Native Linux KVM tool v2

Posted Jun 15, 2011 22:10 UTC (Wed) by aliguori (subscriber, #30636)
Parent article: Native Linux KVM tool v2

Hrm, I think the performance results are a bit too good to be true. Looking more closely at the code and the announcement, I can see why:

1) Native KVM tool is using virtio whereas they're launching QEMU with IDE. IDE only supports a single request at a time (this is how the hardware works).

2) Native KVM tool does not make guest data writes stable by default. QEMU does. While they do advertise the FLUSH feature, they don't advertise a write cache setting to the guest. Unfortunately, many modern guests assume storage with non-volatile write caches and don't use FLUSH's properly.

I would expect QEMU configured in the same way that Native KVM tool would be at least as good. I also would recommend that type of configuration for anyone that cared about their data as a default.

Native Linux KVM tool v2

Posted Jun 15, 2011 22:21 UTC (Wed) by aliguori (subscriber, #30636) [Link] (2 responses)

It's worse than I initially thought :-( Writes are unstable but FLUSH isn't currently implemented. So even if a guest does send FLUSH commands correctly, native KVM tool ignores them...

Native Linux KVM tool v2

Posted Jun 15, 2011 22:46 UTC (Wed) by aliguori (subscriber, #30636) [Link] (1 responses)

Nope, misread the code. It does do an fsync() on FLUSH so for guests that support barriers problem, it will work correctly.

So the behavior is equivalent to cache=writeback in QEMU.

Native Linux KVM tool v2

Posted Jun 16, 2011 6:35 UTC (Thu) by penberg (guest, #30234) [Link]

Yup, we worked pretty hard to make sure we don't eat your QCOW2 images for lunch :-).