Guys,

 

I have 2 SSDs running as a cache pool for my docker and VMs, my VMs use qcow2 for snapshotting. My windows 10 VM is seabios, QT35, passthrough GTX 770, 12800MB Ram, and the CPU has 6 of 8 cores selected in host passthrough mode. The primary vdisk was tried at both /mnt/cache/vstorage and /mnt/user/vstorage (currently set here).

 

After a fresh install tests show me with an average of 25MB/s write, and 300-500MB/s read. I am previously familiar with kvm, so I installed lubuntu with virt-manager, and configured the primary disk to be cache mode: none and IOMode: native, this got me to 70-90 MB/s write, 300ish MB/s read.

 

I must be missing something, I can't imagine this is the performance everyone is getting, I assumed that as a vm direct on the cache drive, I would get near real speeds (minus to btrfs overhead).

 

Can anyone help me troubleshoot?

 

Thanks,

Tom

I know that the KVM documentation specifically mentions that .qcow2 virtual disks hosted on a BTRFS file system may not perform well. Not sure if there are any settings that can help with this.

Itimpi,

 

Thanks, I never made that connection in my head, all i cared was that I had snapshots during this testing process of a win8->win10 upgrade. I am going to create a copy of the disk and convert this one to raw and see what happens.

 

Thanks,

Tom

So after converting the image to raw, updating the xml to reflect such and starting up, still only getting 90 MB/s write. Anything else I can check, can anyone else validate their performance in a similar setup?

 

Thanks,

Tom

So after converting the image to raw, updating the xml to reflect such and starting up, still only getting 90 MB/s write. Anything else I can check, can anyone else validate their performance in a similar setup?

 

Thanks,

Tom

 

If you're familiar with virt-manager, you might want to try playing with IOThreads as that may also help.  Even with RAW image format, you're still doing this as a virtual disk (not physical), so there is a bit of overhead.  Getting full SSD performance to a virtual disk which is a filesystem inside a loopback image on top of another filesystem on top of a physical block device may not be possible.

Jonp,

 

Thanks for the reply, I'll test IOThreads next. I was not hoping for 100% pass through speed, I was hoping for something more than 25%. I am not sure btrfs adds alot of overhead but diskspeed reports the two drives as:

 

/dev/sdb (Cache): 472 MB/sec avg

/dev/sdc: 526 MB/sec avg

 

Thanks,

Tom

Getting better with IOThreads: 111.45 MB/s write on the vm

did you manage to fix/improve the issue? I'm very interested on getting almost bare-metal speeds too!