Optimal Recoverable Mutual Exclusion Using only FASAS

Abstract

Recent research has focused on designing concurrent algorithms that are resilient to process crashes. The idea is to leverage non-volatile memory so that processes can recover from crashes with as little disruption to the normal behavior of the system as possible. We present the first Recoverable Mutual Exclusion algorithm whose Remote Memory Reference (RMR) complexity is optimal for both Cache-Coherent (CC) and Distributed Shared Memory (DSM) machines. If a process fails f times during its attempt to acquire the Critical Section, our algorithm ensures that the process incurs O(1) RMRs on a DSM machine and O(f) RMRs on a CC machine, which we prove is an optimal bound. Our algorithm improves on a recent algorithm by Golab and Hendler in three ways: It has a provably optimal RMR complexity, has a wait-free Exit section, and less reliance on instructions that are not commonly supported on multiprocessors. In particular, Golab and Hendler’s algorithm relies on hardware support for both Fetch-And-Store-And-Store (FASAS) and Double-Word Compare-And-Swap (DCAS), while our algorithm relies only on FASAS. (If X and Y are shared variables and v is a value, FASAS(X, Y, v) writes X’s value in Y and writes v in X, all in a single atomic action.)

The first author is grateful to the Frank family and Dartmouth College for their support through James Frank Family Professorship of Computer Science.

The second author is grateful for the support of Department of Defense (DoD) through the National Defense Science & Engineering Graduate Fellowship (NDSEG) Program.

The third author is grateful for the support of Dartmouth Fellowship.