[mvapich-discuss] run executable with no IB interface

[Steve White]

Hi Sayantan,

It seems we agree that it would be a good feature of MVAPICH for it to be 
ble to launch a single-process run on a node with no interface.

I don't know what I meant about the error message.  As you demonstrated,
it seems fairly clear. (I don't know why "infiniband class devices" and not
"InfiniBand interfaces".)  I just got around to taking care of a note I
took some time ago.  Maybe it had to do with the way in which we were
starting MVAPICH.

Thanks very much for your attention!


On 24.12.06, Sayantan Sur wrote:
> Hello Steve,
> 
> >We have often run into the problem that a user tries to run a test with a
> >single process on a machine that doesn't have an InfiniBand interface, but
> >it fails, and the error message is not helpful to explain why it failed.
> >
> >The situation occurs, e.g. on a cluster with an IB interconnect, whose
> >head node has no IB interface.  The user compiles their executable to use
> >the IB interface, then wants to run a single-process test on the head node.
> >
> >It could be we're doing something wrong.
> >
> >Otherwise, I would recommend:
> >
> >1) Improve the error messages to say explicitly that IB-configured
> >   MVAPICH can't run on a machine with no IB interface.
> >
> >2) If only a single process is started, that it be allowed to run.
> >  
> 
> Thanks for your comments. They certainly make sense. Could you let us 
> know what is the error message you get? I tried to run an executable 
> after unloading IB support on a cluster node, and got this error message:
> 
> [surs at k21-oib:basic] 
> /home/7/surs/projects/mvapich/release/trunk/bin/mpirun_rsh -np 1 k21 ./cpi
> libibverbs: Fatal: no infiniband class devices found.
> [0] No IB devices found
> 
> This is using MVAPICH-0.9.8 using the Gen2 (OpenFabrics) interface. We 
> will aim to enable single process runs on nodes without IB interface in 
> our upcoming release.
> 
> Thanks,
> Sayantan.
> 
> -- 
> http://www.cse.ohio-state.edu/~surs
> 

-- 
Steve White : Programmer
Max-Planck-Institut für Gravitationsphysik      Albert-Einstein-Institut
Am Mühlenberg 1, D-14476 Golm, Germany                  +49-331-567-7195