This preview shows pages 1–2. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: Fairness and Classifications Adam Wierman Computer Science Department Carnegie Mellon University Pittsburgh, PA 15217 [email protected] ABSTRACT The growing trend in computer systems towards using scheduling policies that prioritize jobs with small service requirements has re- sulted in a new focus on the fairness of such policies. In particular, researchers have been interested in whether prioritizing small job sizes results in large jobs being treated “unfairly.” However, fair- ness is an amorphous concept and thus difficult to define and study. This article provides a short survey of recent work in this area. 1. INTRODUCTION Traditionally, the performance of scheduling policies has been measured using mean response time (a.k.a. sojourn time, flow time), and more recently mean slowdown [3, 16, 24] and the tail of re- sponse time [6, 7, 26]. Under these measures, policies that give priority to small job sizes (a.k.a. service requirements) at the ex- pense of larger job sizes perform quite well. For example, Shortest- Remaining-Processing-Time ( SRPT ) is known to optimize mean response time [36, 37]. As a result, designs based on these policies have been suggested for a variety of computer systems in recent years. However the adoption of these new designs has been slow due to fears about the fairness of these policies. Specifically, there are worries that large job sizes may be “starved” of service under a policy that gives priority to small job sizes, which would result in large job sizes having response times that are unfairly long and variable [5, 38, 39, 40]. These worries have recurred nearly everywhere size based poli- cies have been suggested. A first example is the case of web servers, where recent designs have illustrated that giving priority to requests for small files can significantly reduce response times [17, 29]. However, it is important that this improvement does not come at the expense of providing large job sizes unfairly large response times, which are typically associated with the important requests. It is in this setting that Bansal & Harchol-Balter provided the first study of the fairness of SRPT . The same tradeoff has appeared across di- verse application areas. For example, UNIX processes are assigned decreasing priority based on their current age – CPU usage so far. The worry is that this may create unfairness for old processes . Similar tradeoffs can be found in recent designs for routers [27, 28], wireless networks , transport protocols , and beyond. To address these worries, there has been an explosion of theoret- ical work studying the “fairness” of priority-based policies, and it is the purpose of this paper to present a brief survey of this work....
View Full Document
This document was uploaded on 01/05/2012.
- Fall '09
- Computer Science