Using Accelerator as a Base Scheduler

When interfacing to Accelerator, a minimal (zero) configuration for Accelerator Plus is recommended.

Do not configure each Accelerator Plus instance (resources, limits, etc.), and instead rely on the base scheduler's configuration to implement these policies. This approach allows many Accelerator Plus instances to be employed without a corresponding increase in administration overhead.

A notable exception is with queue-specific job classes: jobclasses need to be present in the Accelerator Plus instance otherwise an error is flagged. To avoid duplication of code, it may be desirable to symlink the base scheduler's jobclass directory to the Accelerator Plus instance. It may be necessary to copy or symlink the Accelerator vnc_policy.tcl file to the Accelerator Plus instance, for example: data specific to the job submission environment is required to run the policies.

Also, the jobclasses must be written so that they can be sourced multiple times, (e.g. to prevent duplicate entries in the resource lists), since it is likely that they will be executed once in the Accelerator Plus context and once in the Accelerator context.

The use of preemption within Accelerator Plus is not recommended; it is best to let Accelerator do all necessary preemptions, which will result in job modulation.

FairShare works the same in Accelerator Plus and in Accelerator. You will notice that, in Accelerator, FairShare convergence will be somewhat slower than you would expect, because Accelerator Plus is effectively sending to Accelerator much longer jobs than the jobs in the Accelerator Plus workload, due to the fact that Accelerator sees the "vovtasker" as a single job.