General Purpose Computing Using GPUs on a Linux Cluster:
An Introduction and Programming Practicum
Hosted by HPCMP PET - FMS.
21 through 23 October 2008
ISI, Marina del Rey, CA
Map of the Marina del Rey Area
Course Background The simulation community has often been hampered by constraints in computing: not enough resolution, not enough entities, not enough behavioral variants. High Performance Computing (HPC) can ameliorate those constraints. The use of Linux Clusters is one path to higher performance; the use of Graphics Processing Units (GPU) as accelerators is another. These are often called General Purpose GPUs (GPGPUs). Merging the cluster and GPGPU paths holds even more promise. The ISI team members were the principal architects of a successful proposal to the High Performance Computing Modernization Program (HPCMP) for a new 512 CPU (1024 core), GPU-enhanced Linux Cluster, joshua, for the Joint Forces Command’s Joint Experimentation Directorate (J9). This cluster was awarded to J9 via the Dedicated HPC Project Investment program, DHPI and was configured in such a way to putatively utilize the GPUs to increase performance by a factor of two or more, with concomitant savings in cost, power and space. Offering this course will aid JFCOM as they work to take full advantage of the new JFCOM cluster. One of the unique aspects of the new machine is that there is a state-of-the-art NVIDIA GPU in each node. These GPGPUs can be programmed using the new CUDA programming language (Compute Unified Device Architecture, a "C-like" language). JFCOM has a need to improve simulation performance on the new cluster and will make modifications to simulation programs such as JSAF, enabling them to take advantage of heterogeneous HPC architecture. As JFCOM conducts their Noble Resolve experiments focused on emergency responses in urban areas, performance needs to be improved to adequately model the complexities of these crowded areas and large populations. They will need to make the most of the new GPGPU-enhanced Linux cluster asset. Programming models, code examples and practice problems in CUDA will be presented and implemented in class. |
0. Data from 2007 Course Itself
0.1.1Gene's SLIDES from this version of the Course
0.2 Pictures of Previous Course in Suffolk VA and Participants
2. Date and Location:
21, 22 and 23 October 2008
11th Floor Conference Rooms, 1135 & 1137
Information Sciences Institute
Marina del Rey, CAQuestions?? email Dan Davis
Setting, Room and pictures from previous colloquium there.
Hotels in the Marina del Rey area.
Restaurants in the Marina del Rey area.
3. Agenda and Speakers
3.1 Staff and Participant Speakers
PET FMS FAPOC - Dr. David Pratt
Class Host and Leader - Dan Davis
Practicum Leader - Gene Wagenbreth
3.2 Schedule
Tuesday
(all invitees)
0900 - Welcome and User Needs - Dan Davis
0930 - GP & GPGPU Vision, history, and opportunity - Dr. David Pratt or Chris Bouwens
1000 - Early GPGPU work and uses - Dr. Robert F. Lucas
1100 - Introduction to GPGPU concepts
1200 - Lunch 1300 - Hardware Architecture - Staff
1500 - Software Architecture and Programming Model - Staff
1600 - Community needs and JSAF program architecture -to be announced
1700 - Secure for Day
Wednesday
(serious coders only)
0900 - Programming experience and examples
1030 - Performance tuning and optimization
1200 - Lunch
1300 - Coding Practicum for Journeyman Programmers - Gene Wagenbreth
1700 - Secure for Day
Thursday
(survivors only)
0900 - Exercise in programming model for sample codes
1100 - Work on participant projects
1200 - Lunch
1300 - Coding Practicum for Journeyman Programmers
1500 - Evaluation of Coding Efforts by Attendees
1600 - Discussion of future use of joshua, coding, input for nVidia, etc.
1700 - Evaluations and class wrap up
