Efficient Execution on Long-Distance Geographically Distributed Dedicated Clusters

摘要

Joining, through Internet, geographically distributed dedicated heterogeneous clusters of workstations can be an inexpensive approach to achieve data-intensive computation. This paper describes a master/worker based system architecture and the strategies used to obtain effective collaboration in such a collection of clusters. Based on this architecture an analytical model was built to predict and tune applications' execution performance and behaviour over time This architecture and model were employed for the matrix multiplication algorithm over two heterogeneous clusters, one in Brazil and the other in Spain. Our approach results show that the model reaches 94% prediction, achieving 91% of the clusters' total performance.