[工程科技]DDM.ppt

INTRUDUCTION Use of Finite Element Analysis (cont.): Computers are getting faster Models are getting larger Pre and Post Processing: Need for user friendly input of data Need for user friendly interpretation of results USER INTERFACES Use of Finite Element Analysis (cont.): Increasing speed of computers Larger models are used because of: accuracy: more elements give higher accuracy completeness: simulating complete processes without too many approximations ease of pre processing: automatic generation of finite element meshes with a large number of elements Usage of multiple CPU’s: Solve big problems faster Be able to solve huge problems Parallel processing DDM – Domain Decomposition Method Based upon MPI – Message Passing Interface Why? Speed With more processors the job usually runs faster Memory usage If job is too large for a single machine – use multiple machines on the cluster Parallel processing Supported systems Shared memory machines Single machine with multiple CPUs, all access the same memory. Clusters of workstations Separate workstations over Ethernet Dedicated cluster with high speed network Supported networks for clusters Ethernet (standard, fast) Linux: Gigabit, Myrinet, SCI, Infiniband IBM SP HP Itanium: Gigabit, Hyperfabric Parallel processing MPI versions used MPICH on most Unix systems All Unix systems can use vendor provided MPI (recommended) ScaMPI or MPICH on Linux MP-MPICH from University of Aachen on Windows Parallel processing - Domains Model is subdivided into domains of elements Each element is in one domain only Nodes on domain boundaries are duplicated in the sharing domains – so-called interdomain nodes Each CPU solves the part for one domain Split total model into a number of sub-models (domains) There are three ways of solving the equations in MSC.Marc Partial Conjugate Gradient method, for Direct Solvers – solver 8 Full Conjugate Gradient method, for Iterative Solvers – solver 2 Fully direct method, for Hardware prov