Using runtime measured workload characteristics in parallel processor scheduling

Abstract

We consider the use of runtime measured workload characteristics in parallel processor scheduling. Although many researchers have considered the use of application characteristics in this domain, most of this work has assumed that such information is available a priori. In contrast, we propose and evaluate experimentally dynamic processor allocation policies that rely on determining job characteristics at runtime; in particular, we focus on measuring and using job efficiency and speedup.Our work is intended to be a first step towards the eventual development of production schedulers that use runtime measured workload characteristics in making their decisions. We consider two distinct scheduling scenarios: interactive systems, where minimizing response time is the goal, and batch systems, where maximizing useful instruction throughput is the goal. In both these environments, our experimental results validate the following observations:• Despite the inherent inaccuracies of runtime measurements and the added overhead of more frequent reallocations, schedulers that use runtime measurements of workload characteristics can significantly outperform schedulers that are oblivious to these characteristics.• Runtime measurements are sufficient for schedulers to achieve performance surprisingly close to that possible when efficiency and speedup information is available a priori.• The primary performance loss, relative to the use of a priori information, is due to the transient decisions of the schedulers as they acquire information on the running applications, rather than to measurement and reallocation overheads.Our experiments are performed using prototype implementations running on a 50-node KSR-2 shared memory multiprocessor.

Theory and practice in parallel job scheduling

“…Second, even with very many queues, the resolution in which requirements are expressed is necessarily very coarse, and user estimates are notoriously inaccurate anyway. Recent more theoretical work shows how data can be acquired automatically by the system, rather than relying on the users [59,31,12].…”

“…These characteristics include the user id, the file to be executed, the memory size specified, and possibly others. An estimate of the efficiency [59] or the execution time [31] of a job being scheduled can be obtained from retained statistics on the actual resource usage of jobs from the same (or a similar) class that have been previously submitted and executed. A small database can be kept to record resource consumption of jobs on a class by class basis.…”

Parallel Application Scheduling on Networks of Workstations

“…But both SDF and DIF need some form of execution time information in order to be realized. The approximation method proposed by Nguyen et al [17] for shared-memory machines, where the best partition size of an application is determined by sample executions on different numbers of processors, induces overhead which can become nonnegligible in distributedmemory systems.…”

An empirical evaluation of work stealing with parallelism feedback

“…Adaptive job schedulers have been studied extensively, both empirically [25,34,46,50,51,54,60] and theoretically [22,26,27,36,48]. McCann, Vaswani, and Zahorjan [46] studied many different job schedulers and evaluated them on a set of benchmarks.…”