Optimization of finite element analysis code for software controlled local memory

Kushida, Noriyuki; Takemiya, Hiroshi

In this study, we introduce a novel implementation of FEM targeting the Cell which has a software controlled local (SCLM) memory. In recent days, developers of scientific or engineering numerical simulation code suffer from the big latency of data transfer between a processor and main memory, and power consumption. To overcome these problems, several researchers have proposed new computer architecture, which combines SCLM and SIMD processing unit. FEM is the one of the most famous numerical simulation method especially for the engineering. Such new computer architecture is, however, less effective for FEM than other applications. The Cell is well known as the accelerator of the Roadrunner which is the fastest computer in the Top500 supercomputer list, and is only existing processor which has software controlled local memory. In this study we developed new FEM implementation method which provides faster computation than conventional method by reducing main memory access.