[Beowulf] scheduler policy design

[Mark Hahn]

> And I just spotted that KVM supports migration 
> (http://kvm.qumranet.com/kvmwiki/Migration) and is integrated in the 2.6.21 
> kernel.

I haven't quite figured out whether this is going to wind up acting 
differently from, say, Xen on an un-kvm-enabled kernel.  (assuming a 
normal cluster environment, where no swapping is happening, and jobs
are either entirely compute-bound, or interested mainly in net IO.)

> This would allow the scheduler to schedule on the actual resources being used 
> by an application and suspend and/or migrate a job whenever necessary.

as far as I can tell, the scheduler's logic wouldn't really change much:
instead of starting a bare job, it starts the job in its own container.

> Would'nt such an approach be far more superior (in terms of throughput) and 
> much easier to use ?

the KVM is still a complete OS instance, afaikt.  I don't really understand 
whether this winds up being inefficient in memory - for instance, does the 
kvm have its own pagecache, or all such things merely shared from the host 
kernel instance?

has anyone tried using UML to containerize jobs on a cluster?  I recently 
ran across a moderately old howto on using OpenSSI and UML.