Realistic Models for Predicted Job Times in Queueing

Identify realistic probabilistic models of predicted job service times and associated prediction-error distributions for single-server M/G/1 queues that enable rigorous analysis of prediction-based scheduling policies such as shortest predicted remaining processing time (SPRPT), shortest predicted job first (SPJF), and preemptive shortest predicted job first (PSPJF).

Background

The paper surveys scheduling policies that use predictions of job sizes rather than exact sizes, extending classical M/G/1 analyses to prediction-based variants (SPJF, PSPJF, SPRPT). It models predictions via a joint distribution g(x, y) over true size x and predicted size y and studies some mathematically convenient but artificial prediction distributions (e.g., exponential with mean x, uniform in [(1−α)x,(1+α)x]).

The authors highlight that while these simplified distributions allow tractable analyses, the lack of realistic, broadly applicable prediction-error models limits theoretical guidance and practical relevance, motivating the need to identify natural prediction models suitable across queueing problems.