Why Non-blocking Operations Should be Selfish

Abstract

Non-blocking data structures are often analysed by giving an upper amortised running time bound in terms of the size of the data structure and a measure of contention. The two most commonly used measures are the point contention $c_P$, the maximum number of processes active at any one time during an operation, and the interval contention $c_I$, the number of operations overlapping with a given operation. In this paper, we show that when summed across every operation in an execution, the interval contention $c_I$ is within a factor of 2 of the point contention $c_P$. Our proof relies on properties of interval graphs where at least one simplicial vertex exists, and uses it to construct a lower bound on the overall point contention. We show that this bound is tight.

This result contradicts the folklore belief that point contention leads to a tighter bound on complexity in an amortised context, and provides some theoretical grounds for recent observations that using less helping in non-blocking data structures can lead to better performance. We also propose a linked list algorithm based on Fomitchev and Ruppert’s algorithm but with selfish operations: read-only operations that do not help others but rather execute wait-free. The higher performance of our approach compared to the original list confirms that reducing helping can increase performance, with the same asymptotic amortised complexity.