Research On Parallel Computation Of Shock - Contact Problem

Contact-impact problems widely exist in engineering and real lives. simulation using computer is an efficient way to research those problems. Contact-impact problems have been considered not to be well solved in a few days or months using serial computer because of the strong nonlinearity of this problem and complexity of the contact relation processing. Thus, much attention has recently focused on parallel computing of contact-impact problems to save wall clock time as the increasing requirements of simulation accuracy. It’s worth to design parallel contact algorithms considering the specific architecture of parallel computers in order to make full use of these hardware resources.The explicit finite element computing of contact-impact was parallelized by OpenMP using performance profile based method on shared memory parallel computers. Memory access locality oriented blocked FEM algorithm was used to efficiently use hierarchical memory system. Based on this algorithm, not only the hit ratio of cache was improved, but also the granularity of OpenMP parallel task was coarsened. The parallel program with OpenMP has fairly good parallel efficiency in test suites.Domain decomposition of contact force computing was achieved by parallel auto-contact algorithm which based on bucket sorting. This algorithm was used to parallel global searching of contact interface, and contact nodes with segments relationship construction, and communication topology building. Auto contact force was computed parallel based on this algorithm.Dual domain decomposition strategy was presented according to the peculiarity of explicit FEM that using to solve the contact problems. In the internal force computing phase, graph based static domain decomposition was used because of the computing date dependency relationship is based on element connect topology of mesh. In the contact force computing phase, on the other hand, because of the contact relationship was determined by spatial coordination of mesh nodes and segments, the spatial decomposition methods must be used. Re-decomposition of contact force computing domain must be considered as the contact relations between nodes and segments were changed dynamically along the time steps. The MPI implementation of dual domain decomposition and the numerical experiments show that this method has good parallel speedup on distributed memory computers. Finally, the MPI+OpenMP Hybrid program method was considered in combine with CLUMPS(CLUster of MultiProcessorS) architecture which shared memory inside nodes and distributed memory between nodes. MPI+OpenMP was used to parallelized the contact problem that utilizes MPI for distributed memory parallelism and OpenMP for shared memory parallelism. The hybrid programming model can make full use of the processors that shared memory in a nodes, and costs of MPI communication between nodes did not increased. The test show that this hybrid program can solve10million DOF three dimension contact problems using hundreds processors, and save the wall clock time at least one order of magnitude.