The multi-GPU System with ExpEther
Graduate School of Science and Technology, Keio University, 3-14-1 Hiyoshi Kouhoku-ku Yokohama, Kanagawa 223-8522, Japan
International Conference on Parallel and Distributed Processing Techniques and Applications (PDPTA’12), 2012
@article{nomura2012multi,
title={The multi-GPU System with ExpEther},
author={Nomura, S. and Nakahama, T. and Higuchi, J. and Suzuki, J. and Yoshikawa, T. and Amano, H.},
year={2012}
}
Clusters using multiple GPUs have been already widespread to build a high performance computer economically. However, since the number of plugged GPUs into a CPU is limited, such clusters are consisting of multiple host PCs each of which has a few GPUs. This conventional multi-GPU cluster requires programmers to learn parallel programming skills for controlling communication between nodes as well as GPU programming. In order to show the illusion that a large number of GPUs to a single host, a multi-GPU system with ExpEther is proposed. The multi-GPU system allows interconnecting a single host PC and multiple GPUs by ExpEther which extends PCIe interface to Ethernet. Execution of the application program with two to six GPUs achieved 1.99, 2.96, 3.92, 4.83 and 5.14 times speedup at most, as those with a single GPU. Also, the influence of the bandwidth of the network used in the multi-GPU system is evaluated quantatively.
August 30, 2012 by hgpu
