Poster 145: Improving Data Compression with Deep Predictive Neural Network for Time Evolutional Data

Authors: Rupak Roy (Florida State University), Kento Sato (RIKEN Center for Computational Science (R-CCS)), Jian Guo (RIKEN Center for Computational Science (R-CCS)), Jens Domke (RIKEN Center for Computational Science (R-CCS)), Weikuan Yu (Florida State University), Takaki Hatsui (RIKEN SPring-8 Center), Yasumasa Joti (Japan Synchrotron Radiation Research Institute)

Abstract: Scientific applications/simulations periodically generate huge intermediate data. Storing or transferring such a large scale of data is critical. Fast I/O is important for making this process faster. One of the approaches to achieve fast I/O is data compression. Our goal is to achieve a delta technique that can improve the performance of existing data compression algorithms for time evolutional intermediate data.

In our approach, we compute the delta values from original data and data predicted by the deep predictive neural network. We pass these delta values through three phases which are preprocessing phase, partitioned entropy coding phase, and density-based spatial delta encoding phase.

In our poster, we present how our predictive delta technique can leverage the time evolutional data to produce highly concentrated small values. We show the improvement in compression ratio when our technique, combined with existing compression algorithms, are applied on the intermediate data for different datasets.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF


Back to Poster Archive Listing