Sequential consistency versus linearizability

Abstract

The power of two well-known consistency conditions for shared memory multiprocessors, sequential consistency and linearizability, is compared. The cost measure studied is the worst-case response time in distributed implementations of virtual shared memory supporting one of the two conditions. The memory is assumed to consist of read/write objects. The worst-case response time is very sensitive to the assumptions that are made about the timing information available to the system. All the results in this paper assume that processes have clocks that run at the same rate as real time and that all message delays are in the range [d - u, d] for some known constants u and d, 0 ≤ u ≤ d. If processes have perfectly synchronized clocks or if every message has delay exactly d, then there are linearizable implementations in which one operation (either read or write) is performed instantaneously and the response time of the other operation is d. These upper bounds match exactly a lower bound for sequential consistency, proved by Lipton and Sandberg, on the sum of the response times of read and write operations. If clocks are not perfectly synchronized and if message delays are variable, i.e., u > 0, then such a tradeoff cannot be achieved by linearizable implementations: the response time for both read and write operations is at least Ω(u). In contrast, we present sequentially consistent implementations for this weaker timing model in which one operation (either read or write) is performed instantaneously, and the worst-case response time of the other operation is O(d).