Approaches to Quality of Service in High-Performance Networks

Abstract

In recent years, the topic of quality of service (QoS) has gar- nered much discussion and research amongst network designers. QoS has many definitions, but here we focus on the notion of predictable com- munication performance such as guaranteed bandwidth and maximum latency. Such predictability not only supports continuous media (e.g. au- dio, video, virtual reality) and interactive response (e.g. client response time, real-time control), but also good parallel application performance. We discuss three major approaches for quality of service and evaluate their match for high performance system-area networks.

We classify the wealth of approaches to network quality of service into three classes: virtual circuits, physical circuits, and global scheduling. These approaches differ significantly in their assumptions about the net- work environment and in how they bind communication tasks to physi- cal resources, how physical resources are multiplexed and scheduled, and what information is available to do so. These choices fundamentally affect the cost, complexity, and performance of each approach. Virtual circuit approaches attain the greatest flexibility, but require complex hardware for implementation. Physical circuit approaches reduce hardware com- plexity, but need research to understand to best exploit them to deliver quality of service. Finally, we discuss two promising avenues for combin- ing external control with simple network switches to deliver both low cost and high performance (route based and time-based global schedul- ing). These approaches are in their infancy and many research questions remain about their capabilities and generality.