Live Webcast 15th Annual Charm++ Workshop

Communication is a necessary but overhead inducing component of parallel programming. Its impact on application design and performance is due to several related aspects of a parallel job execution: network topology, routing protocol, suitability of algorithm being used to the network, job placement, etc. This thesis is aimed at developing an understanding of how communication plays out on networks of high performance computing systems and exploring methods that can be used to improve communication performance of large scale applications. Broadly speaking, three topics have been studied in detail in this thesis. The first of these topics is task mapping and job placement on practical installations of torus and dragonfly networks. Next, use of supervised learning algorithms for conducting diagnostic studies of how communication evolves on networks is explored. Finally, efficacy of packet-level simulations for prediction-based studies of communication performance on different networks using different network parameters is analyzed. The primary contribution of this thesis is development of scalable diagnostic and prediction methods that can assist in the process of network designing, adapting applications to future systems, and optimizing execution of applications on existing systems. These methods include a supervised learning approach, a functional modeling tool (called Damselfly), and a PDES-based packet level simulator (called TraceR), all of which are described in this thesis.