Session: Linear Algebra and Its Application

Distributed-Memory Parallel Symmetric Nonnegative Matrix Factorization

Wednesday, November 18, 2020

3:30 PM – 4:00 PM ET

Channel: Track 3

Author(s)

Srinivas Eswar

Georgia Institute of Technology, United States of America
Koby Hayashi

Georgia Institute of Technology, United States of America
Grey Ballard

Wake Forest University, United States of America
Ramakrishnan Kannan

Oak Ridge National Laboratory (ORNL), United States of America
Richard Vuduc

Georgia Institute of Technology, United States of America
Haesun Park

Georgia Institute of Technology, United States of America

We develop the first distributed-memory parallel implementation of Symmetric Nonnegative Matrix Factorization (SymNMF), a key data analytics kernel for clustering and dimensionality reduction. Our implementation includes two different algorithms for SymNMF, which give comparable results in terms of time and accuracy. The first algorithm is a parallelization of an existing sequential approach that uses solvers for nonsymmetric NMF. The second algorithm is a novel approach based on the Gauss-Newton method. It exploits second-order information without incurring large computational and memory costs. We evaluate the scalability of our algorithms on the Summit system at Oak Ridge National Laboratory, scaling up to 128 nodes (4096 cores) with 70% efficiency. Additionally, we demonstrate our software on an image segmentation task.