|Authors||X. Cai and J. Langguth|
|Title||Accelerated high-performance computing for computational cardiac electrophysiology|
|Project(s)||Meeting Exascale Computing with Source-to-Source Compilers|
|Publication Type||Talk, keynote|
|Year of Publication||2017|
|Location of Talk||The University of Tokyo, Tokyo, Japan|
Massively parallel hardware accelerators, such as GPUs, are nowadays prevalent in the HPC hardware landscape. While having tremendous computing power, these accelerators also bring programming challenges. Often, a different programming standard applies for the accelerators than that for the conventional CPUs. For computing clusters that consist of both accelerators and CPUs, where the latter are hosts of the accelerators, elaborate hybrid parallel programming is needed to ensure an efficient use of the heterogeneous hardware.
This talk aims to share some experiences of implementing computational science software for heterogeneous computing platforms. We look at two scenarios: CPU+GPU  and CPU+Xeon Phi  heterogeneous computing. Common for both scenarios is the necessity of a proper pipelining of the involved computational and communication tasks, such that the overhead of various data movements can be reduced or completely masked. Moreover, suitable multi-threading with thread divergence is needed on the CPU host side. This is for enforcing computation-communication overlap, coordinating the accelerators, and allowing the CPU hosts to also contribute with their computing power. We have successfully applied hybrid CPU+Knights Corner co-processor computing  to two topics of computational cardiac electrophysiology, making use of the Tianhe-2 supercomputer. Results  about using the new Xeon Phi Knights Landing processor will also be presented.
. J. Langguth, M. Sourouri, G. T. Lines, S. B. Baden, and X. Cai. Scalable heterogeneous CPU-GPU computations for unstructured tetrahedral meshes. IEEE Micro, 35(4):6–15, 2015.
2nd International Symposium on Research and Education of Computational Science