Litcius/Paper detail

Productive and Performant Generic Lossy Data Compression with LibPressio

Robert Underwood, Victoriana Malvoso, Jon C. Calhoun, Sheng Di, Franck Cappello

202118 citationsDOI

Abstract

In recent years, lossless and lossy compressors have been developed to cope with the ever increasing volume of scientific floating point data. However not all compression techniques are appropriate for all data-sets, and determining which one to use can be time consuming requiring code modifications and trial and error. We present LibPressio - a generic library for the compression of dense tensors that minimizes the code changes scientists need to make to take advantage of new and improved compression techniques. We compare LibPressio to 9 different competing libraries and measure the overhead of their design decisions as well as overall run time overhead showing insignificant overhead. We further show an improvement in usability as measured by a reduction in lines of code compared to native code by 50–90 %. The value of this tool can be seen by integration into Z-Checker and ADIOS2.

Topics & Concepts

Lossy compressionLossless compressionComputer scienceOverhead (engineering)Code (set theory)Data compressionCompression (physics)Volume (thermodynamics)UsabilityComputer engineeringAlgorithmProgramming languageOperating systemMaterials scienceComposite materialSet (abstract data type)Quantum mechanicsPhysicsParallel Computing and Optimization TechniquesComputational Physics and Python ApplicationsNumerical Methods and Algorithms