Computational Foundations for Exascale Simulations of Supernovae

As machine architectures advance beyond the petascale to the exascale, the numerical methods and their implementations in current scientific applications codes will have to be reexamined and likely changed considerably. In addition, given an effective set of numerical methods and implementations for such architectures, the simulations thus enabled will produce data that will challenge, or even break, the hardware and/or software infrastructure currently in place for simulation data analytics (collective parallel I/O, data analysis, visualization, and workflow execution and management). Therefore, the promise of extreme scale science will likely not be realized without a concerted effort, beginning now, to lay the foundations for the effective use of these architectures as they become available over the next decade. We have identified issues specific to our challenge, but relevant to other application areas: (1) The development of algorithms for the solution of (i) the distributed peta- to exa-scale sparse linear systems that arise from our neutrino (radiation) transport simulations, (ii) the elliptic equations describing the gravitational field in nonspherical stellar cores, and (iii) the complicated, multi-wave hyperbolic magnetohydrodynamics equations governing the evolution of the stellar core magnetic fields, that will need to make efficient use of multicore sockets and scale across tens to hundreds of thousands of sockets. (2) Optimization of both single-processor and parallel code performance, including the development of load balancing strategies for adaptive mesh refinement and the distribution of thermonuclear kinetics calculations across entire multi-core (or heterogeneous processor) nodes. (3) The development of strategies and software for data analytics (specifically, regarding what we propose here, collective parallel I/O). (4) The establishment of formalized and automated verification of software. Our approach will be staged. Developments specific to both the CHIMERA code for petascale and sustained petascale supernova simulation and the GenASiS code for exascale supernova simulation are planned.