NFS-cc: tuning NFS for concurrent read sharing

A common file access pattern found in cluster applications is concurrent read sharing: applications running on multiple sites read access the same dataset concurrently. Traditional network file systems are limited by the server’s network bandwidth; therefore cannot satisfy the high-bandwidth concurrent reads that cluster applications typically require. This paper presents NFS-cc: a cooperative caching file system based on NFS and optimized for concurrent read sharing. Cooperative caching used by NFS-cc is a distributed file caching scheme, which allows file system clients to read data from the memory of other clients, instead of the server. If multiple clients cache the same data, they can supply the data to other clients concurrently, which greatly reduces the contention for the server’s network bandwidth. NFS-cc exploits clients’ bandwidth and file caches to improve the aggregate read throughput of the file system. Our preliminary performance measurements show that the aggregate read throughput of NFS-cc was increased with the number of clients and reached as high as 42.6 MB/s with 12 clients connected by a fast Ethernet. NFS-cc also reduced the average block read time by a factor of 50-80% under a variety of concurrent read access patterns compared with NFS.