Modeling and Analyzing the Performance of Lock-Free Data Structures

Rapport, 2014

This paper considers the modeling and the analysis of the performance of lock-free concurrent data structures. Lock-free designs employ an optimistic conflict control approach, allowing several processes to access the shared data object at the same time. The operations on these data structures are typically designed as compositions of retry loops. Our main contribution is a new way of modeling and analyzing a general class of lock-free algorithms, achieving predictions of throughput that are close to what we observe in practice. In our model we introduce two key metrics that shape the performance of lock-free algorithms: (i) expansion in execution time of a retry due to memory congestion and (ii) number of wasted retries. We show how to compute these two metrics, and how to combine them, to calculate the throughput of an arguably large class of lock-free algorithms. Our analysis also captures the throughput performance of a lock-free algorithm when executed as part of a larger parallel application. This part of our analysis leads to an analytical method for calculating a good back-off strategy to finely tune the performance of a lock-free application. Our experimental results, based on a set of widely used concurrent data structures and on abstract lock-free designs, show that our analysis follows closely the actual code behavior. To the best of our knowledge, this is the first attempt to make ends meet between theoretical bounds on performance and actual measured throughput.