Heavy-Tailed Limits for Medium Size Jobs and Comparison Scheduling

TL;DR

This paper analyzes the asymptotic behavior of sojourn times under different scheduling disciplines for jobs conditioned on their size relative to previous arrivals, revealing SRPT's superior performance for smaller jobs and proposing an adaptive grouping approximation.

Contribution

It introduces a novel conditional limit framework for analyzing scheduling performance and proposes an adaptive job grouping mechanism based on relative size comparisons.

Findings

SRPT outperforms PS/FBPS for smaller jobs in asymptotic limits

The adaptive grouping mechanism effectively approximates SRPT performance

Heavy-tailed job requirements are well-modeled by the proposed classification approach

Abstract

We study the conditional sojourn time distributions of processor sharing (PS), foreground background processor sharing (FBPS) and shortest remaining processing time first (SRPT) scheduling disciplines on an event where the job size of a customer arriving in stationarity is smaller than exactly k>=0 out of the preceding m>=k arrivals. Then, conditioning on the preceding event, the sojourn time distribution of this newly arriving customer behaves asymptotically the same as if the customer were served in isolation with a server of rate (1-\rho)/(k+1) for PS/FBPS, and (1-\rho) for SRPT, respectively, where \rho is the traffic intensity. Hence, the introduced notion of conditional limits allows us to distinguish the asymptotic performance of the studied schedulers by showing that SRPT exhibits considerably better asymptotic behavior for relatively smaller jobs than PS/FBPS. Inspired by the preceding results, we propose an approximation to the SRPT discipline based on a novel adaptive job grouping mechanism that uses relative size comparison of a newly arriving job to the preceding m arrivals. Specifically, if the newly arriving job is smaller than k and larger than m-k of the previous m jobs, it is routed into class k. Then, the classes of smaller jobs are served with higher priorities using the static priority scheduling. The good performance of this mechanism, even for a small number of classes m+1, is demonstrated using the asymptotic queueing analysis under the heavy-tailed job requirements. We also discuss refinements of the comparison grouping mechanism that improve the accuracy of job classification at the expense of a small additional complexity.