Why Threads Are A Bad Idea (for most purposes)

Today's summary is about a paper written by John Ousterhout in 1995. This is technically not a paper, but a presentation given by the author at a Usenix conference.

1. Threads are independent execution streams that shared state (memory) and are often preemptively scheduled by the OS/runtime.
2. Threads are too hard for most programmers because:
   • Synchronization: access to shared data must be coordinated with locks.
   • Deadlock: Circular dependencies among locks.
   • Hard to debug: data dependencies, timing dependencies.
   • Threads break abstraction: can't design modules independently.
   • Callbacks don't work with locks.
   • Locks trade-off: Coarse-grain locking yields low concurrency; Fine-grain locking increases complexity and locking overhead.
   • OSes limit performance: scheduling, context switches.
3. Event-driven programming:
   • One execution stream: no CPU concurrency.
   • Register interest in events (callbacks).
   • Event loop waits for events and invokes handlers.
   • No preemption of event handlers.
   • Handlers are generally short-lived.
4. Events are used in most GUIs:
   • One handler for each event (press button, invoke menu entry, etc.).
   • Handler implements behavior (undo, delete a file, etc.).
5. Problems with events:
   • Long-running handlers make the application non-responsive.
   • Can't maintain local state across events (handler must return).
   • No CPU concurrency.
   • Event-driven I/O not always well supported (e.g. poor write buffering).
6. Events vs Threads:
   • Events avoid concurrency as much as possible, threads embrace.
   • Debugging is easier with events.
   • Events are faster than threads on a single CPU because of no context switching or locking overhead.
   • Threads provide true concurrency.
7. The conclusion is to avoid threads wherever possible:
   • Use events, not threads, for GUIs, distributed systems, low-end servers, etc.
   • Only use threads where true CPU concurrency is needed.
   • Where threads are needed, isolate usage in threaded application kernel: keep most of the code single-threaded.