DistOS 2023W 2023-02-06

From Soma-notes

Discussion questions

  • Discuss what you think was interesting about Sprite relative to past systems. What was new? What was old?
  • How does AFS compare to NFS, in terms of their design, implementation, and ambition?
  • What is the role of UNIX in the design and implementation of Sprite and AFS?
  • What else came to mind when reading and discussing these papers?
  • What affects scale in AFS? Sprite?
  • What sort of workloads are these systems designed for?

Notes

Sprite and AFS
--------------

 - Sprite is similar to LOCUS in high-level design
 - but Sprite has optimized for performance in various ways
   - when migrating processes, only copy pages that you need
     (if executable is already on other workstation in memory can just use those pages)
   - caching on both client and server
   - caching of files in memory (we have more RAM)
   - efficient distribution of filesystem namespace
 - still not running on many workstations (100 workstations, 6 servers)
 - note the tradeoff between RAM and disks
   - do as much with RAM as possible, just use disks for persistence
     (very much what we do today)
 - Sprite seems very familiar because its caching architecture is very much how we do things today


AFS
 - Andrew is for Andrew Carnegie, this came out of CMU (Carnegie Mellon Univ)
 - AFS was trying for web scale before the web
   - global filesystem
   - every organization was a "cell" and AFS allowed for inter-cell communication
   - but authentication normally didn't work outside a cell so you
     mostly wouldn't see anything except public files
 - took security seriously, integrated with Kerberos for authentication
 - takes a very different approach to accessing files
   - workstation assumed to have a local disk
   - so on open, a file would be copied to the local disk
   - all work on the file would happen locally
   - on close, file would be copied back to the server
 - quirk of the AFS model - close could fail!
   - do you check close for failure normally?

In AFS, the servers are very different from the clients
 - complex Vice and Venus setup, for managing data blocks and metadata separately
 - designed for large installations

AFS workstations couldn't work disconnected, but later systems tried to fix this (Coda, which didn't get that widely used)