An Acceleration Toolkit of Matlab based on GPU clusters

• Main Authors: Jyun-Kai Wu, 吳浚楷
• Other Authors: Tyng-Yeu Liang
• Format: Others
• Language: zh-TW
• Online Access: http://ndltd.ncl.edu.tw/handle/21445833939470570008

碩士 === 國立高雄應用科技大學 === 電機工程系 === 100 === This research is aimed at developing an acceleration toolkit of Matlab called ATOM based on GPU clusters. With the support of this toolkit, the instructions of matrix operation from Matlab will be captured and redirected to ATOM servers for parallel computing. Because the computational ability of devices are different in GPU-cluster, ATOM supports the load balance mechanism for utilizing resource sufficiently. Each computational device is assigned with a proper amount of computational data to achieve load balance and to increase the execution speed of Matlab by the load balance mechanism. In addition, for decreasing unnecessary communication cost, ATOM imports applies data cache and lazy data-update protocol to minimize the communication cost of distributing data over GPU clusters for parallel computing. The concept of data cache is to let users upload data onto ATOM servers and then the servers need not to fetch data from Matlab during data computation. The lazy-update protocol is not to maintain the consistency of the cached data unless the data is acquired. The experiments show that ATOM can exploit GPU clusters effectively by using the above mechanisms to improve the performance obviously.