Nonblocking Concurrent Data Structures with Condition Synchronization

Abstract

We apply the classic theory of linearizability to operations that must wait for some other thread to establish a precondition. We model such an operation as a request and a follow-up, each with its own linearization point. Linearization of the request marks the point at which a thread’s wishes become visible to its peers; linearization of the follow-up marks the point at which the request is fulfilled and the operation takes effect. By placing both linearization points within the purview of object semantics, we can specify not only the effects of operations, but also the order in which pending requests should be fulfilled.

We use the term dual data structure to describe a concurrent object implementation that may hold both data and reservations (registered requests). By reasoning separately about a request, its successful follow-up, and the period in-between, we obtain meaningful definitions of nonblocking dual data structures. As concrete examples, we present lock-free dualstacks and dualqueues, and experimentally compare their performance with that of lock-based and nonblocking alternatives.

This work was supported in part by NSF grants numbers EIA-0080124, CCR-9988361, and CCR-0204344, by DARPA/AFRL contract number F29601-00-K-0182, and by Sun Microsystems Laboratories.