Design and Implementation of the Sun NFS

Goals:

- Machine and OS independence
  - Not tied to Unix

- Crash recovery
  - Stateless 

- Transparent access
  - Hides server identity, which is used only at mount time

- Unix semantics maintained on client
  - File locking and concurrent access

- Performance

NFS protocol
- How is it specified:

  As a set of RPC procedures with inputs, outputs and their effects.

  - Stateless --> Simple recovery

  - Transport independent.  First implementation uses UDP/IP.

  - File handle.  Opaque.

  - XDR.  External Data Representation.

NFS server side
- Stable writes.  Data and metadata.

- Add generation number in the inode, which is reflected in the fhandle.

Client side

- Transparent access to remote files.
  - Server hostname is only used at mount time.

VFS vs. VNODE operations

Namespace traversal done a component at a time.

- Implementation

  - Hard vs. soft mount

  - nfsiods and nfsds.

- Hard issues

  - Root filesystems
    Architecture-dependent binaries
    /tmp and /dev hard to share
  - Credentials
    Lack of flat uid, gid space over entire network
  - File locking
    Need additional protocol to do the locking.
    Should be mandatory, not advisory.
  - Concurrent access
    Unix does file locking per I/O, so that I/Os serialize.
  - Unix open file semantics
    Allows files to be deleted but still be accessible if not
    closed.  Requires renaming with NFS.
  - Time skew.
    Can be a problem, for example with make.

Performance
  - Care about end-to-end, not benchmarks.

    Initial performance bad.  The graphs are missing units.

    What improvements did they try?

    - Client caching
    - Large UDP datagrams
    - avoid many bcopies
    - cache attributes (getattr)
    - Swap in small executables to take advantage of server readahead
      in case of client readahead.
    - Name cache

- Future improvements

    - Full diskless operation
      - To avoid dependance on ND for root fs
    - Remote file locking
      - Presently, it is advisory.
    - Support other clients.
      - CIFS is a bit different, more emphasis on locking.
    - Performance
      - Build dedicated NFS server.
    - Improved security
    - Automatic mounting
      - See amd (automount daemon)

Things not mentioned in this paper

    - Open-to-close consistency
    - Retransmission cache (at-most-once semantics)
    - Improve write performance
    - v3 add asynchronous reads/writes