Why Set the Stop Flag Using memory_order_seq_cst, If You Check It with memory_order_relaxed?

Context

In his "atomic weapons" talk, Herb Sutter demonstrates an example where a main thread sets a stop flag using memory_order_seq_cst and multiple worker threads check the flag using memory_order_relaxed. Sutter explains that using memory_order_relaxed for checking is acceptable due to the negligible latency impact. However, he suggests using memory_order_seq_cst for the operation that sets the flag, without providing a specific reason.

Memory Ordering

Understanding the concept of memory ordering is crucial here. Memory orders define the visibility and synchronization guarantees provided by atomic operations. In this example, the use of memory_order_seq_cst for setting the flag ensures the following:

• The write to the stop flag is made visible to all other threads in a finite period of time.
• No other memory operations can be reordered before the write.

Performance Implications

While using memory_order_seq_cst for the write operation may seem excessive given that the load operation uses memory_order_relaxed, it actually has no significant performance implications. Implementations are required to make atomic stores visible within a reasonable amount of time, regardless of the memory order used.

Benefits of memory_order_seq_cst

While the latency impact of using memory_order_seq_cst for the write operation is minimal, it provides several benefits:

• Thread Safety: memory_order_seq_cst ensures that the write to the stop flag is visible to all threads in a timely manner, minimizing the risk of race conditions.
• Avoidance of Reordering: It prevents other memory operations from being reordered before the write, ensuring that the intended ordering of events is preserved.
• Optimal Parallelization: Keeping the stop flag up to date allows worker threads to efficiently check its value without waiting for unnecessary memory synchronizations.

Conclusion

In conclusion, using memory_order_seq_cst for setting the stop flag in this example is not for performance optimization but to ensure correctness and thread safety. While memory_order_relaxed is acceptable for the load operation, using memory_order_seq_cst for the write operation provides additional guarantees without compromising performance.