Conclusion

The implementation, as it exists now, is complete and stable. Further investigation of some of the warning messages we log would be useful to gain confidence in the client, but we have not observed data corruption in several weeks of light use. The experimental failed lookup caching can be turned on and off easily at run-time, so does not negatively affect performance or stability.

Ideally, the design and much of the implementation could be integrated into current builds of the development kernel before the next stable Linux kernel is released. The implementation provides an excellent performance improvement when reading already-cached files, and never does significantly worse than the 2.0.x kernel's client. The implementation is reasonably simple and small (the largest memory hits are in the increase in size of the lookup results cache and the storing of page-valid bitmaps for especially large files). It is highly configurable, and can be used with minimal reconfiguring of the rest of the system.

Our enhanced NFS client can substantially improve the performance of Linux boxes that use partitions mounted from a remote NFS server. With a bit of work, our client, along with the improved NFS server already included in development Linux kernels, could greatly increase the usefulness of Linux in a distributed environment.