[Fwd: RE: [openib-general] Loading / unloading IB modules]

Matt L. Leininger
Fri Mar 26 17:24:47 PST 2004


Sandia and Los Alamos are beginning to do side-by-side comparisons of
each IB stack.  With the vast amount of IB code now open sourced, we
need to develop a testing procedure to determine what design criteria
will give us an IB stack that meets the requirements of enterprise, high
performance, and grid computing, as well as the guidelines to eventually
getting into the linux kernel (I'm talking kernel.org not distros
kernels).   There are lots of folks at the DOE national labs, NCSA, etc.
that want to set up IB Linux clusters with thousands of nodes.  We also
need to ensure that the IB system software is scalable to that many
nodes so that a single application (using MPI for example) can run
across the entire cluster.  

For those on the list who aren't aware, Sandia is hosting openib.org.
We should be able to finish up the subversion setup on Monday (sysadmins
off today).  The subversion tree will have the initial openib stack as
well as all the other open source stacks (InfiniCon, SourceForge,
Voltaire, Divergenet) and various MPI's (MVAPICH, LA-MPI, MPICH-VMI, and
MPICH2).  Folks will be able to register accounts to get subversion
commit privileges.  

I agree that the number of modules in the current stack is excessive. 
At some point you have to follow the 'keep it simple stupid' rule.  Some
of the other IB stacks have combined module functions to reduce the
number of overall modules to a minimum.  I think any long term solution
for an IB stack should do the same.

As discussion increases on openib we can easily set up more mail lists
to focus on specific topics (access layer, ipoib, subnet management,
adhering to linux coding standards, whatever).  Drop me or
moderator at openib.org a note and we can set things up.

thanks,

- Matt


On Fri, 2004-03-26 at 13:49, Fab Tillier wrote: 
> Is anyone impartial (i.e. the labs) planning on doing a side-by-side
> comparison of the different IB stacks available to determine what features
> to migrate which way?  A lot of the issues being discussed here are not
> present in other stacks, specifically full RMPP support and load/unload.
> 
> It worries me when I hear comments like:
> >    Greg> Yes, but we are developers, and the fact that these modules
> >    Greg> can't do module unload is a symtom of a real problem.
> >
> > Roland> Fair enough.  I guess all I'm saying is that it's not high on my
> > Roland> personal list because there are a lot of other real problems I'm
> > Roland> trying to fix :)
> 
> - Fab


-- 
To unsubscribe send an email with subject unsubscribe to openib-general at openib.org.
Please contact moderator at openib.org for questions.




More information about the openib-general mailing list