PaSTRI: A Novel Data Compression Algorithm for Two-Electron Integrals in Quantum Chemistry
Abstract: Integral computations for two-electron repulsion energies are frequently used in quantum chemistry. Computational complexity, energy consumption, and the size of the output data generated by these computations scale with O(N4), where N is the number of atoms simulated in the system. In many applications, the same integrals are required to be calculated multiple times. Storing these values and reusing them requires impractical amounts of storage space; whereas recalculating them requires a lot of computations. On the other hand, generated data typically requires much less precision than the built-in floating point data types.
We propose PaSTRI (Pattern Scaling for Two-Electron Repulsion Integrals), a fast novel compression algorithm that makes it possible to calculate these integrals only once, store them, and reuse them at much smaller computational cost than recalculation. PaSTRI is "lossy" compared with floating point numbers but still maintains the precision level required by the integral computations.
PaSTRI is a part of the ECP-EZ project, implemented as one of the compression algorithms in the SZ compressor. We have evaluated our compressor by using the GAMESS dataset and achieved a 17.5:1 compression ratio, whereas the compression ratio for the original SZ was 8.0:1 and for ZFP was 7.1:1.
Bio: Ali Murat Gok is a Ph.D.& student from Northwestern University. He received his master's degrees from Northwestern University and Bogazici University. His research interests are energy-efficient parallel architectures, approximate computing, power efficiency and reliability, hardware characterization, voltage overscaling, and compression algorithms. He is a summer intern at Argonne, working with Franck Cappello, Sheng Di and Dingwen Tao in the data compression team under the ECP-EZ project.