Mis-using network hardware to build efficient parallel runtime systems
Time and Date: 3:00 PM, Wednesday, September 26, 2018
Place: 235 Weir Hall
Speaker: Dr. Brian Larkins
Abstract: Parallel computing systems continue to expand in scale and increase in complexity, making the development of high performance programs for these systems difficult. In this talk, we will discuss the challenges of modeling scientific problems and present approaches for developing solutions that executeon modern parallel systems. We look at the numerous challenges in maintaining high performance at scale including data distribution, remote data access, and load balance. This talk will focus on new parallel programming models and high-level programming abstractions that rely on the novel use of network interface hardware. We will look at two runtime systems, a global-view distributed key/value store and a distributed task collection that performs automatic dynamic load-balancing. Both systems rely on lower-level offload hardware that is available on next-generation network interfaces.
Bio: Dr. Brian Larkins is currently an Assistant Professor of Computer Science atRhodes College in Memphis, TN. Prior to working at Rhodes, he was an Associate Professor at Coastal Carolina University. He received his Ph.D. from The Ohio State University in 2010. Prior to pursuing his doctorate degree, Dr. Larkins worked in industry for ten years, including time at AT&T Bell Laboratories and three technology startups. His current research focus is on developinghigh-performance programming models and runtime systems on large-scale parallel distributed-memory clusters.