[Users] virtio-scsi Windows/Linux VM performance

Markus Stockhausen stockhausen at collogia.de
Thu Sep 19 11:28:03 UTC 2013


Hello,

a few weeks ago I started investigating ovirt as an alterantive to a well
known commercial virtualization center. Up to now I got great support
through this mailing list. Thanks for that. With all the technical things
sorted out I'm getting slowly to the performance related stuff.

Lets start with a small explanation to the test landscape. I have two
hypervisor machines (Intel Nehalem generation) that are attached to
a NFS server via infiniband IPoIB connected mode (MTU 33000). They 
run Fedora 19 (kernel 3.11) and through newest repositories qemu 
version 1.6.0 - 7.fc19.

To get an idea about the overhead of the VM disk layer I run a not very
scientific direct I/O read test on the hypervisor onto a image file that
resides on the NFS share. Access pattern is random, I use one thread,
blocksize is 1K and the file (image that belongs to a VM) is fully cached 
on the NFS server. So no disk activity is going on in the NFS server but 
block reads are always sent over the network. The numbers read 13 
MB/sec or 13000 IOs/sec. Just take that as a very synthetic reference 
value for the latency of the NFS layer.

Now I relocate my test into one of the running VMs. Again a Fedora 19
that uses virtio-scsi based disk images on the NFS. The IOs drop to 
9.5MB/sek or 9500 IOs/sec. I understand that virtualization costs 
performance so I'm still statisfied.

With Windows 7 VM and the virtio-0.1-65 driver I get a real bad experience.
Only 2MB/sek or 2000 IOs/sec left. That is not what one would expect. 

I read through articles/papers about the different approaches in the last
years to increase VM IO performance: virtio-data-plane, virto-blk, ...
But I'm not yet searching to increase the overall performance I'm just
trying to get biggest mistakes sorted out. With the same devices and 
on the same platform I expect the performance of both VMs to be in a 
similar range and not a difference by a factor of 4-5.

Any ideas?

Markus
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: InterScan_Disclaimer.txt
URL: <http://lists.ovirt.org/pipermail/users/attachments/20130919/b8ee94a6/attachment-0001.txt>


More information about the Users mailing list