B9. Develop a “cosmic simulator” capability that integrates increasingly complex astrophysical measurements with simulations of the growth and evolution of structure in the universe, linking the known laws of microphysics to the macro world. Develop large-scale, special-purpose computing devices and innovative algorithm development to achieve this goal.
1:00-1:30 Mike Norman (UCSD)
1:30-2:00 Paul Ricker (UIUC)
2:00-2:30 Tom Quinn (UWash)
2:30-3:00 Paul Woodward (UMinn)
3:30-4:00 Alexei Khokhlov (U of Chicago)
4:00-4:30 Pablo Laguna (Penn State)
4:30-4:45 Bronson Messer/Christian Cardall (ORNL)
4:45 - Free-for-all (Open Discussion)
Questions to be addressed by the breakout session:
- What compelling scientific problem could you address with an exascale computer?
- What external factors make the solution to this exascale problem a pressing concern? E.g., future satellite missions, etc.
- Do you have an extant code you believe might have a path forward to the exascale? I.e., regardless of how the implementation might change, do you now have a coded version of the algorithm you anticipate using to solve the problem in (1)? Or do you believe a new algorithm should/must be used in the future?
- Is there potentially important physics you neglect today for want of computational power that would be feasible to model on an exaflops computer? Why is this physics important in your problem?
- Given what you know about near-future computer architectures, do you see an immediate problem? E.g., will CPUs with 10's-100's of cores but only modest increases in memory bandwidth present a problem?
- Given how you program for today's biggest architectures, do you see an immediate problem for running on machines with 10^6-10^7 cores? I.e., can you clearly see the end of scaling right now?
- Given how much time and energy you expend on the management and analysis of data from simulations now, do you foresee obvious problems when the compute rate and memory available for simulation increases 10,000x?
- Do you anticipate using observational data in/integrating observational data into your simulation or simulation-analysis pipeline? How? Does that require an exaflops computer?