Cache Consistency: Frangipani

http://nil.csail.mit.edu/6.824/2020/papers/thekkath-frangipani.pdf

http://nil.csail.mit.edu/6.824/2020/papers/frangipani-faq.txtnil.csail.mit.edu

• I like they don't shy away from saying they didn't explore something or they haven't got enough experience with something.

• Small data blocks and large data blocks separately. Clever.

• Very decentralized. All the file system logic is on the clients.

  • Challenges with consistency then.

Challenges:

• Cache coherence.

• Atomicity.

• Crash recovery.

Coherence is solved with locks.

Atomic multi-step operations:

• Distributed transactions.

• First acquire all the locks I need. Then do all updates. Then release all the locks.

• But can crash in the middle of doing the updates. With locks.

  • Can't just release its locks.

  • So do write-ahead logging (WAL).

  • Separate per-workstation logs, in the shared storage.

  • Logs contain only metadata-changes, not file contents.

  • Log initially is in-memory; written to Petal only when it has to be.

• In revoke:

  • Write log to Petal.

  • Write modified blocks for locks.

  • Send release.

• If a workstation crashes while holding a lock, another workstation will replay its log! And only then the lock will he released.