Re: [Qemu-devel] VM I/O performance drops dramatically during storage migration with drive-mirror

[Stefan Hajnoczi]

On Mon, May 28, 2018 at 06:17:10PM +0800, Chunguang Li wrote:
> Hi, everyone.
> 
> 
> 
> 
> Recently I am doing some tests on the VM storage+memory migration with 
> KVM/QEMU/libvirt. I use the following migrate command through virsh: "virsh 
> migrate --live --copy-storage-all --verbose vm1 
> qemu+ssh://192.168.1.91/system tcp://192.168.1.91". I have checked the 
> libvirt debug output, and make sure that the drive-mirror + NBD migration 
> method is used.
> 
> Inside the VM, I use an I/O benchmark (Iometer) to generate an oltp workload. 
> I record the I/O performance (IOPS) before/during/after migration. When the 
> migration begins, the IOPS dropped by 30%-40%. This is reasonable, because 
> the migration I/O competes with the workload I/O. However, during almost the 
> last period of migration (which is 66s in my case), the IOPS dropped 
> dramatically, from about 170 to less than 10. I also show the figure of this 
> experiment in the attachment of this email.
> 
> 
> 
> 
> I want to figure out what results in this period with very low IOPS. First, I 
> added some printf()s in the QEMU code, and knew that, this period occurs just 
> before the memory migration phase. (BTW, the memory migration is very fast, 
> which is just about 5s.) So I think this period should be the last phase of 
> the "drive-mirror" process of QEMU. So then I tried to read the code of 
> "drive-mirror" in QEMU, but failed to understand it very well.
> 
> 
> 
> 
> Does anybody know what may lead to this period with very low IOPS? Thank you 
> very much. 

IOPS dropped from 170 to less than 10.  That could be because QEMU or
the storage device is slow at completing requests due to the other
activity (drive-mirror).  But it could also be because the guest simply
isn't submitting many I/O requests!

So I think the first step is to determine how much I/O the guest is
submitting.  There are several ways of doing this.  You can enable the
virtio_blk_handle_write and virtio_blk_handle_read trace events if you
are using virtio-blk (see docs/devel/tracing.txt).  Or you could use
iostat(1) inside the guest to observe the number of completed requests +
queue size.

If the guest is not submitting the expected number of requests then
you'll need to investigate why Iometer is starved.

> Some details of this experiment:
> The VM disk image file is 30GB (format = raw,cache=none,aio=native), and 
> Iometer operates on an 10GB file inside the VM. The oltp workload consists of 
> 33% writes and 67% reads (8KB request size, all random). The VM memory size 
> is 4GB, most of which should be zero pages, so the memory migration is very 
> fast.
> 
> 
> 
> 
> --
> Chunguang Li, Ph.D. Candidate
> Wuhan National Laboratory for Optoelectronics (WNLO)
> Huazhong University of Science & Technology (HUST)
> Wuhan, Hubei Prov., China
>

signature.asc
Description: PGP signature