Nonblocking memory management support for dynamic-sized data structures

Reviewer: Sergei Gorlatch

An in-depth discussion of a new mechanism for lock-free data structures that allow memory to be freed dynamically is presented in this paper. Lock-free algorithms do not rely on locking for thread safety, but instead use certain atomic operations (usually provided in hardware), such as compare-and-swap. While lock-free algorithms avoid many problems caused by locking, a common problem is that data structures cannot easily be shrunk because they can be used by other threads concurrently. The paper proposes a simple mechanism and application programming interface (API) to overcome this problem: prior to using a data item, threads post a guard on the item and release this guard only when they do not use the reference anymore. Before a data item can be released, a special "liberate" method is called to check that the item is not guarded by any other thread. The paper applies the presented mechanisms to two examples: a simple stack implementation, which serves as an initial illustrating example, and the lock-free first in first out (FIFO) queue algorithm of Michael and Scott. The paper provides an implementation for the proposed scheme, and presents a correctness proof for the implementation. The authors also present experimental results, documenting that the cost incurred by the presented mechanisms is negligible under low contention, and modest for high contention. In addition to the FIFO case study, a general methodology (single-word lock-free reference counting) is also presented. The paper provides a great deal of insight into the problems (and solutions) of lock-free data structures, and also provides many pointers to related work. Online Computing Reviews Service