Chapter 17: Rejected Ideas — Futex, Rest Hubber, and Why They Failed

Futex-Based Thread Wakeup (Rejected)

Idea: Replace epoll-based wakeup with futex for faster thread signaling.

Why rejected:

• Profiling showed epoll overhead was NOT significant
• futex_wait overhead was similar to epoll_wait (both = context switch)
• Would require moving all I/O to ext_adapter thread (non-trivial refactor)
• DB interactions aren't thread-safe for splitting across threads
• Actual test showed PERFORMANCE REGRESSION, not improvement

Rest Hubber in PEP (On Hold)

Idea: Move HTTP server to a separate I/O thread (ext_adapter), dispatch work to main EVL via work queue.

Why on hold:

• Only makes sense if futex wakeup is used (which was rejected)
• Adds thread boundary crossing (latency)
• Complex refactor for uncertain benefit
• No measured bottleneck in current HTTP handling

Time-Based Batching in db-proxy (Not Done)

Idea: Add a timer to db-proxy that accumulates commands before flushing.

Why not done: db-proxy is being replaced by db-mux entirely. No point optimizing the old library.

PEP DB Batching Across Sessions (Not Possible)

Idea: Batch DB operations from different PFCP sessions together.

Why not possible: Different sessions hash to different Valkey slots. MSET requires all keys on same slot. Can't batch across sessions in a cluster.

Mark as Complete