Dagstuhl-Perspektiven-Workshop 15342
Power-Bounded HPC Performance Optimization
( 16. Aug – 21. Aug, 2015 )
(zum Vergrößern in der Bildmitte klicken)
Permalink
Bitte benutzen Sie folgende Kurz-Url zum Verlinken dieser Seite:
https://www.dagstuhl.de/15342
Organisatoren
- Dieter Kranzlmüller (LMU München, DE)
- Barry L. Rountree (LLNL - Livermore, US)
Kontakt
- Susanne Bach-Bernhard (für administrative Fragen)
Impacts
- A Power-Aware Cost Model for HPC Procurement : article in 2016 IEEE International Parallel and Distributed Processing Symposium Workshops : pp. 1110-1113 - Gholkar, Neha; Mueller, Frank; Rountree, Barry - Los Alamitos : IEEE, 2016 - (International Parallel and Distributed Processing Symposium Workshops : article).
- Dagstuhl Manifesto : Power-Bounded HPC Performance Optimization : pp. 171-174 : article - Rountree, Barry; Kranzlmüller, Dieter - Berlin : Springer, 2016 - (Informatik Spektrum : 39. 2016, 2 : article).
- Power Tuning for HPC Jobs under Manufacturing Variations - Gholkar, Neha; Mueller, Frank; Rountree, Barry - Raleigh : University, 2016. - 11 pp - (Technical Report, Dept. of Computer Science, North Carolina State University ; 2016-2).
- Power Tuning HPC Jobs on Power-Constrained Systems : article in Proceedings of the 2016 International Conference on Parallel Architectures and Compilation, PACT '16 : pp. 179-191 - Gholkar, Neha; Mueller, Frank; Rountree, Barry - New York : ACM, 2016 - (International Conference on Parallel Architectures and Compilation, PACT '16 : article).
With the goal of limiting its first exascale supercomputer to 20 MegaWatts of electricity, the US Department of Energy has moved the issue of power-limited high-performance computing to the front of its research priorities. But even before the design phase of exascale computing has begun, specifications for new petascale systems (such as CORAL and Tsubame 3) include stringent power bounds. Constructing these systems is, by now, only a moderate engineering challenge. However, writing and optimizing code that will make the best use of power-limited resources poses both scientific and engineering challenges that are only now beginning to be addressed. While the community has sufficient experience to handle porting and optimizing codes for larger numbers of processors and faster networks, optimizing for power requires expertise to be applied from cluster construction and site electrical contracts through job schedulers, load balancing algorithms, dynamic processor configuration and even the choices of algorithms and compiler optimizations.
This Dagstuhl Perspective Workshop will provide a novel setting to advance exascale HPC research. It provides a unique setting to help stakeholders in the HPC community to come to grips with the problem of power-limited optimization. For the first two days, our international panel of experts in facilities, schedulers, runtime systems, operating systems, processor architectures and applications will provide detailed perspectives on open problems in each of their fields that must be resolved in order to reach a useful power-constrained exaflop system. The second half of the workshop will consist of moderated discussion of these issues with the goal of understanding how potential local solutions affect the rest of the software stack.
One of the primary benefits of this workshop will be to align future research with a common set of goals and solutions. We expect it to influence how research is done on power-bounded HPC both in academia and national laboratories. Among others we expect the following results:
- A Strawman architecture of a power-limited supercomputing system as a basis for future work
- A journal paper detailing the Strawman architecture and its rationale
- A Dagstuhl Manifesto describing future funding requirements necessary to meet the goals
- An international collaboration plan across the exascale initiatives in Europe, Japan and the US
This Dagstuhl Perspectives Workshop is unusual in that we will be focusing on the “unknown”, rather than on the “state of the art”. While discussion is a vital component of any Dagstuhl event, here we will be taking the focus away from the presentations and placing it squarely on the interactions among the participants. We expect to forge several new collaborations across vendors, national laboratories and academia. And we would like to invite you to join us.
The Dagstuhl Perspectives Workshop 15342 "Power-Bounded HPC Performance Optimization" has been an interesting experience, as in contrast to other workshops, we focused on the unknown characteristics of future exascale systems rather than on the state-of-the-art of todays petascale architectures. In order to do this, a large fraction of the workshop was spent on in-depth discussions in three working groups, while plenary sessions served to provide impulses on specific topics and to synthesize the findings of the breakout sessions. The key ingredient of this workshop has been the interaction between the participants, leading to several new collaborations across vendors, national laboratories and academia.
The key findings of the workshop can be identified as follows:
- Power-bound performance optimization has different objectives according to the respective targets and operational goals. While infrastructure providers are often bound to a specific spending, users want to utilize a resource at the maximum of its capabilities. As a result, any power-bound optimization must address multiple criteria, and the solution is rarely straight-forward but specific for a given setting.
- The currently available information on each layer of the computing environment is insufficient. Both, the availability of information with respect to its power characteristics, as well as the exchange between different layers, needs to be improved in order to optimize the operation of infrastructures and the execution of applications on a given system.
- Due to the number of dependencies, any optimization needs to find a good balance between "user happiness", total costs, and performance. These characteristics are important for both, providers and users, and a careful balancing strategy needs to be implemented without harming any interests of the actors too much.
The discussions at the Dagstuhl Perspectives Workshop have lead to the identification of a number of technical problems, which need to be addressed in the near future before achieving optimal results in a power-bound environment. As a conclusion, the participants agreed that a strategic and tactical agenda is needed, which identifies the individual problems and technologies as well as their interconnections, such that future systems can utilize this knowledge for new approaches of power-bound HPC performance optimization. The results of this investigations should be made available as a white book, which describes the strategy for future exascale systems.
Creative Commons BY 3.0 Unported license
Dieter Kranzlmüller and Barry L. Rountree
- Michael Bader (TU München, DE) [dblp]
- Natalie Bates (Lawrence Livermore National Laboratory, US) [dblp]
- Pete Beckman (Argonne National Laboratory, US) [dblp]
- Jonathan Eastep (Intel - Hillsboro, US) [dblp]
- Neha Gholkar (North Carolina State University - Raleigh, US)
- Joseph Greathouse (AMD - Austin, US) [dblp]
- Thomas Ilsche (TU Dresden, DE) [dblp]
- Dieter Kranzlmüller (LMU München, DE) [dblp]
- Stephanie Labasan (University of Oregon - Eugene, US) [dblp]
- David K. Lowenthal (University of Arizona - Tucson, US) [dblp]
- Matthias Maiterth (LMU München, DE) [dblp]
- Andres Marquez (Pacific Northwest National Lab. - Richland, US) [dblp]
- Yousri Mhedheb (KIT - Karlsruher Institut für Technologie, DE) [dblp]
- Shirley V. Moore (University of Texas - El Paso, US) [dblp]
- Frank Mueller (North Carolina State University - Raleigh, US) [dblp]
- Andreas Raabe (DFG - Bonn, DE) [dblp]
- David Richards (LLNL - Livermore, US) [dblp]
- Suzanne Rivoire (Sonoma State University - Rohnert Park, US) [dblp]
- Barry L. Rountree (LLNL - Livermore, US) [dblp]
- Martin Schulz (LLNL - Livermore, US) [dblp]
- Kathleen Shoga (LLNL - Livermore, US) [dblp]
- Torsten Wilde (LRZ - München, DE) [dblp]
Klassifikation
- hardware
- modelling / simulation
- operating systems
Schlagworte
- high-performance computing
- supercomputing
- power
- performance optimization
- physics simulation
