Optimizing network throughput: optimal versus robust design

Abstract

Interconnection network performance is usually measured in terms of its latency (time required to deliver a message) and throughput (maximum traffic accepted by the network). At first glance, minimizing average message latency is the main designer goal, because average network traffic is usually far from saturation. However, applications can also generate very high peak traffic. In order to deal with such situations, it is important that network throughput is also high. On the other hand, interconnection network performance depends on several parameters. Some of them can be chosen by the designer: routing algorithm, switching technique, topology and node design parameters. However, there are other parameters that cannot be selected by the designer. Among these, there are parameters that depend on the application, such as message size, message destination distribution and message traffic, as well as parameters defined by the customer, such as network size. Network designer can select the design parameters that maximize average (optimal design) or the design parameters that achieve a good performance under all the feasible combinations of the parameters that cannot be selected by him (robust design). Notice that both alternatives do not always lead to the same parameter configuration. Previously we chose the design parameters of a k-ary n-cube network considering optimize latency. In this case, optimal and robust design lead to the same choice. In this paper, we obtain these design parameters considering optimized network throughput. Unfortunately, there is a discrepancy between optimal and robust design criteria, being the former the best choice.