CAREER:The Impact of Resource Scheduling on Improving Server Performance

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

This proposal tackles fundamental problems in system design using a two-pronged approach including both queueing theory and kernel-level implementation. Particular emphasis is placed on understanding the impact of heavy-tailed properties in compute workloads. The proposal focuses on two specific research questions: (1) Is it possible to reduce the expected response time of every request at a server, simply by changing the order in which requests are scheduled? Todays web servers employ FAIR scheduling, wherein requests are time-shared fairly. This proposal considers instead scheduling which biases towards SHORT requests, and argues analytically and via implementation that biasing towards SHORT requests improves response times for all requests. (2) How should servers be designed to cope with transient overload conditions? Most of client dissatisfaction can be attributed to moments of temporary overload at the server. This proposal aims to understand exactly what happens to systems during overload, both via a formal queueing analysis and via a rigorous systematic study of server/OS internals.