Cluster architecture for sequential applications

Jarrod Smith jsmith at structbio.vanderbilt.edu
Sat Oct 13 09:25:44 PDT 2001


Check out www.mosix.org.  For course-grained process distribution across a
cluster (i.e. each process runs for on the order of minutes to hours), it
will behave exactly like you describe.  You just execute normal jobs on a
single "master" node and it will automatically migrate them to the
least-loaded slave.

We just experimented with it here on a serial code that we use.  It works
very well and seems fairly painless to implement.  In fact, our sysadmin
is working on a "mosix on a floppy" project to make this even more
painless.  The way he implements it, you first configure a single machine
as a mosix "master" and then using his floppy image, simply boot several
diskless machines from floppy as mosix slaves.  As the slaves come up,
they configure themselves from info stored on the master and jobs start
migrating to them.  When they go down, mosix is smart enough to know about
it and stop sending jobs to the defunct slave(s).

-- 
Jarrod A. Smith, Ph.D.
Asst. Director, Center for Structural Biology
Research Asst. Professor, Biochemistry
Vanderbilt University

On Sat, 13 Oct 2001, Jacques B. Siboni wrote:

> Hi all,
> 
> I'd like to put a general question here.
> 
> The use of the beowulf cluster i'm presently building is more to be used for
> sequential applications than parallel ones.
> 
> Let me explain. I want to use a cluster to increase the power of the server.
> That is to automatically distribute processes among the cluster each time the
> server receives a client query.
> This should work for all of the services the server provides. The services
> need not to be specifically designed for parallel processing.
> 
> Therefore I need a cluster architecture, a message passing protocol and a
> supervisor which decides what node is going to process the request. I need
> some kind of distributed process id.
> 
> What I have to find is the transparent manager who can take care of grabbing
> the client requests for services and direct them to a node idle enough to
> handle it.
> 
> In other words I look for the architecture which emulate a multi cpu
> motherboard
> 
> I hope this is clear enough to be understood.
> 
> Thanks for your advises
> 
> Cheers
> 
> Jacques
> 
> 
> 





More information about the Beowulf mailing list