Lock reservation for java reconsidered

Abstract

Lock reservation, a powerful optimization for Java locks, is based on the observation that, in Java, each lock tends to be dominantly acquired and released by a specific thread. Reserving a lock for such a dominant thread allows the owner thread of the lock to acquire and release the lock without any atomic read-modify-write instructions. A recently proposed algorithm has embodied this idea and significantly reduced the synchronization overhead on a reservation hit. However, on a reservation miss, the algorithm stops the owner thread in order to cancel the reservation, which incurs a significant performance penalty. We propose a new algorithm for lock reservation for Java without such penalties. We derive the algorithm in two steps. First, we create a new, reservation-based algorithm for spin lock. Second, observing that the conventional spin lock is embedded in a widely-used Java lock, we attempt to replace it with our new spin lock. We evaluated our algorithm in IBM's production virtual machine and JIT compiler. The results show that our algorithm attained comparable speedups in the SPECjvm98 benchmarks, and that it even improved the performance of two scientific programs which the previous algorithm actually degraded. © Springer-Verlag 2004.