Presentation
Compression of Scientific Simulation Data by Stochastic Basis Expansion – Example on Multiple Computer Systems
DescriptionTargeting scientific data storage and processing of large-scale complex problems, we consider the introduction of stochastic basis expansions for compressing and restoring data with low conversion costs by the effective use of high-performance computer systems and file systems. Here we target Monte Carlo simulation and use high-order stochastic basis expansion to reduce data sizes while suppressing the degradation of accuracy. In order to reduce the analysis cost involved in high-order stochastic expansions, we used a scalable algorithm that can efficiently utilize high-performance computer systems and applied the method to static and dynamic elastic problems. Using the nearly full system of CPU-based Fugaku (147456 nodes), we show that 1300 trillion degrees-of-freedom storage required in Monte Carlo analysis can be compressed by 35-fold to 37 trillion degrees-of-freedom and also show that the same compression rate can be attained using 5120 nodes of AMD GPU-based Frontier (54% of the system) with high efficiency.
