A Data Prefetching Model for Desktop Grids and the Condor Use Case

Data-aware scheduling improves data-intensive applications performance in distributed systems. Desktop Grids have been successfully used for solving scientific applications at low cost. However, since the resources are accessed through wide area networks, the bottleneck comes with the bandwidth limitation and the master-worker paradigm. Data Prefetching has the potential for overlapping the elapsed times to exchange data between nodes. In this work, we propose a decentralized approach for data prefetching for Bag-Of-Tasks and DAG applications. Experimentation, using more than 200 machines on Nancy sites of the Grid'5000 testbed, demonstrate that our model improves the performance of applications execution.