openmosix clusterin

[Gavin McCullagh]

Hi,

On Mon, 23 Oct 2006, James Call wrote:

> Wouldn't it make sense to make edubuntu run on this type of a system?
> It seems that in theory, the server could be anything with a bunch of
> memory and the terminals would provide all the processing power.  That
> means that when Billy is looking at a web page (super low cpu) Susy
> could be making an object in Blender (very high cpu).  It seems that
> in a lot of developing countries (read Mexico) you could use all these
> old P3s to make some super sweet clusters.

My old research group used openmosix a bit a couple of years back to
distribute load of scientific simulations across cluster nodes.  

There may be an argument for this in the LTSP situation, but I'd say it'd
be fairly limited.

 - Stability is all important. If you destabilise your LTSP server(s), the
   cost really has outweighed the benefit.

 - There is a cost to managing a cluster and migrating processes -- mosix
   works best with lots of powerful nodes.  If your thin clients are
   borderline, slow machines, it's possible that they won't contribute
   enough to the cluster to make them worth using as nodes. 

 - If on the other hand, your thin clients are powerful machines (P3s), you
   should probably be looking into using them as diskless clients, which
   run their own processes.  This will be simple, manageable and stable and
   won't load the server.  It may not spread the load quite as well, but if
   your thin clients are mostly powerful machines, you're probably not a
   school on a very tight hardware budget.  Then the server or a few powerful
   workstations can be used for blender et al.

 - In the days we did this, losing a node meant losing the processes which
   had migrated to that node.  If that's still the case and your nodes are
   all over the network, there's the possibility for one user to kill
   another uses processes accidentally by taking down a client.

 - You'll need a pretty fast network for this too.

I'm not saying this could never work or be beneficial, but instinctively I
would wonder if it would really be worth the large effort involved to
maintain and bugfix mosix clusters as a core part of edubuntu if it's
benefits were only really seen by a minority of users.

This guy's justification is:

  "we had to buy diskless new computers for our Xterminals. I then see that
  the cheapest processor was an Athlon 850 (June 2001) and that the less
  possible memory was 64 Mo. What a pity to have an Athlon 850 only used
  for X!"
	http://www.lpmo.edu/~daniau/ltsp-mosix/

but diskless clients running their own processes solves this problem very
adequately, probably with lower network load.

Gavin