2020/01/16 by Robert Underwood, Underwood, Robert, Sheng Di +5
Computer Science · #Advanced Data Storage Technologies #Parallel Computing and Optimization Techniques #Numerical Methods and Algorithms
paper · pdf · doi:10.48550/arxiv.2001.06139
With ever-increasing volumes of scientific floating-point data being produced\nby high-performance computing applications, significantly reducing scientific\nfloating-point data size is critical, and error-controlled lossy compressors\nhave been developed for years. None of the existing scientific floating-point\nlossy data compressors, however, support effective fixed-ratio lossy\ncompression. Yet fixed-ratio lossy compression for scientific floating-point\ndata not only compresses to the requested ratio but also respects a\nuser-specified error bound with higher fidelity. In this paper, we present\nFRaZ: a generic fixed-ratio lossy compression framework respecting\nuser-specified error constraints. The contribution is twofold. (1) We develop\nan efficient iterative approach to accurately determine the appropriate error\nsettings for different lossy compressors based on target compression ratios.\n(2) We perform a thorough performance and accuracy evaluation for our proposed\nfixed-ratio compression framework with multiple state-of-the-art\nerror-controlled lossy compressors, using several real-world scientific\nfloating-point datasets from different domains. Experiments show that FRaZ\neffectively identifies the optimum error setting in the entire error setting\nspace of any given lossy compressor. While fixed-ratio lossy compression is\nslower than fixed-error compression, it provides an important new lossy\ncompression technique for users of very large scientific floating-point\ndatasets.\n