ICC Distinguished Lecturer Series: Kalyan Perumalla
Kalyan Perumalla, PhD
Group Leader, DCS Group, CSE Division, ORNL
Distinguished R&D Staff Member and Manager, CSE Division, ORNL
Adjunct Professor, School of CSE, Georgia Tech
Seminar Talk: Effective Exascale Computing using Computational Cloning
Abstract: "Exascale computing," being pursued by the US and others across the world, is the next level of supercomputing many times larger than the largest petascale parallel systems available today. Among the challenges in achieving exascale computing are the unprecedented levels of aggregate concurrency and the relatively low memory sizes per computational unit offered. Towards meeting these challenges, we present the design, development, and implementation of a novel technique called Cloning. Cloning efficiently simulates a tree of multiple what-if scenarios dynamically unraveled during the course of a base simulation. The lecture will describe the new conceptual cloning framework, and a prototype software system named CloneX that provides a new programming interface and implementation of cloning runtime scaled to supercomputing systems. CloneX efficiently and dynamically creates whole logical copies of what-if simulation trees across a large parallel system without the inordinate cost of full physical duplication of computation and memory. The approach is illustrated with example applications, including epidemilogical outbreaks and forest fire outbreaks, whose aggregate runtime and memory consumption using cloning are decreased by two to three orders of magnitude relative to replicated runs. Performance data on large simulation trees using 1024 GPUs reinforce the promise of cloning as an effective approach that can be adopted to meet the concurrency and memory challenges at exascale.
Refreshments will be served.
Friday, April 21, 2017 at 3:00 pm to 4:00 pm
Rekhi Hall, 214
1400 Townsend Drive, Houghton, MI 49931