LORACX: Low Rank Approximations with Constraints at Exascale

Abstract:

This talk introduces the Low Rank Approximations with Constraints at Exascale (LORACX) framework for computing large-scale Nonnegative Matrix Factorization (NMF) on exascale systems.  LORACX implements state-of-the-art distributed graphics processing unit (GPU)-accelerated NMF algorithms integrated into a modern Python-based high-performance computing (HPC) stack.  Key innovations include communication optimizations via overlapped and blocked algorithms to mitigate latency and memory constraints, alongside GPU-tailored non-negative least squares (NNLS) solvers.  Performance evaluations on up to 8192 Frontier nodes demonstrated scalability, processing a 16.3 x 16.3 million dense matrix in 3 seconds at 0.67 exaflops in double-precision.  Computational verses communication cost analysis for weak scaling results on up to 8192 Frontier nodes are presented whilst baseline comparisons confirmed LORACX- GPU’s superior performance.  To demonstrate its effectiveness, LORACX is applied to a molecular dynamics simulations of Molybdenum disulfide (MoS2) recrystallization, successfully identifying structural motifs and phase transition dynamics, thereby highlighting LORACX's potential for materials science discovery.  Additionally, LORACX is applied to clustering a large-scale 20 million x 400,000 protein k-mer dataset into over 100,000 clusters.

Speaker’s Bio:
Benjamin Cobb is a Computer Science Ph.D. student at the Georgia Institute of Technology, co-advised by Richard Vuduc and Haesun Park, who is currently interning at Oak Ridge National Laboratory under the mentorship of Ramakrishnan Kannan and Piyush Sao.  Ben's research is at the intersection of constrained tensor factorizations and high-performance computing.  In his free time, he enjoys hiking, running, and boardgames as well as spending time with his fiancé and their two Pomeranians.