Re: [PATCH v4 2/2] net/colo-compare.c: handling of the full primary or secondary queue

[Derek Su]

Hello, Zhang and Lukas

Sure, after my re-test, the performance is hurt. Will update it later.

By the way, could I also move the "error_report("colo compare
primary/secondary queue size too big, drop packet");" to trace?
The use of error_report is a little strange and make a flood in log.

May I  also make "MAX_QUEUE_SIZE"  be user-configurable in this series?

Thanks,
Derek Su




Zhang, Chen <address@hidden> 於 2020年4月9日 週四 下午2:59寫道：
>
>
>
> > -----Original Message-----
> > From: Lukas Straub <address@hidden>
> > Sent: Thursday, April 9, 2020 3:19 AM
> > To: Derek Su <address@hidden>
> > Cc: address@hidden; address@hidden; address@hidden;
> > address@hidden; address@hidden; Zhang, Chen
> > <address@hidden>; address@hidden
> > Subject: Re: [PATCH v4 2/2] net/colo-compare.c: handling of the full primary
> > or secondary queue
> >
> > On Sat, 28 Mar 2020 20:46:46 +0800
> > Derek Su <address@hidden> wrote:
> >
> > > The pervious handling of the full primary or queue is only dropping
> > > the packet. If there are lots of clients to the guest VM, the "drop"
> > > will lead to the lost of the networking connection until next
> > > checkpoint.
> > >
> > > To address the issue, this patch drops the packet firstly.
> > > Then, do checkpoint and flush packets.
> > >
> > > Signed-off-by: Derek Su <address@hidden>
> >
> > Hello,
> > I had a look at this again and did some benchmarking.
> > First just qemu 5.0-rc1 with my bugfixes
> > ( https://lists.nongnu.org/archive/html/qemu-devel/2020-
> > 04/msg01432.html ) Then qemu 5.0-rc1 with my bugfixes and this patch
> > series.
> >
> > This commit hurts performance too much:
> > Client-to-server bandwidth falls from ~45.9 Mbit/s to 22.9 Mbit/s.
> > Server-to-client bandwidth falls from ~6.3 Mbit/s to just ~674 Kbit/s.
> > Average latency rises from ~197ms to ~397ms.
> >
> > Meanwhile the packet loss without this commit is negligible, only 1-2 ping
> > packets got lost during each test run.
> >
> > Instead I think we should just turn the error message into a trace so it
> > doesn't flood the logs.
>
> We re-test this patch, Lukas is right.
> Sorry for the original idea, looks like it did not show better performance in 
> the test.
> Agree with Lukas's comments. Derek, can you please change it?
>
> Thanks
> Zhang Chen
>
>
> >
> > Regards,
> > Lukas Straub