Hi!


> ... which is to say that the client programmer better know the set up
> of the cluster and start his processes on machines that can handle
> them. The onus is on the client, not on the cluster, to distribute
> processes intelligently. Obviously, it's not possible to completely
> relieve that, but it is possible to make the cluster a little smarter
> about it.

This is what the project I am working on does.  My strategy is a little
bit of a variant on Gabrials, but enough the same in this context.

> (The point is that you can specify a specific node for each parallelized
> process you fork off, so if you want something back, you'd better pick
> a fast node yourself rather than letting Beowulf pick for you. PVM
> behaves in a similar manner.)

This is kind of interesting in the sense that I bet this discussion happened
some years ago in Sequent labs when they were figuring out how to mix and
match intel processors in their SMP boxes.  From what I have seen there
though is that processor 0 is always the slowest card.  I never understood
that.  If you had a model B board (dual 386/16) and many 'F''s  (Pentium/60)
the 386 would always be the boot processor and if you had a single runnable
process, it would always get scheduled on the 386.  

So now we have moved into nodes.  I hope this is done right this time.

Cheers



Andy

       ~ g r @ eclipsed.net