Supervisor: Jon C. Calhoun (Clemson University)
Abstract: HPC applications require massive amounts of data. The data required is growing faster than memory capabilities. An example of this is pySDC, a framework for solving collocation problems iteratively using parallel-in-time methods. pySDC requires storing and exchange of 3D volume data for each parallel point in time. We evaluate several state-of-the-art lossy compressors such as SZ and ZFP for their applicability to inline compression for pySDC. We evaluate the compressors based on compression ratio, compression bandwidth, decompression bandwidth, and overall error introduced.
This poster evaluates state-of-the-art lossy compressors for use in pySDC; shows lossy data compression is an effective tool for reducing memory requirements for pySDC; and highlights current compression/decompression bandwidth is not fast enough for inline lossy compression yet. Results show using SZ with an error bound of 1e-5, we reduce the memory footprint by a factor of 311.99 while maintaining an acceptable level of loss.
ACM-SRC Semi-Finalist: no
Poster Summary: PDF
Back to Poster Archive Listing