Skip to content

Last updated：2026-10-12

C++ False Sharing and Cache Lines

I recently came across a piece of code that performs alignment manually, so I looked into what impact misalignment can have.

template <typename T>
class AtomicWithPadding {
 private:
  static constexpr int kCacheLineSize = 64;
  uint8_t padding_before[kCacheLineSize];

 public:
  std::atomic<T> value;

 private:
  uint8_t padding_after[kCacheLineSize];
};

This leads straight into a must-know interview topic:

  • Cache line

Here are a few interview questions that follow from it — can you answer them?

  • What is a cache line?
  • What is the MESI protocol?
  • What is false sharing?
  • How do you guarantee avoiding false sharing in C++?
  • What benefits can avoiding false sharing bring?
  • Do you know which two variables C++17 and later provide to solve this problem?

So I found a performance test whose general idea is:

  • Define 4 variables a, b, c, d, shared among multiple threads
  • In the other case, align these four variables, then spawn n threads to be scheduled.

Here is the benchmark:

------------------------------------------------------------------
Benchmark                        Time             CPU   Iterations
------------------------------------------------------------------
singleThread                  4.05 ms         4.04 ms          133
directSharing/real_time       8.78 ms        0.069 ms           77
falseSharing/real_time        8.81 ms        0.083 ms           74
noSharing/real_time           1.16 ms        0.071 ms          578

You can see that noSharing achieves an excellent result! The cache hit rate improves greatly; of course, analyzing it with perf would be even better, to look at the L1 cache misses.

The answers above have been published in the Knowledge Planet — feel free to join the discussion or exchange thoughts in the comments, thanks~

Comments