The Path of the Private FUTEX
summary
Kernel futex paths change to cut latency on contended private futexes by removing a global hash bottleneck, adding a lockless wake mechanism and a futex_nqpi helper for PI handling; IPC MSG and related patches merge upstream and show 10–15% throughput gains on a test MD box. Hashing is rewritten to avoid expensive division (replacing mod with multiply inverses) dropping hash cost from ~7% to ~2% of CPU, and microbenchmarks reach about 6M ops/s with degraded but acceptable scaling to 64 threads.
AI-written summary · about
chapters (23)
- 0:00Intro
- 0:18Wait and Wake
- 2:12Sing EVD
- 5:45Assembly Code
- 7:07Global Hash Table
- 8:47Ping Pong Boost
- 11:49FUTEX Wake
- 13:08RT
- 13:38Global Hesh
- 15:43Few Text Attached
- 18:34What We Learned
- 19:23Hashing Algorithm
- 21:13Performance
- 23:55Hash Queue
- 25:19Hash Prime
- 25:50Multiplication
- 29:03Collision Free
- 34:48Attach
- 35:35Performance Numbers
- 36:30Btree
- 37:03Hash Table
- 37:35Numbers
- 37:55Outcome
