Filesystem performance testing for parallel clusters

Shared filesystem access is the main bottleneck for parallel clusters. This paper presents some of the tests we performed while optimizing different components contributing to the filesystem access speed. Interesting results were obtained while studying how NFS access scales with the number of cluster nodes doing read/write operations at the same time. The scaling curve can be used to optimize the number of nodes and the networking of a cluster depending on the desired NFS I/O needs of the typical job. Another area of interest is live growing of filesystems on RAID arrays by adding more disks, and its effect on the filesystem performance. Aligning the RAID stripe with the filesystem block proved to be important for the access speed to the files. While using more disks on a RAID array increases the access speed, an empty filesystem was as quick as a full one in our tests. The results of the experiments can be used to optimize the performance of the parallel clusters both at the initial design stage and during the fine tuning process after the cluster is running.