Queues 201

Costs that hide behind “put and get.”

Memory and cache lines

A queue is a data structure plus coordination. Two threads sharing a counter on the same cache line bounce that line between cores (false sharing). A ring that puts the head and tail on different cache lines avoids that.

Copies

If enqueue copies the payload, cost grows with payload_bytes. Experiment 2 asks whether ranking flips when the payload grows from 256 B to 4 KiB.

Backpressure

A bounded queue makes put wait (or fail) when full. That is a feature: the producer slows down instead of exhausting memory. Unbounded queues hide the problem until the process dies.

Blocking vs spinning vs yielding

• Mutex + condvar — sleeps; good when waits are long.
• Spin — burns a core; good when the other side is almost ready.
• Async yield — gives the event loop to someone else.

This suite times the handoff. It does not tell you which waiting style fits your latency SLO.