Abstract
Scientific computations are the 'raison d'etre' of supercomputers. These computations have very
large resource requirements, not only in terms of the number of calculations they must make
but also in terms of the amount of storage they must use. The primary goal of a supercomputer
system is to provide these resources as efficiently as possible.
Besides servicing production large scale computations, a supercompter must also accommodate the
steady stream of interacive jobs that arises because of program development and initial
debugging of new applications on test data sets. While the resource requirements of these
jobs are relatively small, by their nature they demand reasonable response times.
Because the batch and interactive workloads compete for resources, one of the objectives
of a supercomputer scheduler is to balance the high resource consumption rate of the batch
workload against the response time requirements of the interactive workload. In doing this
the scheduler must consider both processor and memory demands of the jobs.
In this paper we analyze three approaches to scheduling mixed batch and interactive work loads
on a supercomputer: (i) Fixed Partition, in which memory resources are statically allocated
between the workloads, (ii) No Partition, in which the interactive preempts
resources as needed from the batch workload, and (iii) No Partition With
Grouped Admission, in which the interactive workload preempts resources only when the number
of waiting interactive jobs reaches a threshold value. We also investigate the potential
benefits of using virtual memory instantaneously available to them. Using analytic tools,
we compare the different policies according to the average speedup achieved by the batch
workload given that a mean interactive job response time objective must be met by each. We
show that, under a wide variety of conditions, Fixed Partition performs better then the
other polices.
