Computational Science Community Wiki

GPU Club: Weds 13 Nov


alt="NVIDIA CUDA Research Centre"

  • Software for GPUs inc. compiler/directives, maths libs & tools (debuggers and profilers)


2pm, Weds 13 Nov. Room 4.206, University Place

Application of accelerators to reactive chemistry in WRF-Chem

Slides: Lessons Learned from using Accelerators in Chemical Mechanisms

Numerical solution of reactive processes is the most computationally costly part of atmospheric chemistry simulations. For one standard configuration of WRF-Chem, the atmospheric chemistry version of the Weather Research and Forecast model, chemical kinetics and advection cost many times more than the rest of the model. Although the chemistry module is highly parallel – the solution at each grid cell is independent – the solver itself has very low computational intensity and a large working set, which present challenges for accelerators such as GPUs and Intel’s Xeon Phi. We present a number of approaches, both hand coded and automatically generated, for adapting a chemical kinetics kernel to accelerators and discuss the resulting computational performance.

Authors:
John Michalakes, Environmental Modeling Center, NOAA National Centers for Environmental Prediction, U.S.A.
John Michalakes has twenty-five years experience in high-performance computing applied to atmospheric and related geophysical modeling systems, including HPC software design, parallel libraries, performance analysis, modeling, and optimization, and novel architectures (e.g. MIC and GPU). He developed first distributed memory parallel version of the Penn State/NCAR Mesoscale model and led the software working group that developed the Weather Research and Forecast (WRF) model. Current research involves use of accelerators to improve node-performance and strong scaling of applications. He earned a master of science degree in computer science from Kent State University (Ohio) in 1988. He is currently a scientific programmer/analyst for the Environmental Modeling Center of the U.S. National Oceanic and Atmospheric Administration.
Email: John.Michalakes@noaa.gov

John C. Linford,ParaTools, Inc.,
Dr. John Linford is the author of Kppa: the Kinetic PreProcessor Accelerated. His research interests include emerging multi-core computer architectures, heterogeneous parallelism, automatic code generation and tuning, performance analysis, and numerical simulations of complex Earth systems. He has developed high-performance modeling and simulation software for large-scale supercomputers, Eclipse-based development environments, toolchains for the Cell Broadband Engine Architecture, and airborne real-time signal processing software. Dr. Linford received his Ph.D. in computer science from Virginia Polytechnic Institute and State University where his dissertation was selected as the Outstanding Doctoral Dissertation in Computer Science. He contributed to the Scalasca project as a guest researcher at the Jülich Supercomputing Centre in Jülich, Germany. He is a National Defense Science and Engineering Graduate Fellow and Central European Summer Research Institute Fellow.
Email: jlinford@paratools.com

An Introduction to Dataflow Computing and Its Application to Meteorological Modelling

Slides: Introduction to Dataflow

Whilst dataflow programming has been around for many decades, it's prominence is markedly on the rise as the demand for parallelism increases to handle the complexity of modelling and simulation within scientific computing. Maxeler Technologies has developed dataflow platforms and compiler technology since 2003 to directly target this complexity - the classes of Big Data and Big Science applications - and provide acceleration performance of 20 times to 200 times faster than conventional Control Flow approaches. In this talk, we present an overview of the Maxeler Dataflow compiler technology and paradigms with special emphasis given to its application in Hydrostatic Modelling. Results of an implementation jointly developed by Maxeler and CRS4 Lab in Italy are presented where Maxeler dataflow engines improve performance by 74X over conventional CPU cores.

Speaker Bio
Craig Davies, Business Development, Maxeler Technologies Ltd.
Craig Davies has over thirteen years experience in High Performance and Embedded Computing having previously worked as an Hardware Engineer at Maxeler Technologies and in previous roles at VMETRO and ICS. With a history designing FPGA based solutions for the Aerospace and Defence markets, his move to Maxeler brought exciting new opportunities including the development of the PCIe and memory systems in use on the Maxeler hardware platforms. Today his remit covers both business and technical aspects including instructing for MaxCompiler training courses and administering the MAX-UP Maxeler University Program.