We present a novel DEM (discrete element method) code with explicit time stepping. DEM codes simulateudbillions of small particles that interact with each other primarily through collisions. Different to state-of-the-artudcodes, we rely on triangulated non-spherical particles. This is computationally demanding, and we thus deviseudan asynchronous data exchange communication technique using MPI (Message Passing Interface) on manycoreudsupercomputers, we discuss possible solutions to handle ghost particles that overlap multiple subdomains. Atudthe compute node level, shared memory parallelism as well as vectorised SIMD executions are studied. On theudalgorithmic side, we explore a hybrid parallelisation approach and memory layouts that are suited to combineudrobust exact geometry checks with a fast penalty-based method.
展开▼
