Serial and Parallel, with/out Interactive all on One Hostgroup
- What we got:
-
- A few dozen nodes.
- Each has 12 cores.
- What we want:
-
- Users to be able to run serial, SMP and MPI jobs on the FAT nodes.
- The MPI jobs either fitted wherever they can, so as to start as quickly as possible,
- or fitted to use whole nodes only, to run as efficiently as possible.
- To be able to run interactively too.
- To be sure that a user running an interactive job does not get a shell on a node on which a batch job has been running for a long time and screw the node up — so mutual-subordination is definitely necessary!
- Implementation — three queues. . .
-
C6100-STD.q C6100-STD-serial.q C6100-STD-interactive.q
- . . .a complex attribute. . .
-
#name shortcut type relop requestable consumable default urgency #------------------------------------------------------------------------------------------ interactive inter BOOL == FORCED NO 0 0
- . . .plus a resource quota. . .
-
{ name C6100-STD.rqs description NONE enabled TRUE limit hosts {@C6100-STD} to slots=12 }
- . . .with resource quotas and some subordination
-
- The resource quota allows us to run multiple queues at the same time on the same hosts (without overloading these hosts).
- Each of the three queues subordinates the two others.
- Why all those queues?
-
- We want to be able to impose different limits on each queue from the others.
- We may want the interactive queues to be serial only.
- Usage
-
C6100-STD-serial.q qsub [[no flags needed at all]] ... C6100-STD-interactive.q qrsh -l inter ... C6100-STD.q qsub -pe smp.pe ... qsub -pe orte.pe ... qsub -pe orte-12.pe ...