Towards Application Performance Fairness on Clouds

Abstract

In cloud computing, resource allocation is the key building block. Existing resource allocation strategies are designed for multi-tenant cases and the majority of them target resource fairness and utilization. We consider the problem of resource allocation among multiple applications that belong to a single user. In this case, instead of resource fairness, the user cares about performance fairness. We focus on memory, one of the most universal resources used by applications. We mainly face two challenges. The first one is how to characterize the performance of various applications. We carefully choose User time, the number of CPU jiffies spending on the User mode, as a representative of the application’s progress to reflect the performance of the application. Through a series of processing, we make performance comparable among different applications. The other challenge is to ensure high system memory utilization. We address this challenge by designing an adaptive memory allocation algorithm that guarantees high memory utilization and performance fairness simultaneously through dynamically reallocating memory among applications. We have implemented our algorithm by developing a scheduler on a physical machine where multiple applications share resources. Experimental evaluation shows that our algorithm can achieve good performance fairness on the premise of ensuring high memory utilization with different applications and importance.

Rengan Dou

5th Year Ph.D. Student

Xin Wang

Senior Research Fellow

My research interests include social networks, network economics and distributed systems.

Richard T. B. Ma

Associate Professor

My research interests include cloud computing, big data systems and network economics.