Chapter 20 — False Sharing vs. Cache-Line Padding
Two threads that touch completely disjoint data can still serialize on each other if that data shares a 64-byte cache line.
The full source for this example lives in
examples/concurrency/20-mechanical-sympathy/— clone the repo,cdin, and./demo.sh.
Chapter reference: Doc 20 — Mechanical Sympathy
What it shows
Two threads, pinned to separate CPUs, each fetch_add their own atomic counter N times, with a --pad flag choosing whether the two counters sit on the same 64-byte cache line or each get their own hardware_destructive_interference_size-aligned line. Both counters reaching the exact expected total is checked as a hard assertion in both layouts, since false sharing is a performance bug, never a correctness one; the padded run averaging measurably faster than the unpadded run is checked as a soft assertion against a noise floor, since the effect’s size depends on the two threads actually landing on separate physical cores. This is the reference example the rest of the project’s directory shape, verify.lua contract, and demo.sh contract are copied from.
Running it
cd examples/concurrency/20-mechanical-sympathy
./demo.sh # build + run false_sharing (both modes) + verify.lua
./demo.sh --clean # remove the image