An Efficient Approach to Lossy Compression with Pointwise Relative Error Bound
Workshop: The 2nd International Workshop on Data Reduction for Big Scientific Data (DRBSD-2)
Authors: Sheng Di (Argonne National Laboratory)
Abstract: An effective data compressor is becoming increasingly critical because of the extremely large volumes of data produced by scientific simulations. Many lossy compressors have been developed in the context of absolute error bounds (e.g., minimizing the maximum absolute error or root-mean-squared error). In order to achieve a multiresolution effect, however, scientific applications may prefer to compress the data with a pointwise relative error bound (i.e., the larger the data value, the larger the compression error to tolerate). In this paper, we explore a lossy compression strategy based on the requirement of pointwise relative error bound, under a state-of-the-art compression framework (prediction + quantization + entropy encoding). Specifically, we split the data set into many small blocks and compute a real absolute error bound for each block based on the pointwise relative error bound set by users. We implement this solution and evaluate it using a real-world scientific data set. Experiments show that the compression ratio of our solution is higher than that of other state-of-the-art compressors by 17.2–618%, with comparable compression/decompression times, the same relative error bound, and similar peak signal-to-noise ratio (PSNR).
Workshop Index