High performance data processing system in cloud : implement MARS on multiple GPU

• Main Author: Nguyen, Tran Quoc
• Other Authors: School of Computer Engineering
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/59253
• Institution: Nanyang Technological University

Map-Reduce is a framework for processing parallelizable problem across huge datasets using a large computation power and Graphic Processing Unit (GPU) is suitable to solve parallel problems. MARS has been introduced as one of most effectiveness Map-Reduce framework for GPU. MARS aims to help developer utilize all the compute power without knowing much about GPU programming However, MARS is still not scalable, which can only run on one node with one GPU. This makes MARS not suitable for processing large amount of data – an inevitable problem in nowadays computing world. By using advantage of the new software develop toolkit (SDK) of CUDA which allow GPUs communicates with each other through PCI-E, the student has improved MARS to run on multiple GPUs. Besides, he also collaborated with other student to make MARS can run on multiple nodes. In this report, the student would explain in details how MARS can use multiple GPUs to achieve its goal as well as the benchmark and the difficulties faced during the course of the final year project