The Gittins Policy in the M/G/1 Queue

TL;DR

This paper provides the first fully general proof that the Gittins policy optimally minimizes various mean holding cost metrics in the M/G/1 queue, including mean response time and slowdown, even with batch arrivals or partial service time knowledge.

Contribution

It offers a novel, direct proof of Gittins's optimality in the M/G/1 queue, extending previous results to more general settings and metrics.

Findings

Gittins minimizes mean response time in the M/G/1.

Gittins minimizes mean slowdown with unknown or partial service times.

Gittins's optimality holds under batch arrivals.

Abstract

The Gittins policy is a highly general scheduling policy that minimizes a wide variety of mean holding cost metrics in the M/G/1 queue. Perhaps most famously, Gittins minimizes mean response time in the M/G/1 when jobs' service times are unknown to the scheduler. Gittins also minimizes weighted versions of mean response time. For example, the well-known "$c\mu$ rule", which minimizes class-weighted mean response time in the multiclass M/M/1, is a special case of Gittins. However, despite the extensive literature on Gittins in the M/G/1, it contains no fully general proof of Gittins's optimality. This is because Gittins was originally developed for the multi-armed bandit problem. Translating arguments from the multi-armed bandit to the M/G/1 is technically demanding, so it has only been done rigorously in some special cases. The extent of Gittins's optimality in the M/G/1 is thus not entirely clear. In this work we provide the first fully general proof of Gittins's optimality in the M/G/1. The optimality result we obtain is even more general than was previously known. For example, we show that Gittins minimizes mean slowdown in the M/G/1 with unknown or partially known service times, and we show that Gittins's optimality holds under batch arrivals. Our proof uses a novel approach that works directly with the M/G/1, avoiding the difficulties of translating from the multi-armed bandit problem.