The Commutativity Quotients of Concurrent Objects
share
Concurrent objects form the foundation of many applications that exploit multicore architectures. Reasoning about the fine-grained complexities (interleavings, invariants, etc.) of those data structures, however, is notoriously difficult. Formal proof methodologies for arguing about the correctness – i.e., linearizability – of these data structures are still somewhat disconnected from the intuitive correctness arguments. Intuitions are often about a few canonical executions, possibly with few threads, whereas formal proofs would often use generic but complex arguments about arbitrary interleavings over unboundedly many threads. As a way to bring formal proofs closer to intuitive arguments, we introduce a new methodology for characterizing the interleavings of concurrent objects, based on their commutativity quotient. This quotient represents every interleaving up to reordering of commutative steps and, when chosen carefully, admits simple abstractions in the form of regular or context-free languages that enable simple proofs of linearizability. We demonstrate these facts on a large class of lock-free data structures and the infamously difficult Herlihy-Wing Queue. We automate the discovery of these execution quotients and show it can be automatically done for challenging data-structures such as Treiber's stack, Michael/Scott Queue, and a concurrent Set implemented as a linked list.
READ FULL TEXT