Please enable JavaScript.

Coggle requires JavaScript to display documents.

LLM Systems Seminar (CS7670): Week07a, Fault Tolerance (paper, HotCRP,…

- - - - fail-stop
      - fail-slow
      - Byzantine failure
      - data corruption/bit-flip model
      - consistency violation
      - silent data corruption (SDC)
  - - - state-machine replication
      - primary-backup replication
    - - deterministic record-replay systems
      - N-version programming
      - verifiable re-execution for safety-critical systems
    - - checkpointing and recovery
      - transaction logging and recovery
- - - - application-level and transparent checkpointing
      - periodic and event-driven checkpointing
      - full and incremental checkpointing
    - - save an instant snapshot of the program
        
        instant => consistent states
        
        snapshot => all states
    - - (gloabl) state consistency
      - non-determinism
      - performance: checkpointing time and runtime interference
  - - - training failures?
        
        training instabilities/loss spikes
        
        new types of problems in RLHF/RLVR
        
        software and hardware failures