mpiexec tm_poll error

[Pete Wyckoff]

parnell at msi.umn.edu said:
> Sorry, let me clarify, there was no pbs_demux in the "pbs-mom" rpm
> package of PBSPro, in theory a seperate rpm you can install on
> execution nodes where you don't need the entire distribution.  Our
> University has had an issue with the contract for the source dist
> of PBSPro so we've relied upon the binary dists.  I found pbs_demux
> on the server (it had the full dist of PBSPro) and copied this over 
> to all the execution nodes.  

Aha.  That's definitely a bug in their packaging I suspect.  Noone
likely notices since the only things that try to talk to pbs_demux are
jobs started via tm_spawn.  That's not too popular, but mpiexec does
rely on it to work.

All this is speculation, as I don't have the non-free code.  If anybody
feels like complaining to their PBSPro provider, maybe it would get
fixed.  Until then, ensuring that pbs_demux exists on all compute nodes
sounds like a good fix.

		-- Pete