Asymptotic optimality of statistical multiplexing in pipelined processing

Abstract

The throughput of pipelined processing ofheterogeneous multitasked jobs is computed and optimized in this study. There areK job classes. Each job hasM tasks which have to be processed in a given order (same for all tasks) on a pipeline ofM processors. Tasks have random processing times. The jobs of each class form a stationary and ergodic sequence (with respect to their task processing times). Classes are differentiated by distinct statistics and may not be jointly stationary or ergodic. Thus, the jobs are overall statistically heterogeneous. We are interested in the average execution time per job$\bar \tau$, when the job populations of the various classes become very large (asymptotically). This is shown to depend on the order in which jobs enter the pipeline. Under the natural class-based ordering, where all jobs of the first class enter first, followed by those of the second, third, and so on, the quantity$\bar \tau$ is computed, but is shownnot to attain its minimal value in general. On the contrary, appropriate statistical multiplexing of jobs of different classes on the pipeline is shown to minimize the average execution time per job on every sample path (with probability one). The procedure, calledbalanced statistical multiplexing, is constructed and the minimal$\bar \tau$ is computed in terms of the average execution times of the job tasks.