Running Large-Scale Ultrasound Simulations on Intel Xeon Phi (KNC) accelerators

Large-scale ultrasound simulations has excited high interest in medical applications such as HIFU treatment or Photoacoustic imaging. However, routine execution of such simulations using traditional computational methods such as finite difference methods is considered intractable due to the size of required data sets. The k-Wave (k-space pseudo-spectral based ultrasound wave propagation solver) toolbox alleviates the computational and memory requirements by significantly reducing the spatial and temporal grid resolution, however, for the cost of a large amount of global all-to-all communication. To improve data locality, reduce data movements and allow an efficient use of accelerators, we introduce a local domain decomposition using local Fourier basis. In this article we investigate behavior of k-Wave in local domain decomposition variant running on the Salomon cluster using the Intel Xeon Phi (KNC) accelerators which constitute about 50\% of cluster's 2 Pflop/s R-Peak performance.