Size-independent vs. size-dependent policies in scheduling heavy-tailed distributions

Abstract

We study the problem of scheduling jobs on a two-machine distributed server, where the job size distribution is heavy-tailed. We focus on two distributions, for which we prove that the performance of the optimal size-independent policy is asymptotically worse than that of a simple size-dependent policy. First, we consider a simple distribution where incoming jobs can only be of two possible sizes. The motivation is that with two largely different sizes, the simple distribution captures the important aspects of a heavy tail. Second, we extend to a bounded Pareto distribution, which has an actual heavy tail. For both cases, we analyze the performance with regards to slowdown (waiting time divided by job size) for several size-independent and size-dependent policies. We see that the size-dependent policies perform better, and then go on to prove that even the best size-independent policy cannot achieve the same performance. We conclude that as we increase the variance of our job size distribution, the gap between size-independent and size-dependent policies grows.