Z-checker: A framework for assessing - arXiv

1 Research PaperZ-checker: A framework for assessinglossy compression of scientific dataDingwen Tao1, Sheng Di2, Hanqi Guo2, Zizhong Chen1,3and Franck Cappello2,4 AbstractBecause of the vast volume of data being produced by today s scientific simulations and experiments, lossy data com-pressor allowing user-controlled loss of accuracy during the compression is a relevant solution for significantly reducingthe data size. However, lossy compressor developers and users are missing a tool to explore the features of scientific datasets and understand the data alteration after compression in a systematic and reliable way. To address this gap, we havedesigned and implemented a generic framework called Z-checker.

2 On the one hand, Z-checker combines a battery of dataanalysis components for data compression. On the other hand, Z-checker is implemented as an open-source communitytool to which users and developers can contribute and add new analysis components based on their additional analysisdemands. In this article, we present a survey of existing lossy compressors. Then, we describe the design framework ofZ-checker, in which we integrated evaluation metrics proposed in prior work as well as other analysis tools. Specifically,for lossy compressor developers, Z-checker can be used to characterize critical properties (such as entropy, distribution,power spectrum, principal component analysis, and autocorrelation) of any data set to improve compression lossy compression users, Z-checker can detect the compression quality (compression ratio and bit rate) and providevarious global distortion analysis comparing the original data with the decompressed data (peak signal-to-noise ratio,normalized mean squared error, rate distortion, rate-compression error, spectral, distribution, and derivatives)

3 Andstatistical analysis of the compression error (maximum, minimum, and average error; autocorrelation; and distribution oferrors). Z-checker can perform the analysis with either coarse granularity (throughout the whole data set) or finegranularity (by user-defined blocks), such that the users and developers can select the best fit, adaptive compressors fordifferent parts of the data set. Z-checker features a visualization interface displaying all analysis results in addition to somebasic views of the data sets such as time series. To the best of our knowledge, Z-checker is the first tool designed to assesslossy compression comprehensively for scientific data , lossy compression, assessment tool, data analytics, scientific data, visualization1.

4 IntroductionOne of the most challenging issues in performing scien-tific simulations or running large-scale parallel applica-tions today is the vast amount of data to store in disks, totransmit on networks, or to process in postanalysis. TheHardware/Hybrid Accelerated Cosmology Code (HACC),for example, can generate 20 PB of data for a single one-trillion-particle simulation(Habib et al., 2016); yet a sys-tem such as the Mira supercomputer at the Argonne Lead-ership Computing Facility has only 26 PB of file systemstorage, and a single user cannot request 75%of the totalstorage capacity for a simulation. In climate research, asindicated by Gleckler et al. (2016), nearly PB of datawere produced by the Community Earth System Model(CESM) for the Coupled Model Intercomparison Project(CMIP), which further introduced 170 TB ofpostprocessing data submitted to the Earth System Grid(Bernholdt et al.)

5 , 2005). Estimates of the raw data require-ments for the CMIP6 project exceed 10 PB (Baker et al.,2014). Particle accelerator based research also deals with1 Department of Computer Science and Engineering, University ofCalifornia, Riverside, CA, USA2 Division of Computer Science andMathematics, Argonne NationalLaboratory, Lemont, IL, USA3 Beijing University of Technology, Beijing, China4 Parallel Computing Institute, University of Illinois Urbana Champaign,Champaign, IL, USAC orresponding author:Franck Cappello, Mathematics and Computer Science Division, ArgonneNational Laboratory, 9700 Cass Avenue, Lemont, IL 60439, : International Journal of HighPerformance Computing Applications1 19 The Author(s) 2017 Reprints and.

6 Large volume of data during experiment and postanaly-sis. The Advanced Photon Source (APS) Upgrade (Austin,2016) is the next-generation APS project at Argonne; oneof its targets is mouse brain images of by APS-based researchers at Argonne, the instru-ments can produce 150 TB of data per specimen, with atotal of at least 100 specimens for storage. Also involvedare 35 TB of postprocessing data. Obviously, effectivedata compression is significant for the success of today sscientific compressors are commonly considered the rele-vant solution for significantly shrinking the data size forscientific research, while still carefully controlling the lossof data accuracy based on user demand.

7 However, scien-tific data sets are often composed of floating-point dataarrays, which lossless compressors such as GZIP (Deutsch,1996) cannot compress effectively (Baker et al., 2016).Arguably, many floating-point data-based lossy compres-sors have been proposed, some of which are consideredrelevant solution for climate simulations by guaranteeingthe validity of data after decompression (Baker et al., 2016;Sasaki et al., 2015). However, many other types of scien-tific research issues need lossy compression, and the impactof lossy compressors on their results is still unclear. More-over, users want to be able to select the best-fit compressorsfor their specific needed, then, an easy-to-use, generic, compre-hensive assessment tool for users to understand the effec-tiveness of lossy compressors and their impact on theirscientific results.

8 The key challenges in designing such atool are the masses of evaluation metrics to deal with andthe various design principles across lossy the one hand, scientific researchers often have differ-ent targets with different requirements on the accuracy ofdata. For instance, graph-processing researchers focusmainly on peak signal-to-noise ratio (PSNR) (Taubmanand Marcellin, 2012), while climate researchers may alsobe concerned with other metrics such as the correlationcoefficient of the original data and decompressed data(Baker et al., 2014). Cosmology researchers, as indicatedby the HACC development team at Argonne, are con-cerned with the distribution of the compression errors andtheir autocorrelations.

9 On theother hand, lossy compres-sors may have largely different designs, such that theircompression results can be different even with the samedata sets. In particular, we observe that (Tao et al.,2017) has a uniform distribution of the compression errorsbut ZFP has a normal distribution and that ZFP has higherautocorrelation of the compression errors than (see the experimental evaluation section fordetails). Thus, it is a nontrivial task to design a genericassessment tool for assessingmultiple lossy compressorsin a comprehensive and fair this work, we present a flexible, generic, open-sourcelossy compression assessment tool called Z-checker,which can be used to evaluate the compression qualityof various lossy compressors.

10 Z-checker is developed inthe context of the Exascale Computing Project (ECP) torespond Exascale application needs. These applicationsare currently not using lossy compression. Also, sincethere is no tool to provide a comprehensive assessmentof compression quality (such as compression error),Z-check is designed to address this gap. Our key contri-butions are as follows: We design a flexible assessment framework , calledZ-checker, that integrates a series of assessmentmodules and visualization tools. In particular, thisframework allows users to assess the lossy compres-sion qualities in multiple ways (such as readingstandalone data files, library calls in the compressioncodes, and external calls of compression executablesfrom Z-checker).