Generating an Efficient Dynamics Multicast Tree under Grid Environment

The use of an efficient multicast tree can substantially speed up many communication-intensive MPI applications. This is even more crucial for Grid environment since MPI runtime has to work on wide area network with very different and unbalanced network bandwidth. This paper proposes a new and efficient algorithm called, GADT (Genetics Algorithm based Dynamics Tree) that can be used to generate an efficient multicast tree under Grid environment. The algorithm takes into consideration the unbalanced network speed of Grid system in order to generate a multicast tree. The experiments are conducted to compare GADT with binomial tree and optimal algorithm. The results show that GADT can produce a multicast tree that has communication performance close to the optimal multicast tree deriving from exhaustive search. Moreover, the multicast tree generated results in a substantially faster communication than a traditional binomial tree algorithm. Therefore, GADT can be used to speed up the collective operation for MPI runtime system under Grid environment.