The Power of Less: Harnessing Sparsity for Performance Optimization

The seminar can be joined online using this link.

Speaker: Dr. Kazem Cheshmi, McMaster University, Canada

Abstract: Sparse computations are a crucial category of algorithms characterized by minimal interaction between system components. These algorithms are essential for various scientific simulations, including computer graphics, weather forecasting, data analytics, and machine learning. The effectiveness of these simulations heavily relies on the efficient execution of sparse computations. However, irregular memory access patterns and the inherent nature of sparse computations limit their ability to fully utilize high-performance computing systems. In this presentation, I will address the fundamental question of how to optimize complex sparse codes and algorithms for practical applications. I will introduce innovative strategies for automating and redesigning sparse linear algebra computations by decoupling the symbolic information from the numerical manipulations within sparse calculations. I will demonstrate how these novel approaches in symbolic decoupling can automatically generate high-performance code that significantly outperforms even highly-tuned code from state-of-the-art linear solvers and numerical optimization libraries.

Biography: Dr. Kazem Cheshmi, an Assistant Professor in Electrical and Computer Engineering at McMaster University, is a leading researcher in high-performance computing. His academic journey began with a B.Eng. in computer engineering from Ferdowsi University of Mashhad, followed by an M.A.Sc. in the same field from the University of Tehran. He then earned a Ph.D. in computer science from the University of Toronto. Dr. Cheshmi's research passion is evident in his experience at prestigious institutions like Microsoft Research, Adobe Research, and Rutgers University. His dedication has been recognized through numerous awards, including the ACM-IEEE CS George Michael Memorial HPC Fellowship (2020), Adobe Fellowship, and the 2017 ACM Student Research Competitions grand prize. Dr. Cheshmi spearheads the SwiftWare Lab, where he collaborates with students to develop high-speed software systems for machine learning and scientific simulations. These systems leverage custom compiler techniques and libraries to handle complex data and computation patterns beyond the capabilities of traditional compilers.

Host: Prof. Olaf Schenk