Re: Bug#97336: ITP: dsh -- dancer's shell, or a distributed shell.
On Tuesday 22 May 2001 11:34, Junichi Uekawa wrote:
> Eray Ozkural (exa) <erayo@cs.bilkent.edu.tr> cum veritate scripsit:
>
> Hi Eray,
>
> I think we should start with the first part;
> managing the list of machines centrally regardless of
> what use it might have.
>
> Then, go for HA/HPC related stuff.
>
> However, I can't really think up of anything HPC specific.
>
Sure. There are issues specific to HPC.
1. Kinds of programs related to HPC: message passing programs,
HPC libs...
2. Beowulf installation: Inside a beowulf you go mostly for a private
network and don't care about all the security mumbo jumbo. Performance
over everything.
a. private net
b. program choices: benchmarks, libs, devel stuff, math...
c. naming schemes: how do you want it?
3. Tools to do synchronous operations over all nodes. HPC specific stuff:
clean up /scratch partitions, reset daemons, etc.
4. Allocate temporal/spatial subsets of the cluster to certain individuals,
queue parallel programs to subclusters...
5. Topology matters! People wanna embed nets or optimize for topology, etc.
6. Virtual clusters make sense... (although for general purpose, too)
Well, of course some of these are bogus, but I think you could make the
list arbitrarily long :) What I mean is that one should think of the
operational aspects of an actual beowulf cluster.
Starting with a list of machines is just fine. :)
Thanks,
__
Eray
Reply to: