Re: [Qemu-devel] [PATCH] augment info migrate with page status

[Anthony Liguori]

Dor Laor wrote:
> >  static ram_addr_t ram_save_threshold = 10;
> > +static ram_addr_t pages_transferred = 0;
> >   
>
> It would be nice to zero pages_transferred each migration operation.
> ram_save_threshold is really to small. From Uri's past measurements, 
> as value of 50 is a
> better suite. Alternately it can be parametrized by the monitor command.
>
> In general there is small drawback in the current approach:
> The way bandwidth is capped,  iirc,  in every second you start 
> consuming migration
> bandwidth. If the bandwidth allocation was consumed after 100msec, 
> you'll wait 900msec.
> In this period, mgmt app reading the ram_save_remaining will notice 
> that migration does
> not progress and might either increase bandwidth or stop the guest.
> That's why #of no-progress-iteration has advantage.

If I were implementing this in libvirt, here's what I would do:

B = MB/sec bandwidth limit
S = guest size in MB
C = some constant factor, perhaps 4-5

T = S / B * C

1) Wait for T seconds or until migration completes.
2) If timeout occurred:
  a) M = actual transfer rate for migration in MB/sec
  b) If M < B, T1 = S / M * C
  c) T = T1 - T
  d) If T <= 0, migration failed
  d) else goto 1

Basically, this institutes a policy that a migration must complete after 
transferring C * guest_size amount of data.  It adjusts for observed 
bandwidth rate vs. capped.  It makes sense from an administrative 
perspective because you are probably only willing to waste so much 
network bandwidth on attempting to migrate.  Obviously, C and B are 
tunables that heavily depend on the relatively priority of the migration.

Whether you force a non-live migration after failure of a live migration 
is an administrative decision.

--
Regards,

Anthony Liguori