The iSCSI protocol communicates over existing IP infrastructure, making it an intriguing storage alternative for cost-conscious HPC users. Two approaches have emerged for using iSCSI in an HPC context. With the first approach, I/O servers — also called data movers — access iSCSI storage at the block level and export it to the cluster nodes via a distributed file system. With the second approach, cluster nodes connect directly to the storage at the block-level via local iSCSI initiators. This paper describes both architectures and compares their performance and scalability across several workloads at increasing cluster sizes. Our results show that cluster and workload characteristics dictate the suitable approach for each cluster. We conclude with design recommendations based on our measured performance results.
展开▼