Within caches, memory is stored in contiguous chunks called cache lines, which are always a power of two in length. If two CPUs write to two separate variables that share the same cache line, they are forced to take turns. This phenomenon is called False Sharing.
To prevent it, pad cache lines with empty data to ensure each independently-written variable occupies its own cache line.
TAKEAWAY: check if any data structure is being written to by multiple threads and check for false sharing.
See also
- Prefer sequential memory access, CPUs predict and prefetch based on locality — false sharing is the dark side of spatial locality: proximity that helps reads hurts concurrent writes - The Actor Model Isolates State Behind Message Passing, No Shared Memory — the actor model eliminates false sharing structurally: no shared memory means no shared cache lines
- System Programming - Compile utilities — ThreadSanitizer detects race conditions on shared data; false sharing is a performance pathology of the same category
- The store-buffer litmus test proves mutual exclusion is not enough — false sharing lives at the same hardware layer (per-core caches, cache lines, store buffers) that makes that test’s surprising outcome possible; it is the performance face of the machinery the memory model reasons about