Forecasting Lock Contention Before Adopting Another Lock Algorithm

Locks are often used in parallel applications to ensure correct behavior, but frequently induce performance bottlenecks caused by contention at frequently accessed locks. There are state-of-the-art locking algorithms, whose performance under high contention is superior to that of the most commonly used ones. However, some of these have a different interface, and adopting them incurs a reimplementation effort. It is therefore important to efficiently estimate the potential benefit of adopting another lock algorithm, prior to investing effort for reimplementation. We use the Queue Delegation Locking algorithm (QD locking) as an example, which to our knowledge has the best performance among such algorithms. For high-contention scenarios, delegation-based locking algorithms perform well due to their ability to delegate critical sections to a single thread, which requires using a different interface than the traditional lock()/unlock(). In the paper, we present a method that allows predicting the contention under the QD locking algorithm which does not require any reimplementation. In our method, one only needs to gather a set of machine-specific parameters and profile the target application in a single-threaded run to obtain the lock access patterns and holding times. Given these inputs, our method uses queueing networks to predict the contention at each lock, aiding programmers in judging the impact of adopting QD locks. We validate the model using generated benchmarks with multiple locks and showcase the accuracy on an in-memory database application, which is 24% for all 3500 configurations of our benchmark and 14% for critical sections sized no smaller than 1000 nanoseconds.