Parallel graph generation algorithms for shared and distributed memory machines

In this paper we give an overview and a comparison of two parallel algorithms for the state space generation in stochastic modeling on common classes of multiprocessors. In this context state space generation simply means constructing a graph, which usually gets extremely large. On shared memory machines, the key problem for a parallelization is the implementation of a shared data structure which enables efficient concurrent access for retrieving the nodes of the graph. In our realization this search structure is a B-tree with special locking mechanisms assigned, leading to significant speedups. For distributed memory machines a dynamically adaptive partitioning strategy to distribute the state space onto different processors together with load balancing mechanisms is implemented. Thus sequentially not manageable problem sizes can be solved by combining the main memories of clustered workstations.