 fds to poll in qemu_poll_ns before loading guest
virtio_console.ko, whereas 76 once modprobe virtio_console.

Looks like the ppoll takes more time to poll more fds.

Some trace data with systemtap:

12 fds:

time  rel_time      symbol
15    (+1)          qemu_poll_ns  [enter]
18    (+3)          qemu_poll_ns  [return]

76 fd:

12    (+2)          qemu_poll_ns  [enter]
18    (+6)          qemu_poll_ns  [return]

I haven't looked at virtio-serial code, I'm not sure if we should reduce the
number of ioeventfds in virtio-serial-pci or focus on lower level efficiency.

Haven't compared with g_poll but I think the underlying syscall should be the
same.

Any ideas?

Fam


> 
> > >So, I think it has no business with legacy interrupt mode, right?
> > >
> > >I am going to observe the difference of perf top data on qemu and perf kvm 
> > >stat data when disable/enable virtio-serial in guest,
> > >and the difference of perf top data on guest when disable/enable 
> > >virtio-serial in guest,
> > >any ideas?
> > >
> > >Thanks,
> > >Zhang Haoyu
> > >>If you restrict the number of vectors the virtio-serial device gets
> > >>(using the -device virtio-serial-pci,vectors= param), does that make
> > >>things better for you?
> 
> 
> 
>               Amit
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to address@hidden
> More majordomo info at  http://vger.kernel.org/majordomo-info.html