-
Notifications
You must be signed in to change notification settings - Fork 3
Teams
Other teams may be interested in our prototype code which runs on the GPUs (with PGI 15.7) and has a ~4.4x speedup compared to OMP. It can be found here: https://github.com/BoxLib-Codes/ode-openacc
The 4.4x speedup is achieved in the code in the first-order/ directory. We implement here a simple first-order ode integrator operating on a simple network of 3 isotopes. first-order/NOTES contains some notes about the code and lessons learned.
Team Members
- Adam Jacobs Stony Brook University
- Philip Jones Los Alamos National Lab
- Max Katz Stony Brook University
- Chris Malone Los Alamos National Laboratory
- Donald Willcox Stony Brook University
Mentors
- Oscar Hernandez ORNL
- Markus Wetzstein CSCS - Swiss National Supercomputing Centre
The Geophysical Fluid Dynamics Lab team is representing one of NOAA's OAR Coupled Climate model shops for long term climate and high impact weather research. Our model has been in development for many decades, widely supporting research in this domain. It has demonstrated good strong and weak scaling on x86 and BGQ platforms with our hybrid MPI+OpenMP model. We have been exploring systems and methods to leverage emerging novel hardware architectures and Fortran-friendly compiler developments such as OpenACC into our code base. Our efforts at the Hackathon has highlighted challenges with our code and the state of available compilers as applied to our code. However, we were able to demonstrate as a side project the raw power of modern GPUs as applied to a completely fresh implementation of a Radiation domain specific computation using simple unoptimized CUDA C and believe that OpenACC can deliver similar successes going forward.
Team Members
- Jeffrey Durachta GFDL NOAA, Princeton NJ
- Zhi Liang GFDL/Engility
- Garrett Wright NOAA GFDL Affiliate
Mentors
- Eric Dolven Cray, Inc
- Simon Layton NVIDIA
We are hacking the Accelerated Climate Model for Energy, a coupled Earth system model for DOE climate research. The focus for this years' Hackathon will be continued work on the CAM atmosphere component (especially the dynamics) and an initial effort to port the new MPAS-Ocean model using OpenACC.
Team Members
- Rick Archibald ORNL
- Peter Bosler Sandia National Laboratories
- Glenn Brook University of Tennessee
- irina Demeshko Sandia National Laboratories
- Kate Evans ORNL
- Salil Mahajan ORNL
Mentors
- Dave Norton The Portland Group
- Matt Norman ORNL
NUMA is the Nonhydrostatic Unified Model of the Atmosphere (see http://faculty.nps.edu/fxgirald/projects/NUMA). NUMA was developed at the Department of Applied Mathematics of the Naval Postgraduate School in Monterey California. NUMA is a highly scalable model that has been shown to scale at 99.1% strong scaling efficiency on Mira (ALCF) using the entire 3 million MPI threads available. The results of this study can be found in this paper: https://www.researchgate.net/publication/282887372_Strong_Scaling_for_Numerical_Weather_Prediction_at_Petascale_with_the_Atmospheric_Model_NUMA
NUMA also runs on GPUs thanks to the OCCA library. Progress reports on this portion of the work can be found here: https://www.earthsystemcog.org/projects/espc-numa. NUMA, with OCCA spitting out CUDA kernels, has been shown to run at about 90% weak scaling efficiency using up to 16384 GPUs on Titan (OLCF).
Our goal at this hackathon is to compare our OCCA implementation of NUMA with an OpenACC version. We wish to measure 1) performance and 2) amount of time required to port.
Team Members
- Daniel Abdi Naval Postgraduate School
- Frank Giraldo Naval Postgraduate School
Mentors
- Mat Colgrove NVIDIA
- Kshitij Mehta Total E&P (onsite at ORNL)
- Matthew Otten Cornell University
Team Members
- Sven Binder University of Tennessee, ORNL
- Andreas Ekstrom University of Tennessee
- Gaute Hagen ORNL
- Gustav Jansen ORNL
- Thomas Papenbrock University of Tennessee
- Micah Schuster ORNL
Mentors
- Wayne Joubert ORNL
- Jeff Larkin NVIDIA
We are working to speed up a 2D physics-based model of the ionosphere [Sami2 is Another Model of the Ionosphere] (http://www.nrl.navy.mil/ppd/branches/6790/sami2). We have had some success in using OpenMP to improve the performance of the code and at the Hackathon we are targeting the same areas of the code using OpenACC.
Team Members
- Koushik Ghosh Engility Corporation
- Raymond Menzel Engility Corporation
- Michael Stokes Engility Corporation
- Christopher Kung Engility Corporation
- Sarah McDonald Naval Research Laboratory
- Sam Cable Naval Research Laboratory
- Douglas Drob Space Science Division, Naval Research Laboratory
- Guido Klingbeil Okinawa Institute of Science and Technology
Mentors
- Carl Ponder NVIDIA
- Brent Leback PGI
Team Members
- Ewart Timothee EPFL Blue Brain Project
- Stuart (Sam) Yates EPFL Blue Brain Project
- fabien delalondre EPFL Blue Brain Project
Mentor
- Jakob Progsch Nvidia
Our application is a Toolbox for Urban Mobility Simulations (TUMS), which models urban transportation system. TUMS is a data-driven simulation platform that utilizes large volumes of heterogeneous data sources as an input to model and simulate a suite of very complex scenarios (such as evacuation and special event population). It estimates the impact of such scenarios on traffic engineering, planning, and dynamics of urban systems. There are three main components to this toolset - i) Data processing for generating initial input, ii) Simulation using TRANSIMS software, iii) Middleware and visualization. At Hackathon, we intend to scale TUMS to use OpenACC directives to program GPUs. This in turn will helps us to utilize TUMS to use large geographical area (and eventually continental USA) simulating millions of vehicles at a finer granularity and at the same time.
Team Members
- Rajasekar Karthik ORNL
- Cheng Liu ORNL
- Dilip Patlolla ORNL
- Gautam Thakur ORNL
Mentors
- Mark Berrill ORNL
- Adam Simpson ORNL
Application area:
XGC1 is a gyrokinetic particle-in-cell code, which can include the magnetic separatrix and the biased material wall. Lagrangian equation of motion is used for time advancing the particles, while conserving the mass, canonical angular momentum, and energy.
Approach:
The electron particle push and multi-species collision are two of the computational expensive computational modules. The team has extracted simplified serial versions of the two modules for porting to GPU using OpenACC.
Team Members
- Jianying Lang Princeton Plasma Physics Laboratory
- Eisung Yoon Rensselaer Polytechnic Institute
- Ed D'Azevedo ORNL
Mentor
- Levi Barnes NVIDIA