Lossy Compression in PDE Solvers for Effective Memory Bandwidth Use
Eileen Martin, Stanford University
Lossy compression of a PDE’s state at a given time has been suggested as one way to cope with memory bandwidth limitations, both between main memory and cache, and between main memory and writing to disk. We investigate the role of lossy compression in two applications that require huge amounts of data movement: simulating laser-plasma interactions and the reverse time migration algorithm for seismic imaging.
To explore the utility of possible hardware compression in cache, we explore the use of a small-scale compression algorithm, fpzip, between time steps in an acoustic wave propagation model used for seismic imaging, and in pF3D, a multiphysics model of laser-plasma interactions leading to fusion ignition at the National Ignition Facility.
Only modest small-scale compression rates are acceptable for seismic imaging, but higher compression rates can be achieved by compressing full wave fields before writing those fields to disk. Reverse time migration is a popular imaging algorithm based on summing cross-correlations of simulated wave fields from a well-characterized source and known receiver data, so it is relatively robust to errors introduced by lossy compression. We investigate wavelet and curvelet compression and explore acceptable compression ratios for both two-dimensional and three-dimensional domains.
Abstract Author(s): Eileen Martin, Steve Langer