Preemptive scheduling for approximate computing on heterogeneous machines: Tradeoff between weighted accuracy and makespan

Abstract

Approximate computing is a research area which investigates a wide spectrum of techniques to trade off computation time for better accuracy. This paper models the accuracy of the computation as a concave, possibly non-differentiable function of job size to decide the accuracy of each job as well as the scheduling decision of the jobs on heterogeneous servers. The two metrics of interest are the makespan of the completion time of the jobs and the weighted accuracy of the different jobs. For a given makespan, an algorithm is provided that achieves the optimal weighted accuracy of computations, thus providing an optimal tradeoff between the two objectives for preemptive scheduling of approximate computing jobs on multiple servers. The proposed algorithm decides the task sizes of each of the jobs, which gives the accuracy of the jobs in addition to the schedule of these partial jobs on the servers.