A short lock can still build a queue
A lock convoy is a pattern in which many operations queue behind one contended lock and progress is repeatedly limited by the owner or by handoff delays. The protected work can be short while waiting, waking and scheduling still dominate the overall cost. Merely finding a mutex in the program is not evidence of a convoy. Look for a queue of operations whose progress depends on the same serial bottleneck. Its severity depends on the mutex implementation, scheduler behavior and arrival pattern of the workload being measured.
A stalled owner delays its followers
Suppose sixteen workers repeatedly remove items from one shared queue protected by a mutex. If an owner stops running while holding the lock, the other workers cannot access the queue. After release, handoff and wakeup costs may keep a line of waiters moving slowly even though each removal does little work. This is a hypothetical execution scenario, not a benchmark result. The worker count alone proves nothing: observations must connect actual waiting workers, the lock owner’s progress and the queue operations they are trying to perform.
Separate hold time from wait time
Separate acquisition wait time from lock hold time. Also examine throughput, context switches and CPU scheduling. Long waits with short holds suggest contention or handoff overhead, but interruptions, CPU quotas and unrelated descheduling may contribute. Fair acquisition order can prevent starvation, so replacing a fair lock with an unfair one is not automatically an improvement. Compare a distribution of delays rather than relying only on a mean that hides occasional stalls. Instrumentation itself can perturb a hot lock; keep its cost in mind when interpreting very short critical sections.
Reduce contention and reevaluate
First reduce unnecessary I/O, computation and shared-state scope inside the critical section. Sharded queues, batched removals or per-worker state can reduce acquisition frequency. Each changes something important: sharding may affect global order and load balance, while batching can increase the delay of an individual item. Spinning is not a universal remedy when the owner cannot run; it may simply spend more CPU waiting. Evaluate throughput and tail latency together under the intended scheduler and workload, and preserve the queue’s correctness and ordering requirements during the experiment.