Supervisor: Jon C Calhoun (Clemson University)
Abstract: Per-node memory capacity limits the maximal problem size of HPC applications. Naïve data compression alleviates the memory impact, but requires full decompression before the data is accessed. ZFP compressed arrays reduce the memory footprint, by independently compressing data in fixed sized blocks. Thus, decompressing individual blocks and caching them enables random access and a reduction in decompressions on the critical path. The performance of ZFP compressed arrays is dependent on several key variables: software cache size, cache policy, and compression rate. In this poster, we explore the sensitivity of these ZFP parameters on runtime performance for the matrix-matrix multiplication algorithm. Results show that selection of cache size, policy, and rate yields 8% performance improvement over the default ZFP configuration.
ACM-SRC Semi-Finalist: no
Poster Summary: PDF
Back to Poster Archive Listing