School of CSE Seminar Series: Aydin Buluc

Speaker: Aydin Buluc, senior scientist at Lawrence Berkeley National Laboratory and adjunct assistant professor at the University of California, Berkeley
Date and Time: October 18, 2:00-3:00 p.m.
Location: Coda 114
Host: Helen Xu

Title: Supercomputing Scale Graph Neural Network Training

Abstract: Graph neural networks (GNNs) have demonstrated unprecedented success in numerous challenging scientific problems such as weather prediction, material design, and protein structure prediction. However, training a large-scale GNN is both memory intensive and computationally expensive. Consequently, solving grand challenge scientific problems require supercomputing scale GNN training capabilities. My talk will focus on distributed-memory parallel algorithms for GNN training. We will start by describing how to map full-batch GNN training to communication-avoiding sparse matrix operations. We will then focus on utilizing sparse matrix primitives to parallelize mini-batch GNN training based on node-wise and layer-wise sampling. Finally, we will illustrate techniques that are based on sparsity-aware sparse matrix multiplication algorithms to accelerate both full-graph and mini-batch sampling based GNN training.

Bio: Aydın Buluç is a Senior Scientist at the Applied Math and Computational Research Division of the Lawrence Berkeley National Laboratory (LBNL) and an Adjunct Faculty at EECS department of UC Berkeley. His research interests include parallel computing, combinatorial scientific computing, high performance graph analysis and machine learning, sparse linear algebra, and computational genomics. He received his Ph.D. in Computer Science from the University of California, Santa Barbara in 2010. After that, he was a Luis W. Alvarez postdoctoral fellow at LBNL. Dr. Buluç is a recipient of the DOE Early Career Award in 2013 and the IEEE TCSC Award for Excellence for Early Career Researchers in 2015. He recently led a team that was chosen as a finalist for the 2022 ACM Gordon Bell Prize. He was a founding associate editor of the ACM Transactions on Parallel Computing. He is currently directing a DOE Mathematical Multifaceted Integrated Capabilities Center named Sparsitute.