Design And Implementation Of Parallel Graph-partitioning And Hypergraph-partitioning Methods For OpenFOAM

OpenFOAM is a widely used open source CFD software. When we use OpenFOAMtocomputealarge-scaleCFDapplicationinparallel, itisnecessarytodecomposethegen-erated mesh into several subdomains and assign them to the target processors for parallelexecution. Aclassicalapproachtosolvethisproblemisusinggraph-partitioningmethods.Graph-partitioning methods available in OpenFOAM are all serial, which have two maindisadvantages in application: the serial partitioning procedure is a bottleneck of the appli-cation execution because of its low partitioning efficiency; these methods are not capableof adaptively re-decompose the mesh when information of the mesh changes. In addition,descriptions of the communication volume of common graph-partitioning methods arenot accurate which weakens their performance. On the basis of the research background,this paper researches into the following four aspects:Firstly we analyzed execution efficiencies, partitioning qualities and capabilities tobeparallelofseveralclassicalgraph-partitioningalgorithmsandconcludedtheadvantagesof using multilevel partitioning algorithm to solve the mesh decomposition problems inOpenFOAM. This work formed the theoretic foundation of our research.Secondlywedesignedaparallelgraph-partitioningmethodinOpenFOAM-FPMetis.The generated mesh was described as a graph and then partitioned by parallel coarsen-ing, initial partitioning and parallel uncoarsening. The partitioning procedure aimed atminimizing the total edge-cut while maintaining the load balance constraint. FPMetis im-proved the partitioning efficiency compared with serial methods and was able to performadaptively dynamic partitioning. Considering that the traditional global match algorith-m used in the parallel coarsening procedure introduced a huge amount of communicationoverhead,webroughtforwardanovelmatchschemenamedFHEMwhichcombinedglob-al match with local match to avoid introducing the unnecessary communication overhead.The complexity analysis of FPMetis showed that the new match scheme can reduce thecommunication overhead.Thirdly, we designed a hypergraph-partitioning method in OpenFOAM-FHMetis.Hypergraph is more applicable and could describe communication volume more accu-rately than graph. The generated mesh was described as a hypergraph. The partitioningprocedure included hypergraph coarsening, initial partitioning and uncoarsening, aiming to minimizing the total hyperedge-cut while maintaining the load balance constraint. Thepartitioning quality can be improved by using hypergraph-partitioning to replace graph-partitioning.The above two partitioning methods were implemented in OpenFOAM. Firstly, thestoring structures of distributed graph and hypergraph and the transition algorithms ofgenerated mesh into these structures were designed. The whole implement contained thecore partitioning algorithms of FPMetis and FHMetis, parallel mesh decomposition inter-face, mesh decomposition interface based on hypergraph and task-decomposition moduleofOpenFOAM.AnevaluationmodelofmeshdecompositionmethodswerebuiltinOpen-FOAM including execution time, load balance and communication volume. The validityand efficiency of the presented method in this paper were analyzed and testified.