Job resource manager

The part of OAR that is usually responsible for setting up and cleaning up the operating system of resources given to a job is the job resource manager.

This is a Perl script that is part of the OAR configuration files: located on the OAR server in /etc/oar (see oar.conf). The default one provided with the OAR is named: /etc/oar/job_resource_manager_cgroup.pl.

This Perl script is broadcasted (thanks to TakTuk) to be executed on all the nodes of a job to initialize the resources at the beginning (creating the job cgroups and more), and to clean-up the resources at the end (killing the remaining processes of the job, removing the job cgroups, and more).

This script can be modified or just augmented by admins: it is meant for that, despite exposing a good part of the complexity of OAR.

Notes:

• Contrary to the job prologue/epilogue which executes only on the head node of the job and contrary to the job server prologue/server epilogue which executes on the OAR server, the job resource manager executes on all the nodes of the job hosting resources of type default.
• It is possible to configure OAR do not use the job resource manager, but this is not usual (no cgroup used) and has not been tested recently.
• The job resource manager does not execute on the deploy frontend (for jobs of type deploy) nor on the cosystem frontend (for jobs of type cosystem): indeed OAR does not set up cgroups on frontends unlike on nodes.
• For jobs of type deploy or cosystem, the job prologue/epilogue however executes on the deploy/cosystem frontend, and server prologue/server epilogue on the OAR server.
• For jobs of type noop, the job prologue/epilogue do not execute, but the server prologue/server epilogue do execute on the OAR server, as well as the job resource manager on all nodes hosting resources of type default.