Re: Extreme Linux

[Drake Diedrich <Drake.Diedrich@anu.edu.au>]

On Sun, Nov 22, 1998 at 10:49:49AM +0000, Andrew Martin Adrian Cater wrote:
> There's an interesting thread over on the Extreme Linux and Beowulf lists.
> 
> I need help and support from the other Debian developers, particularly
> any using Debian for clusters.  There is talk of reorganising the Extreme
> Linux CD to include .tar.gz / .debs / .rpms but to be a minimal drop in
> addition to a standard distribution rather than simply a slightly altered
> RedHat 5.0.
> 
> Please, if you're interested, read the threads following on from
> Robert Brown's "Rant".  I've expressed an interest in helping get .debs
> made and running as has Martin Wheeler - see http://www.startext.co.uk/beowulf.

   I just read the thread, but I'm not sure how to contribute to it.

   Our pvm package produces both static and dynamic libraries.
I believe the EL PVM only has static libraries.  For compatibility it would
be beneficial for us if they had a dynamic library with the same SONAME.
   The povray we ship has modified PVMPOV patches applied.  The
modifications involved making all PVM code #ifdef'd so a non-PVM binary
could be produced from the same sources.  I'm not sure if EL is interested
in this capability as PVM is a pretty standard library on any cluster.
   EL needs a queueing system, but the non-commercial distribution clause in
DQS (our only queueing system at present) probably precludes their use of
it. PBS looks like the Beowulf favorite, but it's been in beta to U.S.-only
sites for years.  Who knows if there will ever be a public license.  I wish
GNQS was further along, but can't dedicate the time needed to give it all
the capabilities it needs for my use when DQS is almost DFSG-free and
currently sufficient.

   Debian already distributes most everything EL has, and several things EL
is missing (eg. a queueing system).  EL might be able to make good use of
some of our software though.  Our PVM package was modified to produce both
static and shared libraries.  Our Povray source produces both PVM and
non-PVM binaries (the upstream pvmpov patches are PVM-only).  Thanks
particularly to John Lapeyre Debian already has more scientific packages
than any other distribution.  EL could scavenge his work.  Scientific
packages are nasty to compile.
   We could use a DFSG compliant parallel renderer and queueing system
though: DQS and Povray both fail due primarily to commercial restrictions. I
don't think EL is that interested in pursuing the free software angle when
there are more capable non-DFSG alternatives, particularly the U.S.-only
PBS.  GNQS (GPL) is apparently a nice queueing system for a dedicated pool
of non-parallel compute boxes, but would need a lot of work to drive a
non-SMP cluster.  If you've got two boxes (any size) and a lot of time on
your hands I could give you a wishlist.  :)


> 
> Neither of us are currently running clusters :( but both of us are keen to
> see that Debian doesn't get ignored here.

   What Debian really needs to become the prefered cluster distribution is
non-interactive postinsts.  EL isn't in a position to give us this, and
we've already got everything else we need, even if some of it is non-free.
   An automatic 'N' to the config-file handler would allow a cluster admin
to predistribute most config files before upgrading a package, cutting out
about 90% of the interaction with well-chosen packages.  For really big
farms nothing short of 100% would be sufficient though.
   Perhaps a policy of checking for existence of an /etc/batch-install
before doing anything interactive (update-mime, config-file handling,
postinst reads), and choosing the simplest alternative when in batch mode
would do.  The work on batch installs seems to have stalled (unless it's on
a list I'm not subscribed to- debian-dpkg is all bug reports lately).  The
configuration database proposed a few months ago may be too complex to do
when so few need it. I've been meaning to try some simple batch changes, but
can't on my workstation cluster (users) and my 486 mini-cluster is
completely disassembled at the moment.
   When a non-interactive method exists we should write a Cluster HOWTO. 
The DQS README.Debian would be a good start, as it already covers a few
related clustering issues (NFS, NIS, automounting, firewalling). 
Documentation is another missing feature of Extreme Linux.
   So ... I think we're on track to take over the Beowulf market.  The only
reason I can think of to follow up a non-Debian distribution discussion on
the beowulf lists is marketing, and we're not quite ready to market to large
clusters yet.  When we are ready, I think the big clusters will come to us
without any prodding for the enormous pool of high quality cluster-ready
apps.

-- 
Dr. Drake Diedrich, Research Officer - Computing, (02)6279-8302
John Curtin School of Medical Research, Australian National University 0200
Replies to other than Drake.Diedrich@anu.edu.au will be routed off-planet