Distributed computing has brought about promising new possibilities, both by increasing the computing power to which people have access and by supporting new technologies such as real-time data analysis and collaborative applications. The power of distributed systems is offset by the tremendous complexity of developing applications for dynamic, heterogeneous environments. An important way to manage distributed applications is designing them to adapt their computing and networking needs to their environment. To support adaptation, a number of systems provide resource information obtained using active benchmarks. Benchmarks provide support for many applications, but their effectiveness is limited by low scalability, invasiveness, and the inability to derive network topology. I have examined the use of low-level network information to support adaptive applications without the shortcomings of active benchmarking.

The low-level details obtained directly from network components provide the information needed by distributed applications to adapt themselves to modern network environments. Low-level access overcomes the limitations inherent in benchmarking by providing a scalable, non-invasive measurement technique that provides network topology information while continuing to support the predictions of end-to-end application performance available through benchmarking. In this dissertation, I address the need for low-level information, the feasibility of providing it through an application-level interface, the accuracy of end-to-end predictions made provided by low-level information, and the topology discovery capabilities using low-level information. The topology discovery algorithms I present are the first to use the incomplete information available through network components and are provably good with minimal knowledge. My research demonstrates that violating the end-to-end networking abstraction by providing applications with access to low-level network knowledge meets the needs of many applications and is feasible on modern networks.

148 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu