[ale] beowulf clustering

[hbbs at attbi.com]

I think you're more or less correct.  I think the distinction lies in what
constitutes a "thread" in Linux.

Just to rehash/recap, Beowulf-class clusters can be thought of as just a bunch
of boxes that run code that is written specifically for clustering and contain
calls to message-passing libraries that facilitate interprocess communication. 
This benefits code written to solve problems that can be broken down into
parallelizable bits whose computations depend on the results from other bits. 
One must pay special attention to the networking between the nodes with the
dynamics of the message-passing specific to the code being run when building the
cluster lest it bog itself down.

By contrast, Mosix-class clusters run ordinary code and primarily benefit
solving problems that can be broken down into separate processes that do not
need to intercommunicate, although the processes may communicate with a common
file system or database server.  In the classic "node zero" Mosix cluster, one
of the nodes is the one on which processes are started but the kernel patches
and config data cause the processes to migrate to and start running on the least
busy node.  In the baseline or idle case, there is very little network traffic
to a Mosix cluster, however, in practice, this can change once NFS mounts and/or
the Mosix common file system are implemented.  Under Mosix, you get your benefit
not just from the multiplication of CPUs but the width of the aggregate memory
and disk I/O pipes that grow wider and wider with each added node.  

There are other kinds of loosely-coupled clusters you can make, such as a Web
server cluster with a front end that redirects incoming HTTP requests among a
set of nodes that share a common Web document space.  That's what TurboLinux
Cluster Server did and I'm sure there are other mechanisms as well.

- Jeff
> On Fri, 2003-03-21 at 08:06, da Black Baron wrote:
> > You know you gotta have special code to run stuff on Beowulf?  Openmosix
> > lets you run anything, and will migrate any thread generated by a
> > program to a remote node for processing (with some exceptions).
> > 
> 
> Aren't the exceptions related to shared mem and I/O?
> 
> If you'r sharing memory between 2 threads I would think that you could
> not migrate.  If you have an FD to a local file, I also think you would
> not migrate.
> 
> Problem is that most threaded applications shared memory.
> 
> > http://openmosix.sourceforge.net
> > 
> > It's pretty easy to install, too.  Just patch your kernel, compile, and
> > there ya go...
> > 
> > 
> > On Thu, 2003-03-20 at 21:04, Christopher Bergeron wrote:
> > > -----BEGIN PGP SIGNED MESSAGE-----
> > > Hash: SHA1
> > >  
> > > Have any of you guys set up a Beowulf cluster?  Is it difficult?  Can it 
> > > be done with off-the-shelf components?
> > 
> > -- 
> > -------------------------------------------------------------------------
> > -------------------------------------------------------------------------
> > "Here's my family tree.  As far as I can tell, my relatives were 
> > carnival folk who were touring this place called Hiroshima in 
> > the summer of 1945.  Because they lost most of their hair, they 
> > mostly married each other.  And here I am"
> > 
> > -------------------------------------------------------------------------
> > -------------------------------------------------------------------------      
> >                                                         (,,oBo,,)       
> >  Wylde Bill                                               ||||          
> > __________________________________________           http://lyst.org    
> >                 
> > http://mrpooter.sytes.net                              
> > wyldechylde at geocities.com
> > -------------------------------------------------------------------------
> > -------------------------------------------------------------------------  
> > 
> > _______________________________________________
> > Ale mailing list
> > Ale at ale.org
> > http://www.ale.org/mailman/listinfo/ale
> 
> 
> _______________________________________________
> Ale mailing list
> Ale at ale.org
> http://www.ale.org/mailman/listinfo/ale
_______________________________________________
Ale mailing list
Ale at ale.org
http://www.ale.org/mailman/listinfo/ale