#3 Attach to subset

[Mark O'Connor]

When dealing with very large scale applications, a "attachToSubsetAndSpawnDaemons" function would be valuable. The only way for an API consumer to attach to a subset of the running job at present is to attach to all and kill off any unwanted daemons.

[Dong Ahn]

I think this is a reasonable interface addition which CAN be done most of platforms I've seen. On SLURM-based systems, LaunchMON should be able to launch back-end daemons on an actual target subset. On BlueGene, however, LaunchMON will have to launch daemons to all the IONs first and to force exit for those daemons that do not belong to the subset. BlueGene RM isn't flexible to allow launching of daemons only onto a target subset. Please comment on how one can do this on CRAY if you know ALPS behavior.

In any case, one option to move forward might be to study all the nuance for the platforms we want to support first and to propose the actual interface and semantics for attachToSubsetAndSpawnDaemons?

-Dong

[Nobody/Anonymous]

As far as I know, ALPS requires you to launch a daemon on all nodes and force the ones you don't want to exit - as on BlueGene. I think this is a pretty sane behaviour for LaunchMON as a fallback - any client application would have to implement the same code anyway (DDT currently does this by default on BG/P and with LaunchMON).