Recoverable, Abortable, and Adaptive Mutual Exclusion with Sublogarithmic RMR Complexity (2011.07622v3)

Abstract: We present the first recoverable mutual exclusion (RME) algorithm that is simultaneously abortable, adaptive to point contention, and with sublogarithmic RMR complexity. Our algorithm has $O(\min(K,\log_W N))$ RMR passage complexity and $O(F + \min(K,\log_W N))$ RMR super-passage complexity, where $K$ is the number of concurrent processes (point contention), $W$ is the size (in bits) of registers, and $F$ is the number of crashes in a super-passage. Under the standard assumption that $W=\Theta(\log N)$, these bounds translate to worst-case $O(\frac{\log N}{\log \log N})$ passage complexity and $O(F + \frac{\log N}{\log \log N})$ super-passage complexity. Our key building blocks are: * A $D$-process abortable RME algorithm, for $D \leq W$, with $O(1)$ passage complexity and $O(1+F)$ super-passage complexity. We obtain this algorithm by using the Fetch-And-Add (FAA) primitive, unlike prior work on RME that uses Fetch-And-Store (FAS/SWAP). * A generic transformation that transforms any abortable RME algorithm with passage complexity of $B < W$, into an abortable RME lock with passage complexity of $O(\min(K,B))$.