Performance Improvement Of Multi-scale Modeling Of Concrete Performance Solver Using Hybrid Cpu-gpu System

The multi-scale modeling of concrete performance solver is a tool developed at University of Tokyo. It is used to help engineers predict the material and structural properties of concrete structures under coupled mechanistic and environmental actions. The current implementation of multi-scale modeling of concrete performance solver is developed for Central Processing Unit (CPU) and CPU clusters and has longer runtimes. In this work, we present a CPU- Graphics Processing Unit (GPU) hybrid implementation of the tool. Changes were made to the LU factorization module by moving the computations to the GPU after profiling revealed it to be compute intensive. Tests conducted on the modified tool using real world sample concrete structure show that using the hybrid CPU-GPU implementation improved the performance by 35.29% on single node and 37.5% on cluster.rnKey Words:, Hybrid CPU-GPU system, General Purpose Computation on Graphics Processing Units, Multi frontal Massively Parallel Solver, Matrix Algebra on GPU and Multicore Architectures

Get Full Work

Report copyright infringement or plagiarism