Thomas Sterling (Center for Research in Extreme-scale Computing, Indiana University)
Computational Physics at Extreme Scale
The future of high performance computing is jeopardized by the end of Dennard scaling and the approaching end of Moore’s Law, eliminating the two principal sources of exponential performance gain that has dominated the field for more than two decades. Future science applications demand increased fidelity, physics phenomenology, multi-scale formulation, efficiency, and scalability. It is further recognized that these must be achieved within a context that makes system use easier and therefore more productive than even today’s systems while delivering two to three orders of magnitude performance advantage. The ParalleX execution model suggests an alternative paradigm for the representation and management of extreme scale physics applications that dramatically improves efficiency, scalability, and productivity through the exploitation of dynamic adaptive methods within the context of global address space. The HPX runtime system is an experimental resource management and task scheduling system which, when combined with the low level programming interface, XPI, offers an alternative strategy that promises new opportunities in science exploration and discovery through computation. ParalleX integrates advanced concepts including lightweight multi-threading, message-driven computation, multi-nodal processes, dynamic introspection, and heterogeneous resource control into a single computational framework to address the challenges of starvation, latency, overhead, and contention. Additional emerging capabilities provide promising paths to addressing the challenges of power consumption and reliability. By offloading much of the resource management decision from the user to the runtime system, ParalleX offers a means of simplifying methods for programming. This presentation will discuss the ParalleX model and its proof-of- concept implementations of HPX and XPI. A number of physics codes under development will be examined using these advanced concepts including wavelets and finite element codes for shockwave physics with reactive materials, adaptive mesh refinement (AMR) for multimaterial modeling, Fast Multipole Method codes for molecular dynamics, cosmology simulations, and computational electromagnetics, Barnes-Hut N-body codes for cosmological structure simulations and galaxy evolution, and LULESH for shock hydrodynamics.