A Dynamic Task Graph Parallelization Approach

The general purpose Finite Element system PERMAS [1] has been parallelized by a dynamic task graph approach capable to support shared and distributed parallel computer ar-chitectures as well as workstation clusters. The method used to parallelize this large application software package is of high generality and has the capability to parallelize all mathematical operations in a FEM analysis – not only the solver. Utilizing the existing hyper-matrix data structure for large, sparsely populated matrices, a programming tool called PTM was introduced that automatically parallelizes block matrix operations on-the-fly. The programming model totally hides parallelization from higher order algorithms, thus giving the physically oriented expert a virtually sequential programming environment. This is a key issue for an automatic paralleliza-tion as well as for the re-use of existing software. The approach is based on dynamic scheduling concepts for an operation graph of sub-matrix operations that is asynchronously build and executed. A clustering algorithm distributes the work, performing a dynamic load balancing and exploiting data locality by data distribution strategies applied in the case of distributed memory systems. Furthermore a distributed data management system allows free data access from each node. The generality of the approach is demonstrated by some benchmark examples dealing with different types of FEM analyses.