Scyld, local access to nodes, and master node as compute node

[Brian C Merrell]

To be honest, I agree that this would be the way to go.  But they have
fairly specific requirements.  Also, they want to use the PGI CDK because
of its fortran 90 compiler.  This presents a problem since the PGI stuff
(v 3.1) won't install on the Scyld nodes.  Also, I don't seem to be able
to install any PVM stuff and have it work with the Scyld nodes, although
I've only been tinkering with that so far.  I've started installing a mix
of full nodes and partial (Scyld) nodes to see how I can resolve my
issues, since I would prefer to use the scyld stuff too (the BProc stuff
is pretty slick).  I expect two more days of hard work and maybe a few
hours on top of that of little things.  (BTW, I'm doing this on contract :)

Thanks for everyone's help.  Hope to have either a successful report soon,
but if not then just more questions. :)

-brian

-- 
Brian C. Merrell      P a t r i o t  N e t      Systems Staff
brian at patriot.net    http://www.patriot.net    (703) 277-7737
          PatriotNet ICBM address: 38.845 N, 77.3 W

On Thu, 24 May 2001, Art Edwards wrote:

> For what it's worth:
>
> I have been using a truly serial code on a very high performance cluster that
> supports MPI. For some political and some actual reasons, even though I am
> running seperate serial jobs, I actually run them under MPI. I'm also starting
> to use a Scyld cluster and the analogies are heartening. MPI is part of
> Scyld so I would like to state some obvious things.
>
> 1. You can assign as many nodes as you want using MPI (Read you can assign one)
>
> 2. Installing MPI into existing serial code is very easy. There are about three
> or four calls.
>
> 3. Thanks to Sean Dilda I can now write files to /tmp on each node. I haven't
> tried to read from /tmp files but I'm guessing this will be straight
> forward.
>
> So, rather than work on some system modification that could be clumbsy, why
> not modfiy the research code to take advantage of the features of Scyld?
>
> I think you could use a combination of bash shell scripts and MPI calls to
> accomplish what you say you want. Use the bash script to put the appropriate
> input files on the various nodes and use it to start the mpi job. Then let
> the computational codes open the files on the various nodes and use them in
> the specific calculations.
>
> This is a slightly different paradigm and I know from my own experience that
> one would rather simply keep doing what they know. Once I got use to it there
> was no big deal.
>
> Sorry if I wasted your time.
>
> Art Edwards
> On Thu, May 24, 2001 at 02:07:04PM -0400, Brian C Merrell wrote:
> > On Thu, 24 May 2001, Sean Dilda wrote:
> >
> > > Is there any reason the program itself can't run itself in the special
> > > way they want?  Anything you can do with rlogin or rsh can be done with
> > > bpsh, except for an interactive shell.  However, this can be mimiced
> > > through bpsh.  If you can give me some idea of what they are wanting to
> > > do, I might be able to help you find a way to do it without requiring an
> > > interactive shell.  Scyld clusters are designed to run background jobs
> > > on all of the slave nodes, not to run login services for users on the
> > > slave nodes.
> > >
> >
> > Hmmm.  I guess this warrants some background info.
> >
> > The cluster is not a new cluster.  It was previously built by someone else
> > who is now gone.  The cluster master node crashed, taking the system and
> > most of their data with it.  I am now trying to rebuild the cluster.  The
> > cluster previously used RH6.1 stock and followed more of a NOW model than
> > a beowulf model, although all the hardware was dedicated to the cluster,
> > not on people's desks.  I'm now trying to use Scyld's distro to bring the
> > cluster back up.  I'm pretty happy with it, and managed to get the master
> > node up with a SCSI software RAID array, and a few test nodes up with boot
> > floppies.  Seems fine to me.  BUT....
> >
> > There are three reasons that they want to be able to rlogin to the
> > machines:  1) first, there are a number of people with independent
> > projects who use the cluster.  They are used to being able to simply login
> > to the master, rlogin to a node, and start their projects on one or more
> > nodes, so that they take up only a chunk of the cluster.  2) Also, at
> > least one researcher was previously able to and wants to be able to
> > continue to login to separate nodes and run slightly different (and
> > sometimes non-parallelizable) programs on his data.  3) ALSO, they have
> > code that they would rather not change.
> >
> > > It is possible to use BProc with a full install on every slave node
> > > however this reduces a lot of the easy administration features we've
> > > trying to put into our distro.
> > >
> >
> > I just set this up, and realize what you mean.  I had to statically define
> > IP addresses, users, etc.  At first it wasn't a pain, but I realized after
> > the first two that doing all 24 would be.  Even though it is now possible
> > to rlogin to different nodes, it wasn't what I was hoping for. I imagine
> > it will be particularly unpleasant when software upgrades need to be
> > performed. :(
> >
> > I'm still hoping to find some happy medium, but I'm going to present these
> > options to the group and see what they think.  The problem is that they
> > are mathematicians and physicists, not computer people.  They really don't
> > want to have to change, even though it seems to be the same.
> >
> > Also one thing I'm still trying to find a solution to: how can the nodes
> > address each other?  Previously they used a hosts file that had listings
> > for L001-L024 (and they would like to keep it that way) I guess with the
> > floppy method they don't have to, because the BProc software maps node
> > numbers to IP addresses,
> >
> > -brian
> >
> >
> > _______________________________________________
> > Beowulf mailing list, Beowulf at beowulf.org
> > To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>