Day 3: Conferences: Knfs

Up to 2.0 kernels, the NFS server was a user level mode UID/GID mapping, full root export. The problems with it where:

• Extra memory copies
• context switches all the time
• Not multi-threaded
• No support for file or record locking
• Server must be killed each time /etc/exports is modified

With Knfs (or Kernel NFS), which was started by Olaf Kirch, never fully finished and then handed from person to person is now maintained by HJ. Its strenghts are:

• Implemented in the kernel
• Much more efficient (no extra copies or context switches)
• Multi threaded
• Support NLM and NSM for locking
• Server doesn't have to be restarted when /etc/exports is modified

One of the problems was inter-operability between the different NFS implementations, and KNFS participated to the Connectathon sponsored by Sun and passed the test suite. Actually the user level version also passed the test suite, except for the file locking parts.

One of the problems was that KNFS uses dentries on linux, and it causes problems when the server was rebooted. HJ with the help of others worked on those and now KNFS 1.4.2 and later fix some of those.

When testing KNFS against the old user level server, the kernel threads offer a definite win against the old server, but the most noticeable part is the option of increasing the number of threads and the sweet spot seems to be at 8 threads.

For now, there is no stable NFS v3 or TCP support, but most of the code is there, and there are patches. HJ is hoping that other people will step in and work on those.

As for Linux as an NFS client, there are well known performance issues, as well as operability issues with some NFS servers, like 64 bit SGI servers.

You may be able to decipher (sorry) his slides, which you can find in the middle of the picture library

Picture library

Back to Main Page

Next page

99/08/13 (10:02): Version 1.0