[Users] Live migration by using network swap works well?

[Kir Kolyshkin]

On 07/16/2014 02:10 AM, Junliang Li wrote:
> Hi kir,
>
> 	I learn your opinion about live migration by using network swap from
> your presentation (7problems.odp). But right now I check openvz kernel
> and can not find anything about this mechanism (Solution 1).  How is it
> going ? Or totally give up ?

Network swap is a dead end, it should not be used. Not sure it's explained
in the slides, but the problem is simple. While migrating there should
always
be a way to roll back to a good state. The rollback is to start (unfreeze)
a container on the source node, if anything fails during the migration.

In case of network swap, we have already started a system on the
destination, but since not all memory is there yet, we swap it in from the
source. Now, if a network connection between source and destination
breaks, we need to roll back but it's too late as we have already started
something on the destination.

What should be used is memory pre-migration, when a destination system
is "pre-heated" with the memory pages from the source system before
migration.
Currently we don't have it in OpenVZ, but it is already implemented in CRIU
(see http://criu.org/Iterative_migration) and will be used in future OpenVZ
releases.

Kir.