Design and implementation issues of high performance cluster computing

Abstract

The emerging Grid Computing paradigm enables researcher, scientists and engineers to build so called Hypercluster, or cluster of clusters, from the available resources of the Grid. However, Hypercluster built from the resources of the Grid inherits the unpredictable resource behaviors, which result in variable performance, even when using the same set of the resources but at different times. The unpredictable resource behaviors also violate the assumptions like, homogenous processors and communication delay in traditional parallel environment. The parallel computing models in the past are no longer reflecting the circumstances that we have in Grid Computing today. In this thesis, a design of a Reservation Aware Operating System (RAOS) and a parameterized model, Hypercluster on Grid (HOG) are proposed to tackle the problems. RAOS is able to stabilize the performance of applications running on the unpredictable Grid environment. We show this by stabilizing the communication cost (connection stup time, message packing time) and the computation time, which are the important parameters required to optimize parallel applications and distributed applications, under the loaded computation nodes in cluster environment. HOG is a parameterized model, which is abstract enough to reflect the heterogeneous Grid environment, regardless of the underlying implementation technologies. It benefits the development of parallel algorithms in Hypercluster environment. The major contribution of this thesis is its advancement of the understanding of tools and methodology that can be used in resource reservation, resource sharing and also modeling parallel programs in the heterogeneous Hypercluster environment.