Modeling of Exascale Applications on Exascale Systems Summit (Washington, D.C.)
March 6, 2012 from 8 a.m. to 5:30 p.m.
American Geophysical Union (AGU), 2000 Florida Ave, N.W. Washington, D.C. 20009
Landscape and motivation:
Exascale computing will be increasingly complex due to the evolving sophistication of the science problems being computed, the exponential growth in concurrency, and the need for achieving reliable high performance under severe power constraints. Specifically, Exascale systems will exhibit billion-way parallelism, deep and narrow memory hierarchies, heterogeneity in compute and storage, near threshold operating voltages, and complex node level architectures with a lot of "dark silicon" (not everything can be turned on all the time to stay in energy envelope). Moreover, Exascale computing may require new programming models, system software, or even novel execution models to address the challenges of the performance/power/reliability triad.
Exascale applications will increasingly include coupled physics and/or sophisticated data analysis exhibiting non-local patterns of data exchange and synchronization. Such applications already utilize dynamic and "unstructured" data structures. These applications will have to continue their trend towards higher parallelism but lower locality across a broad spectrum of science and even business applications. These trends taken with the architecture trends imply the need to radically evolve in order to operate efficiently under severe energy constraints, and exploit vastly more parallelism without relying on the same order of magnitude increase in the available memory. Applications will need to exploit multiple degrees of heterogeneity, map efficiently on increasingly deeper narrower memory hierarchies, all while minimizing data movement, the main source of energy inefficiency. As a consequence, the workload characterization of applications that will run on Exascale machines is expected to be significantly more complex from that of applications that run on current machines.
Exascale platforms in turn will display execution behaviors that use asynchrony and irregular structures to overcome the increasingly heterogeneous nature of the applications and platforms themselves and will have to dynamically adapt to a number of hardware, system software, and application events, continuously optimizing energy and performance in the presence of very high error and failure rates. As a result, increasingly more tradeoffs will have to be negotiated.
Performance modeling will be needed across the spectrum of scales from helping to evaluate design tradeoffs on machines 5 to 10 years out down to helping smart self-aware applications and runtimes to make the right resource allocation decisions on-the-fly. Future Exascale applications on future Exascale platforms will require extensive design space exploration as well as revolutionary modeling and analysis techniques. Moreover, modeling techniques need to encompass a broader space, including execution and programming models, and an increasing number of boundaries among the architecture/hardware, system software, execution models, and algorithms and applications.
This Summit will explore the challenges and promising alternatives associated with the performance modeling of future Exascale applications on future Exascale platforms. A high-level goal of the summit is to brainstorm out-of-the-box modeling alternatives and techniques for Exascale. Details of the research work developed by participants are out of scope for this meeting. We anticipate technically provocative statements that instigate a lot of discussion, new thinking and revolutionary new approaches to the problem. At the end of the summit, we need to have a very good understanding of the sub-problems that should be tackled, if any. We also need to identify issues associated with performance modeling in the Exascale era, clearly articulating challenges. Revolutionary modeling approaches for Exascale need to be identified, together with objective criteria to assess them. The final goal is to outline a road map for this research area in the path to Exascale.
Summit Goals Summary:
1. Brainstorm out-of-the-box modeling alternatives and techniques for Exascale.
2. If applicable, decompose the modeling problem into sub-problems that should be tackled.
3. Identify issues associated with this problem, clearly articulating and prioritizing associated challenges.
4. Identify alternative Exascale modeling approaches and objective criteria to assess them.
5. Outline a roadmap for this research area of Exascale computing.
The public is welcome to participate in this informal conversation about the Modeling of Exascale Applications. However, space at the AGU is limited and on-site registration will be used to accommodate as many participants as possible. To register in advance, please email firstname.lastname@example.org