Caching in the Sprite Network File System
Nelson, Welch, Ousterhout
ToCS 1988

Why do they want a file server?
  Performance?
  Cost?
  Reliability?
  Sharing?
  Use my files from any workstation?
  What was the alternative?

What consistency model do they provide?
  "same semantics as if all of the processes ...
   were executing on a single ... system"
  What does that mean?
  Is it strict consistency?
  Sequential consistency?

What do they claim about NFS and AFS consistency?
  Why is the NFS/AFS model good?
  Why is it bad?

What's the intuition behind the assumption that write-sharing
  is usually sequential?
  And rarely concurrent.

How does Amoeba deal with concurrent write sharing?

How does Sprite's cache consistency algorithm work?
  They have one file server, and a bunch of client workstations.
  Clients tell the server when they open and close each file.
  Server keeps track of which clients have each file open.
    And also remember whether each is open read-only, or read/write.
  Server views a file in one of three states:
    If no read/write opens, clients may read from caches.
    If exactly one open and it's read/write,
      that client may read/write locally in its cache.
    If multiple opens and >= 1 is for writing.
      no caching is allowed, all reads/writes go to server.

How does a client make sure its cached data is fresh?
  How does client know some other client hasn't modified the file?
  How does client get latest copy of the data?

Why is this arrangement good?
  Common case is only one client workstation is using each file.
    Client can keep all the files in its cache.
    Never has to send/recv data from the file server.
  Multiple clients writing a file is rare.
    Sprite gets this correct, but it's slower.

Look at Figure 4.
  Caching decreases run-time and network traffic by 5x.
  But decreases server utilization by only 1/3.
  Why?

What happens when a Sprite server crashes and restarts?  

There's clearly tension among:
  1. Durability (vs crashes)
  2. Strictness of consistency
  3. Amount of write traffic
  4. Amount of cache validation traffic
  5. Complexity (vs stateless server)
  It's mostly #1 that makes a file system different from a DSM.
    And maybe you have more freedom with #2.
    Since you have a notion of open() and close().