Design and Implementation of Parallel Processing Platform on Multi-core System Based on Shared Memory Architecture

• Main Authors: Shao-Chun Kuo, 郭紹群
• Other Authors: JIh-ching Chiu
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/k7d8b9

碩士 === 國立中山大學 === 電機工程學系研究所 === 106 === Abstract Nowadays, there are two kinds of parallel programming models for developing parallel applications which are Procedure-Oriented models and Behavior Description models. Using directives offering from Behavior Description models, the programmer can easily parallelize the code Semi-Automatically. There is restriction of parallelism because of the rules of Behavior Description models. As a result, performance of the code parallelized by Behavior Description models is restricted. Procedure-Oriented models offer full-custom development platform so they have much more flexibility than Behavior Description models. Programmer can arrange how to parallelize the code via Behavior Description models. In addition, with user’s knowledge of parallelism, the code parallelized by Procedure-Oriented models can have high performance. This thesis propose Parallel Processing Platform on Multi-core System Based on Shared Memory Architecture (MSSM) which is a kind of Procedure-Oriented models to make programmer can easily develop parallel processing programs in C language on Windows. Programmer can use MSSM’s API to create multiple processes easily to do multicore processing. Besides, via API offering from MSSM, programmer can allocate physical memory as shared memory in user mode in order to decrease Inter-Process Communication latency. MSSM is composed of Main Program、Computing Process、Kernel Mode Driver. Main Program creates multiple Computing Processes to do multicore processing and use MSSM’s API to order Kernel Mode Driver allocate physical shared memory. For testing the performance of MSSM, this thesis implemented three kinds of algorithms which are Advanced Encryption Standard (AES), Sum of Absolute Differences (SAD) and K-Means Clustering by MSSM OpenMP, Parallel Patterns Library (PPL), Message Passing Interface (MPI) and compared the performance of these models. In average, MSSM is 30% faster than PPL, 13% faster than OpenMP and 14% faster than MPI , proving that MSSM has high performance of programming parallel processing Applications.