Small Scale project, 2020 Jove Matrix Performance

Researcher

Benjamin Thomitzni, PhD Student in the Theoretical and Computational Chemistry Group

Initial Problem

• Matrix multiplication code has poor single core performance
• Performance also doesn't scale beyond 4 threads.

What we did

• Exploit symmetry to reduce required operations
• Change data layout and re-order loops to avoid cache misses
• Use a specialized linear algebra library to generate optimized code
• Add thread-safe and performant parallelism using OpenMP

Results

• Optimized code runs more than 4x faster on a single core with small test dataset
• With a larger dataset this increases to 8x performance improvement on a single core
• Near perfect parallel scaling on a 12-core machine with small test dataset
• Near perfect parallel scaling on a 56-core machine with larger dataset