Porting Tissue-Scale Cardiac Simulations to the Knights Landing Platform

To study the performance difference between the two generations of Xeon Phi, as well as the respective programming techniques, we port and optimize a simulation code for 3D tissues of the human cardiac ventricle to the new Knights Landing (KNL) platform. The amount of computation arises from a large number of cardiac cells and a physiologically realistic model adopted for each cell, which is resolved as having \(10^4\) calcium release units and controlled by \(10^6\) stochastically changing ryanodine receptors and \(1.5 \times 10^5\) L-type calcium channels. The programming challenge arises from the fact that the involved computational tasks have various levels of arithmetic intensity and control complexity, requiring in some cases hardware-specific manual optimizations. We also study how the new memory system of KNL can be properly used to allow larger simulations beyond the capacity of the 16 GB MCDRAM. The combined advancements in hardware and software result in an almost ninefold increase in performance on the KNL over the previous generation.