Multi-level parallelization with parallel computational services in BeesyCluster

The paper presents a concept, implementation and real examples of dynamic parallelization of computations using services derived from MPI applications deployed in the BeesyCluster environment. The load balancing algorithm invokes distributed services to solve subproblems of the original problem. Services may be installed on various clusters or servers by their providers and made available through the BeesyCluster middleware. It is possible to search for services and select them dynamically during parallelization to match the desired function the service should perform with descriptions of services. Dynamic discovery of services is useful when providers publish new services. Costs of services may be incorporated into the selection decision. A real example of integration of a given function using distributed services has been implemented, run on several different clusters without or with external load and optimized to hide communication latency.