Improved Data-Aware Task Dispatching for Batch Queuing Systems

Abstract

This paper describes a data-aware task dispatching strategy called Improved Data-Aware Task Dispatching (IDAD). This approach exploits the high-performance of local file access in non-uniform storage-access (NUSA) file systems and is based on our previous work, Data-Aware Dispatch (DAD). In IDAD, the method of calculating data placement is revised, and the CPU factor is removed, as it has no major impact on performance but significantly reduces the difficulty for tweaking parameter.We evaluated our approach in comparison with DAD and the stock FIFO Torque scheduler using BLAST benchmarks. We observed makespan reductions of 10.40% and 35.05% compared with DAD and stock FIFO schedulers, respectively.