We have proposed a workload model for batch jobs on
space-sharing parallel computers, based on a uniform-log
distribution of lifetimes. We show that the proposed model captures
the behavior of real workloads in two different environments. This
model can be used to generate synthetic workloads for simulating and
evaluating scheduling policies for similar environments.
We have used this workload model to develop statistical
techniques for predicting queue times for waiting jobs. The proposed
techniques worked well in simulations of both of the environments we
tested. These techniques should be useful for several applications,
especially selection of cluster sizes on space-sharing parallel
computers.
Many scheduling policies for supercomputing environments depend
on information provided by users about the resource requirements of their
jobs. We observe that this information is often unreliable, but show that
the proposed techniques are able to distill this information in a useful
way.
