RMR-Efficient Randomized Abortable Mutual Exclusion

Recent research on mutual exclusion for shared-memory systems has focused on local spinalgorithms. Performance is measured using the remote memory references (RMRs) metric. Ascommon in recent literature, we consider a standard asynchronous shared memory model withN processes, which allows atomic read, write and compare-and-swap (short: CAS) operations.In such a model, the asymptotically tight upper and lower bounds on the number of RMRsper passage through the Critical Section is Θ(logN) for the optimal deterministic algorithms [27,7]. Recently, several randomized algorithms have been devised that break the Ω(logN) barrierand need only o(logN) RMRs per passage in expectation [16, 17, 8]. In this paper we present thefirst randomized abortable mutual exclusion algorithm that achieves a sub-logarithmic expectedRMR complexity. More precisely, against a weak adversary (which can make scheduling decisionsbased on the entire past history, but not the latest coin-flips of each process) every process needsan expected number of O(logN/ log logN) RMRs to enter end exit the critical section. If aprocess receives an abort-signal, it can abort an attempt to enter the critical section within afinite number of its own steps and by incurring O(logN/ log logN) RMRs.