Scaling NFS through RDMA for cluster computing

Parallel applications combine the power of a large number of commodity processors to solve a single problem, distributing their global data structures throughout the memory of all participating clustered computers. As such, they can benefit greatly from the ability to share data using a single file system that can provide a common view to the filesystem's namespace. Due to its ease of implementation, ubiquitous interface in nearly all operating systems, and low cost, many cluster platforms adopted the Network File System (NFS) as a common filesystem. Unfortunately, NFS suffers from specific performance bottlenecks, limiting it's suitability in large scale cluster computing. Combining NFS with emerging network technologies such as Remote Direct Memory Access (RDMA) may improve NFS performance to meet the expectations of high performance computing (HPC). We will benchmark and analyze NFS over RDMA using an I/O intensive combustion code developed at Sandia.