[00748] New probabilistic algorithms for scientific supercomputing
Session Time & Room : 5D (Aug.25, 15:30-17:10) @E705
Type : Contributed Talk
Abstract : Sustained strong scalability is hard to sustain beyond 10K processors due to the communication and synchronisation involved in domain decomposition for PDEs. Seeking to overcome them, Spigler and Acebrón introduced probabilistic domain decomposition, which inserts stochastic calculus in the formulation---however with slow error convergence. I will present a hybrid idea, SpAc, which retains most of the scope for embarrassing parallelism, while being orders of magnitude faster. Proof of concept supercomputing simulations will be discussed.