Re: [Qemu-devel] I/O performance degradation with Virtio-Blk-Data-Plane

[Stefan Hajnoczi]

On Thu, Sep 05, 2013 at 10:18:28AM +0900, Jonghwan Choi wrote:

Thanks for posting these details.

Have you tried running x-data-plane=off with vcpu = 8 and how does the
performance compare to x-data-plane=off with vcpu = 1?

> > 1. The fio results so it's clear which cases performed worse and by how
> >    much.
> >
> When I set vcpu = 8, read performance is decreased about 25%.
> In my test, when vcpu  = 1, I got a best formance.

Performance with vcpu = 8 is 25% worse than performance with vcpu = 1?

Can you try pinning threads to host CPUs?  See libvirt emulatorpin and
vcpupin attributes:
http://libvirt.org/formatdomain.html#elementsCPUTuning

> > 2. The fio job files.
> > 
> [testglobal]
> description=high_iops
> exec_prerun="echo 3 > /proc/sys/vm/drop_caches"
> group_reporting=1
> rw=read
> direct=1
> ioengine=sync
> bs=4m
> numjobs=1
> size=2048m

A couple of points to check:

1. This test case is synchronous and latency-sensitive, you are not
   benchmarking parallel I/Os so x-data-plane=on is not expected to
   perform any better than x-data-plane=off.

   The point of x-data-plane=on is for smp > 1 guests with parallel I/O
   to scale well.  If both those conditions are not met by the workload
   then I don't expect you to see any gains over x-data-plane=off.

   If you want to try parallel I/Os, I suggest using:

   ioengine=linuxaio
   iodepth=16

2. size=2048m with bs=4m on an SSD drive seems quite small because the
   test would complete quickly.  What is the overall running time of
   this test?

   In order to collect stable results it's usually a good idea for the
   test to run for a couple of minutes (e.g. 2 minutes minimum).
   Otherwise outliers can influence the results too much.

   You may need to increase 'size' or use the 'runtime=2m' option
   instead.

Stefan