Re: MOSIX

[Eray Ozkural <erayo@cs.bilkent.edu.tr>]

On Sat, 16 Sep 2000 01:33:37 Dariush Pietrzak wrote:

> > 1000s of boxen of course. :)
> it's been said that it can support up to 6500 boxes, but I think that it
> would be way cheeper and faster to buy parallel machine from SGI

Well, no. :) COTS is the way to go, and Beowulf is getting hotter as days
pass. COTS = Commodity Off The Shelf Components. Have a look at www.beowulf.org
to learn about the history of the project. Beowulf machines are among
the TOP 500 fastest supercomputers. (That has a web page, too, but where)
Stay subscribed to this list, I'll post some of my lousy bookmarks.

> Amoeba is quite strange thing, although i think i like it's filesystem
> very much. and i don't like it's licence.
> there is linux version built on Mach microkernel, do you think this would
> be any better? 

What I meant is supporting some kind of message passing, RPC OO, etc. is
a nice thing to have. You're not done with that sure. You'll probably need
higher level software, like languages, runtime systems, etc. This is
one of the places where the distributed / parallel research distinction
starts to disappear.

> 
> > for scientific computing. Linear algebra, mostly. blas, scalapack, etc.
> blacs. yes. but I ain't going to do much scientific computing, am thinking
> about some more general solutions.
> 

Well, go ahead but first check what's been done before. A textbook on parallel
programming will be instrumental. I advise "Introduction To Parallel Programming
- Design And Analysis Of Algorithms by Kumar, Grama, Gupta and Karypis". Those
are the guys we're playing catch-up with :) For newer stuff, you need to see
the research projects which you can access on the internet. You may be interested
for instance in POOMA.

> > People also have tried automatic parallelization of sequential code, but that
> > doesn't work. :)
> Hmm? and what about modern processors? don't they do this and do this
> quite nicely ?
> 

It's basically because sequential code is too sequential :) Traditional languages
don't have nice semantics that are well-parallelizable. :) A language with side-effects
is a nightmare for any parallelizing compiler.

> > be sure that it is [Fortran] going to be viable for at least a couple of decades. :)
> so are assemblers, but I ain't going to do much coding with asm either.

You aren't making the right analogy here. Assembly is low-level, C++ is high level.
In parallel programming, passing messages with C or C++ is low-level, using a data-parallel
language is high level. HPF is one of the (or the) most advanced compilation systems,
considering what's under the hood. So, if you don't like low-level stuff, avoid anything
that exposes architecture specific details.

> > Some physicists like Charm++ also, but I don't find that very efficient.
> what is Charm++?
> 

Charm++ is some extensions to C++ that allow you to do OO message passing. Invoking a remote
member function is like executing a normal member function. It might be nice if you're
doing coarse grained code.

Regards,

-- 
Eray (exa) Ozkural
Comp. Sci. Dept., Bilkent University, Ankara
e-mail: erayo@cs.bilkent.edu.tr
www: http://www.cs.bilkent.edu.tr/~erayo