Beowulf [& VMWare]

[Timothy H. Keitt]

On Thu, 2002-11-07 at 13:45, Donald Becker wrote:
> On 31 Oct 2002, Timothy H. Keitt wrote:
> > On Thu, 2002-10-31 at 13:30, Donald Becker wrote:
> > > 
> > > As most list readers know, I consider Mosix to be an academically
> > > interesting system and its transparent process migration an excellent
> > > touchstone to compare process migration approachs.
> > > 
> > > Mosix is not a good system for scalable or performance-oriented
> > > clustering, and I agree with your conclustion that Mosix+VMware
> > > application serving is not a good match.
> > 
> > VMware aside, I'm not aware that openmosix precludes the use of mpi or
> > pvm or standard batch processing. Are these not good solutions for
> > scalable or performance-oriented clustering?
> 
> Mosix
>    moves the migration policy into the kernel,
>    has (deliberately) slow job migration to avoid scheduling hot-spots
>    uses in-kernel communication forwarding
> 
> While these allow unaware, and unmodified, applications to be migrated,
> they are exactly the wrong decisions for MPI and PVM applications.
> MPI and PVM
>    explicitly and simultaneously starts a known number of subprocesses,
>    knows where those subprocesses should be, and
>    builds new, explicit communication paths
> 
> The total effect is that Mosix takes much longer to start MPI/PVM jobs,
> and has much higher overhead.
> 
> The features Mosix provides are
>   single point application updates
>   single process space view
> Scyld Beowulf provides these at much lower CPU overhead.

My point was that you don't _have_ to use the automatic process
migration on an openmosix cluster, you can do whatever load balancing
you want in your mpi (or other) code. Clearly, the "scalable
applications" you refer to will require custom code and I'm saying that
that custom code will run on an openmosix cluster (if told not to do
migration for those processes) exactly as on a standard cluster without
openmosix. Unless I'm missing something, there's no cost to having it in
the kernel. And for many people who don't have dedicated resources for
writing custom code, openmosix makes a lot of sense (to me anyway).

Tim

> 
> -- 
> Donald Becker				becker at scyld.com
> Scyld Computing Corporation		http://www.scyld.com
> 410 Severn Ave. Suite 210		Scyld Beowulf cluster system
> Annapolis MD 21403			410-990-9993
> 
-- 
Timothy H. Keitt
The University of Texas at Austin
Section of Integrative Biology
1 University Station C0930
Austin, Texas 78712-0253 USA