mpich and scheduling issues

Alexis Zubrow azubrow at galton.uchicago.edu
Tue Oct 29 11:24:16 PST 2002


Hi all.

I was hoping to get some advice on different scheduling scenarios for
dealing with multiple users on small cluster.  I have done some reading
of howto's and searching the web, but I wanted to poll your collective
experience on how various schemes work in the real world. I know that this
is very hardware and application specific, so I included a small summary
of our system:

hardware-
	4 nodes (dual Athlon MP 1800+)
	gigabit connection 
	may expand to as many as 8 nodes

Number of users:
	presently around 8 active users
	around 30 infrequent to very infrequent users

Applications-
	majority are serial applications
	alot of R, splus, matlab calculations
	aperiodically some intense parallel programs (e.g. MM5, parallel
	version of CMAQ, possibly some parallel development)

present scheduling scheme-
	self-monitoring (i.e. politeness)


Some ideas I had:

1.) Make the head node non-scheduled.  Put a scheduling system (openPBS,
Grid Engine, etc) on nodes2 through N.  People can do their debugging and
compiling on the head node, and they can run their more intense processes
through the scheduler.

2.) Mosix on all nodes.  For serial users they will not have to worry
about anything.  The openMosix documentation suggests that mpi and mosix
work great together.  But, I'm a bit concerned about what happens when you
mix mpi and serial processes.  For example, MM5 tries to run at 100% of
the available cpu.  So if you have a series of serial processes being
distributed across all of your nodes, what happens when you try and run a
MM5 job across 3 nodes (i.e. 6 processors)?  Will all the other processes
get shifted to 1 node, and the MM5 job will jug along normally?  Will some
of the processes get left behind, meaning that the whole program will wait
for the slowest node?

3.) Some combination of #1 and #2.  For example, if we expanded to 8
nodes, we could have mosix running on the first 2 nodes and a scheduler
running on the back 6 nodes.  

Basically, I'm trying to balance ease of use for my users who want to
primarily use serial applications, with a few users who need more
intense parallel applications.  For many of the users, I want to keep it
as easy as possible, so the inexperienced user can use the system w/o much
assistance.  But, I want to have the control to efficiently run big mpich
jobs. Any suggestions would be much appreciated.

Thanks,
Alexis







More information about the Beowulf mailing list