2014年4月25日学术报告--Discovery of Graphs for

题目：Discovery of Graphs for Designs of Supercomputer Networks
报告人：邓越凡教授 (美国Stony Brook University)
主持人：姚道新教授、梁世东教授
时间：2014年4月25日(星期五) 14:30-16:00
地点：中山大学南校区高等学术中心（冼为坚堂）211报告厅

报告摘要：
Supercomputers, capable of performing 10^16 floating-point operations per second (34PFlops), connect millions of computing cores by complex networks. In the case of the Tianhe-2, 3.12 million cores require networking. In 2019 when Exascale systems emerge, more than 100 million cores will need to be connected and the traditional intuition with simple networks will unlikely survive to produce scalable systems. To advance, we must leverage on the, mature and active, graph theory, conceptually and computationally, to seek for new breakthrough in network topologies and routing protocols.

For a regular graph of N vertices each with k edges, we note it as Nkk, e.g., for a regular graph with N=32 vertices each with k=5 edges, we note it as 32k5. After discovering a series of perfectly optimized (in terms of the graph diameter and average vertex-vertex edge distances) regular graphs with N=4,8,16,32 for the appropriate corresponding node degreesk=2,3,4,5,we embed themto generate much larger composite hierarchical graphs, e.g., 16k4⨂32k5 or 8k3⨂16k4⨂32k5. These embedded graphs, with tens of thousands of vertices, are used to design supercomputer interconnection networks. With the metrics we introduced to measure the network-performance-relevant properties of graphs, we compare our quasi-optimal embedded graphs with many widely adopted networks for supercomputers.

Joint work with M. T. Michalewicz and his staff of Singapore’s Agency of Science, Technology, and Research during the author's sabbatical.

报告人简介：Yuefan Deng is Professor of Applied Mathematics at Stony Brook University, a Visiting Scientist at Brookhaven National Laboratory, and has worked at IBM in the design of BlueGene supercomputers. His research is in Parallel Computing, Molecular Dynamics, Monte Carlo Methods, and Computational Science. He published more than 70 papers and supervised 20 doctoral theses. He is the architect of Galaxy Beowulf Supercomputer at Stony Brook built in 1997 and of NankaiStars Supercomputer which was China's fastest computer when it was completed in 2004. He lectured widely in US, Germany, Russia, Brazil, South Korea, Saudi Arabia, Turkey, as well as the Greater China region. US DOE, NSF, NIH，as well as China's Ministry of, and Shanghai's Commission of, Science and Technology have supported his research. Professor Deng earned his BA in 1983 in Physics from Nankai University and his PhD in 1989 in Theoretical Physics from Columbia University.