False Sharing

struct Foo { int x; int y; int z; }; std::vector<Foo> arr(100'000); # Run in 2 threads for (int i = thread_idx; i < arr.size(); i += 2) { arr[i].x = i; arr[i].y = arr.size() - i; arr[i].z = arr.size() + i; }

struct alignas(64) Foo { int x; int y; int z; }; std::vector<Foo> arr(100'000); # Run in 2 threads for (int i = thread_idx; i < arr.size(); i += 2) { arr[i].x = i; arr[i].y = arr.size() - i; arr[i].z = arr.size() + i; }

False sharing happens when 2 cores read/write to different variables in the same cache line. In the example, Foo struct only has 3 integers, which is smaller than 1 cache line (64 bytes). Therefore, multiple instances of Foo will fit into a single cache line, causing the possibility for 2 cores to read/write at the same cache line.

Adding alignment for Foo struct so that 1 cache line only has 1 Foo instance (at the cost of more and make the program run faster. Alternatively, you can divide the work between 2 threads differently (e.g. 0-size/2 to 1 thread, and the rest to another thread).

To measure effect of false sharing, you can use perf c2c.

References: