Research On Memory Access Characteristics For Memory Optimization

Improving the performane of computers is always the final goal of researchrers. With the development of large scale integrated circuit, computing power of the processor is rising rapidly. The bottleneck of computer performance has transformed from computing to memory accessing. Both memory access characteristics of programs and memory structure affect the memory performance. The study of memory access characteristics has been accompanying the development of processors, providing direction for the storage structure.This thesis selects seceral applications with typical memory access characteristics, analyzing the characteristics and making memory optimization:1) In the Cache-shared memory architecture, we use Simics+GEMS architecture simulator to analysis memory access characteristics of different data sets of on-line transaction processing, which is based on PostgreSQL database, and create a data classification model into three types: protected, free-competitive and abandoned; then propose a software co-translucented shared Cache structure to distinguishing three types of data sets. Experimental results show that Cache missing rate drops 12% at most.2) With characteristics of matrix rank-alternated access and features of DRAM memory, we propose alternated optimization memory access model--window access for matrix row-column access; the window access model is implemented on our reconfigurable experimental platform into a new memory controller, getting that meory performance of matrix rank-alternated access increases to 73.6% and one-dimensional parallel FFT algorithm performance can be improved by 45.1%.3) We use window access principles to guide the layout of the data matrix on CPU and GPU. After optimizing the programs, matrix rank-alternated access on CPU gets an imcrement of 58.4% in memory performance mostly; Banding window access and FFTW to one-dimensional parallel FFT algorithm, computing performance increases up to 7%, comparing with FFTW. With analysis of CUDA programming framework and the corresponding structural features of GPU, the window access principle is applied to the warp,which is the unit of GPU memory access process, the experiments show that matrix rank-alternated access achieves 1-fold performance improvement.