Deconstructing Redundant Memory Synchronization

In multiprocessor systems with weakly consistent shared memory, memory fence (also know as barrier) instructions are necessary to establish memory consistency at synchronization points in a parallel program. Programs that follow an acquire-release synchronization protocol (e.g., Java) make frequent use of such fence instructions and hence the thrifty use and efficient implementation of such instructions is an important performance aspect. Perhaps surprisingly, this paper demonstrates on the example of multithreaded Java benchmarks that most fences occurring in a program execution are ex post unnecessary according to the rules of memory consistency demanded by the software (max. 99.9%, avg. > 98%). We conclude that the model for memory synchronization offered by current multiprocessor systems is not well aligned with the requirements of the widely used acquire-release software synchronization protocol.