GPU-Aware MPI on Modern Supercomputers
vimeo
Abstract: General-purpose Graphics Processing Units (GPGPU or GPU) have been widely adopted by modern supercomputers due to their superior performance and energy efficiency. However, the separate host and device requires programmers to use multiple programming models, such as CUDA and Message Passing Interface (MPI), to operate on data in different memory spaces. This challenge becomes more difficult to tackle when non-contiguous data in multidimensional structures is used by real-world applications. In this talk, I will introduce the GPU-Aware MPI to unify host memory and device memory using standard interfaces of MPI. The benefits of using GPU-Aware MPI, including simplify programming and optimize performance, will be presented. In the second part of my talk, I will introduce how to redesign applications with the GPU-Aware MPI using a production application, i.e., the multiphase 3D Lattice Boltzmann Method (LBM) in Computational Fluid Dynamics, as an example. I will illustrate the effectiveness of the GPU-Aware MPI with the performance and scalability results of the largest-scale simulation of LBM ever conducted on Titan supercomputer.
Speaker: Hao Wang is a research associate at the Department of Computer Science, Virginia Tech. He got his Ph.D. from the Institute of Computing Technology, Chinese Academy of Sciences. His research focuses on High Performance Computing, in particular, parallel computing architecture and algorithm, and distributed system for big data analytics. He has published more than 30 papers in conferences and journals related to these research areas. Hao designs and implements the GPU-Aware MPI, which has been adopted by open-source MPI libraries MVAPICH2, OpenMPI, and MPICH2, and used by production applications in the supercomputing centers all around the world, including Oak Ridge National Laboratory, Texas Advanced Computing Center, San Diego Supercomputer Center, Swiss National Supercomputing Center, etc.
Last download moments ago
