Further Exploit the Potential of I/O Forwarding by Employing File Striping

摘要

I/O forwarding has become a standard I/O layer in today's top HPC systems to alleviate the mismatch of computing power and I/O bandwidth. Each I/O node is exclusively utilized by a small group of compute nodes to forward their I/O requests to storage servers. Since I/O workloads on production HPC systems are bursty both in time and space, the load on I/O nodes are highly unbalanced. Rank 0 I/O pattern that commonly exists in scientific applications aggravates the load imbalance as well. A small number of I/O nodes can easily become hot spots, which hinders the potential of I/O forwarding layer to be fully realized. In addition, the lack of coordination between I/O forwarding layer and storage system layer causes serious I/O interference and impacts the performance significantly. In this paper, we propose to employ file striping on I/O forwarding layer to apportion the heavy workloads to multiple idle I/O nodes. We also coordinate the file striping on I/O forwarding layer and storage system layer to minimize the I/O interference on storage servers. Our ideas are implemented and evaluated on an open-source I/O forwarding software IOFSL, and the results show that the I/O performance of applications can be greatly accelerated by our approach.