Browse Theses

When multiple jobs compete for processing resources on a parallel computer, the operating system kernel's processor allocation policy determines how many and which processors to allocate to each. This dissertation investigates the issues involved in constructing a processor allocation policy for large scale, message-passing parallel computers supporting a scientific workload.

First, the issues that affect the performance of scheduling policies for message-passing parallel systems are examined. We argue that reasonable policies must provide nearly equal resource allocation to all runnable jobs and allocate, to a single job, processors that are in close proximity to one another.

Second, the concept of efficiency preservation is defined as a characteristic of processor allocation policies. Efficiency preservation captures the impact of scheduling overheads such as reallocation cost and system-induced load imbalance on processor efficiencies. We show how efficiency preservation can be used in a first-order evaluation to identify promising scheduling policies.

Third, we address short-term scheduling, that is, how to allocate processors among a set of jobs where resource requirements allow them to be executed concurrently. The details of two families of processor allocation policies, Equipartition and Folding, are specified. Both policy classes employ dynamic allocation, but differ in the way they address their costs. Folding achieves good application load balance at the cost of higher reallocation overhead, while Equipartition achieves low reallocation cost at the expense of higher system-induced load imbalance. Through performance evaluation of these policies, this dissertation shows that while maintaining low reallocation overhead is imperative to the good performance of dynamic allocation policies in message-passing systems, load balancing is also a dominant factor in the policy performance.

Fourth, this dissertation investigates medium-term scheduling policies, that is, how to choose subsets of the runnable jobs to execute concurrently when the total resource requirements of all the jobs exceeds the system's capacity. We argue that, in general, an efficient optimal solution to this problem is unlikely. As a result, this dissertation concentrates on restricted problem domains, and shows that, for these cases, optimal solutions exist.