[sheepdog-users] [sheepdog] Qemu Live Migration and Cache

[Bastian Scholz]

Hi Yuan,

Am 2012-06-07 11:38, schrieb Liu Yuan:
> Without object cache, the answer is NO, the data will be known to the
> cluster as soon as the request is completely.
>
> With object cache enabled, probably 'nope' again, but I am not 100%
> sure, live migration mechanism from QEMU will issue a flush request 
> to
> sheep and block for it, then will migrate memory to other machine. 
> For
> the last try with live migration, it did the right thing for me, but
> you'd better test it with your own. When I have time, I'll try it 
> again
> and give you a solid answer.so

Is there a kind of "standard" test for checking this kind of
cache consistency?

I try it with a virtual machine, running Flexible I/O tester
fio (with randrw, verify) inside and a script, migrating this
virtual machine from host A to B and back in a loop...

I have made only a few short checks at the moment, but fio
show no verify failure, whatever cache strategy is used.

If I increase fio blocksize, the migration takes longer, I
guess, because kvm/qemu needs more time for cache flush. If
blocksize is too big, migration stops at all as long as fio
runs, after finishing, migration works as expected again.

Cheers

Bastian