[Beowulf] Kill zombies after a parallel run

David Mathog mathog at caltech.edu
Tue May 2 14:23:31 PDT 2006

mg <mg.mailing-list at laposte.net> wrote:
> I use MPICH- to generate and run an FEM parallel application.
> During a parallel run, one process can crash, leaving the other 
> processes run and OS commands have to be used for kill these zombies. 
> So, does someone have a solution to avoid zombies after a failed 
> parallel run: can the crashed process kill the other processes?

I think what you're saying is one compute node dies and this causes the
master and processes on the other nodes to run forever, or at least
not exit even if they have stopped using CPU. Or are you really
asking about processes that show up in the unix "Zombie" state?

Assuming the former, I've only ever dealt with this on PVM but
I'm assuming MPICH has similar functions.  In general you want
the slaves to send an "I'm DONE" message before they exit, at
which point the parallel system generally delivers a
"Slave Exited" message.  If the master sees the EXITED message 
without having first seen the DONE message the master node can then
send all the other nodes a "Time to Die" message so that they
can clean up as best they can and then exit.  For this to work, of
course, the slaves have to check for this message once in a while.
The master could also send a direct "kill" message to those
processes, causing them to go away, but probably leaving scratch files
and other detritus around.


David Mathog
mathog at caltech.edu
Manager, Sequence Analysis Facility, Biology Division, Caltech

More information about the Beowulf mailing list