HOD refactoring to ease integration with scheduler/resource managers other than
torque
--------------------------------------------------------------------------------------
Key: HADOOP-5441
URL: https://issues.apache.org/jira/browse/HADOOP-5441
Project: Hadoop Core
Issue Type: Improvement
Components: contrib/hod
Environment: All
Reporter: Nate Woody
Situation: HOD currently uses the pbsdsh (a distributed shell that works via
Torque's TM interface to start remote processes) command to start processes on
all nodes in the job. This call is provided as part of a torqueInterface class
that is meant to abstract interactions with the torque resource managers (RMs).
However, this is not functionality typically provided by other RMs, and is
instead typically performed by an distributed command available on the HPC
system, mpiexec, ssh, or site-specific scripts. The specificity of pbsdsh to
Torque makes writing HOD interfaces to other RMs somewhat difficult as it
forces the implementer to choose the remote start method on a somewhat faulty
per-RM basis.
Proposal: Refactor the torqueInterface and nodePool classes so that the choice
of remote start method is available as a configuration option in hodrc. This
involves fairly simple changes to remove the pbsdsh command from the Scheduler
class and addition configuration step of starting the appropriate remote start
wrapper. The selection of the nodePool class will be altered to allow dynamic
loading of classes, so that new interfaces people choose to write will not
require altering HOD code. Provide remote start classes for pbsdsh, mpiexec,
ssh, as well as custom scripts (sites often provide mpiexec wrappers that
ensure proper selection of network interfaces, etc). Provide interface classes
to SGE and Moab, as well as updated Torque class.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
