Performance Tuning SLATE
Abstract
Software for Linear Algebra Targeting Exascale (SLATE) 1 is being developed as part of the Exascale Computing Project (ECP) 2, which is a joint project of the U.S. Department of Energys Office of Science and National Nuclear Security Administration (NNSA). SLATE will deliver fundamental dense linear algebra capabilities for current and upcoming distributed-memory systems, including GPU-accelerated systems as well as more traditional multi coreonly systems. SLATE provides coverage of existing LAPACK and ScaLAPACK functionality, including parallel implementations of Basic Linear Algebra Subroutines (BLAS), matrix norms, linear systems solvers, least squares solvers, and singular value and eigenvalue solvers. In this respect, SLATE will serve as a replacement for ScaLAPACK, which, after two decades of operation, cannot be adequately retrofitted for modern, GPU-accelerated architectures. This working note focuses attention on several performance issues that existed in SLATE, and how these were resolved.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 02, 2020
- Accession Number
- AD1112729
Entities
People
- Ali Charara
- Asim Yarkhan
- Dalal Sukkari
- Jack Dongarra
- Mark Gates
- Mohammed Al Farhan
Organizations
- University of Tennessee system