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. As a result, a vast search space may need to be explored to find configurations that are optimal for different applications. This talk presents our recent work on two prediction methods that can help in such an exploration at lower cost: functional modeling and trace-driven simulation. For each of the methods, I will describe the underlying principles and discuss a few case studies to demonstrate their efficacy.

Bio:
Nikhil Jain is a doctoral candidate with Prof. Laxmikant Kale in the Department of Computer Science at the University of Illinois at Urbana-Champaign. His current research interest spans high performance computing networks, performance prediction methods, and interoperation of parallel languages. As part of the Charm++ group, he is among the lead developers of the Charm++ software and OpenAtom application. Nikhil was awarded the IBM PhD fellowship in 2014 and the Andrew and Shana Laursen fellowship in 2011. More information about his publications and projects can be found at his home page.