Co-Designing Scientific Applications and Performance Tools for High-Performance Computing

Scientific applications must adapt to today’s highly heterogeneous high-performance computing systems. Understanding performance measurements is a crucial first step toward getting a holistic view of a code’s performance, identifying code sections for optimization, and mapping the code sections to target advanced architectures such as Summit. The speaker will discuss three topics:

(1) a co-design implementation effort of a core computational kernel, the Kronecker multiplication, in the DMRG++ application with parallel programming languages/APIs and the performance findings parallel-programming approaches used to address the challenge of load imbalance;

(2) recent work in assisting with the development of accelerated algorithms and efficient implementation of DCA++ on Summit (e.g, the nvprof, scoreP, and HPCToolkit profiling tools) and how they render useful, complementary performance information; and

(3) ongoing research collaborations with the Barcelona Supercomputing Group in using advanced task-based parallel models to drive future parallel programming standards and in driving the development of new performance-measurement features with the TAU/HPCToolkit profiling teams.

About the Speaker:

Arghya Chatterjee a Post-Master’s Research Associate at ORNL and a part-time graduate student at the Georgia institute of Technology. His research focuses on using asynchronous programming models to optimize performance of legacy condensed-matter physics applications at ORNL.