Using Runtime Measured Workload Characteristics in Parallel Processor Scheduling

We address the design of practical scheduling policies for scalable, shared memory multiprocessors. In particular, we propose and evaluate experimentally processor allocation policies that use information about parallel job execution characteristics in making their decisions. In contrast to existing processor allocation work, which for the most part has relied on perfect information supplied before job execution, we require no a priori information, instead relying on runtime measurement of executing jobs. The experimental results we present validate the following observations: The use of runtime measurements of workload characteristics can signi cantly improve performance relative to disciplines that are oblivious to these characteristics, even given the inherent inaccuracies in the measurements and the overhead of the dynamic reallocations that approaches based on them require. Runtime measurements are su cient for the scheduler to achieve performance surprisingly close to that possible when perfect, a priori information is available. The primary performance loss, relative to the use of a priori information, is due to the transient poor decisions of the scheduler as it acquires information on the running applications, rather than to the overhead of measurement itself. Additionally, we consider both interactive environments, in which a response time directed scheduler is appropriate, and batch environments, in which maximizing useful instruction throughput is the primary goal. Our experiments are performed on a prototype implementation running on a 50-node KSR-2 shared memory multiprocessor. A nal novel aspect of our work is the use of both hand and compiler parallelized programs in our test workloads.