Provably-Efficient Online Adaptive Scheduling of Parallel Jobs Based on Simple Greedy Rules

Abstract

Scheduling competing jobs on multiprocessors has always been an important issue for parallel and distributed systems. The challenge is to ensure overall system efficiency while offering a level of fairness to user jobs. Although various degrees of successes have been achieved over the past decades, few existing schemes address both efficiency and fairness over a wide range of work loads. Moreover, in order to obtain analytical results, many known results [22, 24, 7, 8, 17, 20, 23, 25, 33] require prior information about jobs such as jobs' release time, amount of work, parallelism profile, etc, which may be difficult to obtain in real applications. This chapter describes a scheduling algorithm GRAD, which offers provable efficiency in terms of makespan and mean response time by allotting each job a fair share of processor resources. Our algorithm is non-clairvoyant [10, 6, 18, 12], i.e. it assumes nothing about the release time, the execution time, and the parallelism profile of jobs. A parallel job can be classified as adaptive or non-adaptive. An adaptively parallel job [34] may change its parallelism, and it allows the number of the allotted processors to vary during its execution. A job is nonadaptive if it runs on a fixed number of processors over its lifetime. With adaptivity, new jobs can enter the system by simply recruiting processors from the already executing jobs. Moreover, in order to improve the system utilization, schedulers can shift processors from jobs that do not require many processors to the jobs in need. However, since the parallelism of adaptively parallel jobs can change during the execution and the future parallelism is usually unknown, how a scheduler decides the processor allotments for jobs is a challenging problem. We describe GRAD that effectively addresses such an adaptive scheduling problem. Scheduling parallel jobs on multiprocessors can be implemented in two levels [14]: a kernellevel job scheduler which allots processors to jobs, and a user-level thread scheduler which maps the threads of a given job to the allotted processors. The processor reallocation occurs periodically between scheduling quanta. The thread scheduler provides parallelism feedback to the job scheduler. The feedback is an estimation of the number of processors that its job can effectively use during the next quantum. The job scheduler follows some processor allocation policy to determine the allotment to the job. It may implement a policy that is either space-sharing, where jobs occupy disjoint processor resources, or time-sharing, where O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg