Distributed systems have the power and resources needed to solve many interesting problems, in areas ranging from physics to market analysis. However, the complexities of distributed systems prevent many who could benefit from making full use of distributed computing’s power.
My research has focused on providing distributed applications with the information they need to make effective use of the resources available to them. Specifically, I have focused on measuring and predicting network capabilities through the use of network-layer information. Obtaining resource information directly from the network components provides the same accurate performance predictions currently obtained through application-level benchmarking. Furthermore, it offers the scalability and topology discovery needed to support emerging distributed systems. Topology discovery is particularly important for providing performance predictions for parallel applications and building models of distributed systems.
In this talk, I will provide an overview of the network-based performance prediction technique. I have experimentally verified the accuracy of this technique, and it has been implemented in the Remos system at Carnegie Mellon. I will discuss its accuracy and factors affecting its usability, as well as several applications that can benefit from the additional information provided by this technique.