Home > Workload Solutions > High Performance Computing > White Papers > Enhance Availability of Storage Services with an NFS Storage System > Input/Output performance
This section presents the results of the I/O performance tests for the Dell Technologies Validated Design for NFS Storage system. All performance tests were conducted in a failure-free scenario to measure the maximum capability of the solution. These tests focused on three types of I/O patterns: large sequential reads and writes, small random reads and writes, and three metadata operations (file create, stat, and remove). Like the previous version of the solution, this Design for NFS Storage uses deadline I/O scheduler, and 256 NFS daemons.
An XFS file system with 569TiB of usable space (840TB raw storage size) is configured and exported via NFS 4 and benchmarked with IPoIB connectivity over HDR100, using datagrams and MTU of 4096 for clients and servers (via Subnet Manager on the Mellanox QM8700 switch). A 16-node compute cluster is used to generate workload for the benchmarking tests. Each test is run over a range of threads on the client nodes to test the scalability of the solution as the number of users increased.
Since the cluster has only 16 nodes, the 32 and 64-thread data points are obtained with each of the 16 clients running 2 and 4 threads respectively to obtain sustained performance values. Client caching is prevented by each client node unmounting the NFS file system and dropping all the operating system caches before remounting NFS between every operation such as, sequential writes and reads or dropping all OS caches on clients and servers when all files are removed (metadata tests).
The IOzone and MDtest benchmarks are used in this study. IOzone is used for the sequential and random tests. For large blocks sequential tests, a request size of 1024KiB was used on large files. The total amount of data transferred is 512 GiB divided among the client threads used. Random tests used a 4KiB request size and each client thread, reading and then writing randomly 4GiB on a pre-created file after dropping caches as described, and the server-available memory for cache is restricted to obtain sustained performance without using 2x server RAM. Metadata tests are performed using the mdtest with OpenMPI and included file create, stat, and remove operations.