[Beowulf] SGI to offer Windows on clusters

[Mikael Fredriksson]

Mark Hahn wrote:
> ........ what I'm interested in is whether Mosix
> functions well in a more-than-toy cluster (say, at least 100p).
>
> I guess I also am uncertain where Mosix's competitive advantage lies.
> my experience is that a serial-job workload is so undemanding that 
> migration is unnecessary - it's mainly for parallel jobs where you 
> really want migration.  (in SHARCnet, there is a strong correlation
> for serial jobs to also be relatively short and quite small in memory
> use.  not _all_ are, of course, but certainly most.)

Well, one big advantage is that the cluster will be open to a much wider
range of people who can do some programming but are not parallell
programming experts.  This in it self is a very powerful argument when
you want to motivate the funding you need to buy a (large) cluster.
Several departments may have to share the cluster and not all have the
skill, time or the money to develop specialized parallell software for
their applications.  But what they usually have is a *good* knowledge to
handle, for instance, math programs e.g. MATLAB.  So if they do, for
example, a type of "loop unrolling" in their MATLAB scripts (and also
ingrease detail), they can then start say 100 instances of MATLAB and
let them migrate.  And if necessary, they can implement simple types of
syncronization barriers.  This is not optimal HPC, but it is a way to
adapt to the reality for many organisations.

So, if we return to the SGI/Windows cluster case, with a MOSIX extension
there is a possibility that the cluster can also be used to run e.g.
serial administrative jobs.  Again, with a MOSIX extension you have a
extra argument to motivate the purchase (or selling) of a *good* cluster.


MF