On Thu, 2006-02-02 at 10:19 +0000, Stephen Childs wrote:
> However, the default walltime for the cluster is 1 minute, and jobs that run for
> longer than this get killed. With globus, you can set the walltime
> required in an RSL so it gets fed through to Torque: how do you do this in
> a JDL?
You can't. JDL does not provide this facility.
I have a similar local problem; our local GridEngine cluster is
effectively queueless. We use the attributes assigned to each
individual job (ie memory requirements, upper limit on wallclock time,
etc) to guide scheduling.
This is, as you have discovered, incompatible with the current JDL model
which depends on end-users selecting the queue(s) which best fit their
My current solution is to cheat; I advertise virtual queues that *do not
exist* on the batch system. When a job arrives for a particular virtual
queue (eg my jobmanager-sge-10min target) then my custom JobManager
remaps the job's properties based on the queue that was requested.
For example, jobs that arrive at my local jobmanager-sge-10min target
gets mangled in the following ways:
- Change queue specification to '*' (the SGE queue wildcard)
- Change wallclock timelimit to 10 mins.
Modifying our SGE JobManager to perform this mapping was fairly trivial;
writing a new information reporter to calculate the appropriate values
to report for each virtual queue was a little harder..
David McBride <[log in to unmask]>
Department of Computing, Imperial College, London