Re: MOSIX

[Eray Ozkural <exa@ttnet.net.tr>]

Fabrice Lorrain wrote:
> 
> 
> >  > Something to look for would be compatibility with current kernel
> >  > versions I guess. Do patches work all right for 2.2.14, etc?
> >
> >  AFAIK, they do.  They should work with 2.2.15 and 2.2.16, too.
> Not completely, work with 2.2.15 but doesn't compil with 2.2.16 AND
> gcc-2.95.2-12 (last night tests). The answer from Ariel Eizenberg from
> the mosix team :
> "MOSIX compiles flawlessly with gcc 2.7.2 and egcs 1.1.2 (gcc-2.91)."
> Some other guy has the same pb (debian 2.2, gcc-2.95.2) and swapping to
> 2.7.2 solve it.

That's what I meant. It was mentioned that getting mosix to run with the
latest gcc and kernel versions was not easy.

> >  The beowulf community?  What's that?  h/j.
> Doesn't know that either ... And what's that stuff on free software and
> MOSIX ???
> >From what I've seen in the last version (Mosix-0.97.6), every thing is
> GPLed ...
> It seems to be "free" enough for me.

All right. Beowulf community is us. :) I didn't refer to some concrete
organization or company. Hope you don't mind :)

In other words, beowulf people are generally not very keen on using MOSIX
patches. Why not? Because if you want HPC, you run message passing programs.
If you want storage, or you know the stuff database people do, then you
wouldn't be needing MOSIX patches still, it seems. I have never had a
lasting interest in databases, so my opinions would be from HPC and distributed OS
perspective.

Now, linux is a monolithic kernel, so it isn't very well suited to distributed/open
system design in the first place. (Open doesn't mean open source... ;)
Those modules are just not the thing. You know, the holy grail of a distributed
system is one that gives you a single system image, it's got no "home node", it's
rootless, dynamic, blah blah. Even if so, MOSIX *is* free software and right
here. So, you don't even have to retain the name MOSIX. :) Go ahead,
fork the latest MOSIX release, and make it suit your needs. :) OTOH, what I
would first like to see would be some of kind of message passing subsystem
that works fast. Then, I'd like to see some shared object support, "industrial
strength" distributed memory and filesystem implementation. And what else
do we have? Oh, the nasty "process migration" issues, and those other things.
Of course some of these things are being addressed by MOSIX, but I guess there
is much left to do. :)

Caveat Emptor: these things aren't trivial. I personally think no major problem
has been solved by the parallel programming or distributed OS communities.
Years ago, when I first heard HURD, I thought that it would be a distributed OS
(yeah it was MACH after all!), and provide POSIX services on top. Now I see why
they didn't go that route :)

Did you realize that there is not even one usable distributed file system? Why not?
The only distributed file system I used was on Windows NT, which crashed and ruined
our whole project. I don't know. I feel like I could bundle up one in 2 weeks, but then
is Linux the right thing? My "coder sense" says that you need to go all the way down
and start with a microkernel and a very clean design for such capability. Linux is
probably not the right answer to every question out there.

BTW, I think a true distributed OS will need to deviate a lot from C/UNIX
tradition. You know, traditions are there to smash. That dream OS, I hope it will
be free :), will be able to run multiple parallel programs, and provide a single
system image on massive clusters [ May I say > 10K nodes? ]

Thanks,

-- 
 ++++-+++-+++-++-++-++--+---+----+----- ---  --  -  - 
 +  Eray "exa" Ozkural                   .      .   .  . . .
 +  CS, Bilkent University, Ankara             ^  .  o   .      .
 |  mail: erayo@cs.bilkent.edu.tr                .  ^  .   .