The Introduction of Multi-level Parallelism Solvers in Multibody Dynamics

The paper deals with the problem of the scaling speed-up of time-consuming computations on the example of the dynamic stress-strain state (DSS) modelling for the multibody systems. Speed-up is achieved through the introduction of multilevel parallelism at the levels of decomposition of the model and the distribution of calculations on the computational nodes of the cluster, at the level of multithreaded computations within the part of the model at one node and at the level of vectorization of calculations and their transfer to GPU/ FPGA accelerators within a single computational flow. Vectorization is considered for both conventional multicore CPUs and the MIC architecture. Accelerators are used using OpenCL technology for heterogeneous systems. The possibility of computations speed-up up to 50 times for the considered problem due to the introduction of multilevel parallelism is shown.