libvirt users - Feb 2014 - Re: [Qemu-devel] Adjust disk image migration (NBD)

On Fri, Feb 14, 2014 at 11:58:56AM +0100, Michal Privoznik
wrote:
> On 14.02.2014 10:40, Joaquim Barrera wrote:
> > Hi all,
> > 
> > As I am doing some tests with qemu, I realized that the way it does 
> > 'migrate -i tcp:DEST:444' is not the same as 'libvirt
migrate
> > --copy-storage-inc'. Basically qemu uses the same stream as RAM 
> > migration and libvirt takes advantage of NBD transfer.
> > 
> > With virsh migrate-setspeed I observed that one can only control the 
> > transfer throughput of RAM, but not disk synchronization. At least
this
> > is what I can see in bmon when doing a migration with incremental
copy.
> > 
> > The question is: Am I missing something or it is not implemented?
> > 
> > Thank you guys!
> > 
> 
> I think this is actually a qemu bug. Libvirt passes the correct values:
> 
> 
> 2014-02-14 10:52:08.010+0000: 27701: debug : qemuMonitorIOWrite:504 :
QEMU_MONITOR_IO_WRITE: mon=0x7f06cc00ea20
buf={"execute":"drive-mirror","arguments":{"device":"drive-virtio-disk0","target":"nbd:masina:49153:exportname=drive-virtio-disk0","speed":1048576,"sync":"full","mode":"existing"},"id":"libvirt-15"}
> 
> 
> ...
> 
> 2014-02-14 10:53:51.169+0000: 27701: debug : qemuMonitorIOWrite:504 :
QEMU_MONITOR_IO_WRITE: mon=0x7f06cc00ea20
buf={"execute":"migrate_set_speed","arguments":{"value":1048576},"id":"libvirt-221"}
> 
> 
> 2014-02-14 10:53:51.204+0000: 27701: debug : qemuMonitorIOWrite:504 :
QEMU_MONITOR_IO_WRITE: mon=0x7f06cc00ea20
buf={"execute":"migrate","arguments":{"detach":true,"blk":false,"inc":false,"uri":"fd:migrate"},"id":"libvirt-223"}
> 
> 
> However I observe what you do - disk migration is not shaped, while
internal state is.
Thanks for raising this.

I noticed that mirror_run() does not throttle the first loop where it
populates the dirty bitmap using bdrv_is_allocated_above().  The main
copy loop does take the speed limit into account but perhaps that's
broken too.

Paolo, Jeff: Any ideas?

Stefan

> Thanks for raising this.
> 
> I noticed that mirror_run() does not throttle the first loop where it
> populates the dirty bitmap using bdrv_is_allocated_above().
This is on purpose.  Does it causes a noticeable stall in the guest?
> The main
> copy loop does take the speed limit into account but perhaps that's
> broken too.
Yeah, it looks broken.  Each iteration of the loop can write much more
than sectors_per_chunk sectors, but here:

            if (s->common.speed) {
                delay_ns = ratelimit_calculate_delay(&s->limit,
sectors_per_chunk);
            } else {
                delay_ns = 0;
            }

the second argument is fixed. :/

Paolo

On 24/02/14 23:26, Paolo Bonzini wrote:
>> Thanks for raising this.
>>
>> I noticed that mirror_run() does not throttle the first loop where it
>> populates the dirty bitmap using bdrv_is_allocated_above().
> This is on purpose.  Does it causes a noticeable stall in the guest?
>
>> The main
>> copy loop does take the speed limit into account but perhaps that's
>> broken too.
> Yeah, it looks broken.  Each iteration of the loop can write much more
> than sectors_per_chunk sectors, but here:
>
>              if (s->common.speed) {
>                  delay_ns = ratelimit_calculate_delay(&s->limit,
sectors_per_chunk);
>              } else {
>                  delay_ns = 0;
>              }
>
> the second argument is fixed. :/
>
> Paolo
>
Thanks for the answer. Something is still not clear to me. Are we in 
front of a bug (that means, something that could be fixed) or is this 
behaviour somehow expected for some reason? More and more tests I am 
doing, I get allways the same throughput chart: unlimited bandwidth when 
syncronizing the disk, and smooth bandwidth limit when migrating RAM.

Joaquim