Browse Theses

This dissertation presents a new network file system, called Zebra, that provides high performance file access and is highly available. Zebra stripes file data across its servers, so that multiple servers may participate in a file access and the file access bandwidth therefore scales with the number of servers. Zebra is also highly available because it stores parity information in the style of a RAID (Patterson88) disk array; this increases storage costs slightly but allows the system to continue operation even while a single storage server is unavailable.

Zebra is different from other striped network file systems in the way in which it stripes data. Instead of striping individual files (file-based striping), Zebra forms the data written by each client into an append-only log, which is then striped across the servers. In addition, the parity of each log is computed and stored as the log is striped. I call this form of striping log-based striping, and its operation is similar to that of a log-structured file system (LFS) (Rosenblum91). Zebra can be thought of as a log-structured network file system: whereas LFS uses a log abstraction at the interface between a file server and its disks, Zebra uses a log abstraction at the interface between a client and its servers. Striping logs, instead of files, simplifies Zebra's parity mechanism, reduces parity overhead, and allows clients to batch together small writes.

I have built a prototype implementation of Zebra in the Sprite operating system (Ousterhout88). Measurements of the prototype show that Zebra provides 4-5 times the throughput of the standard Sprite file system or NFS for large files, and a 15-300% improvement for writing small files. The utilizations of the system resources indicate that the prototype can scale to support a maximum aggregate write bandwidth of 20 Mbytes/second, or about ten clients writing at their maximum rate.