PRACTICAL PERFORMANCE ESTIMATION ON SHARED-MEMORY MULTIPROCESSORS
Paul E. McKenney · 2002
There has been much work done modeling, simulating, and measuring the performance of locking primitives under high levels of contention. However, an important key to producing high-performance parallel programs is to maintain extremely low levels of contention. Despite its importance, the low-contention regime has been largely neglected. In order to fill this gap, this paper analyzes the performance of several commonly used locking primitives under low levels of contention. The costs predicted by a number of analysis methodologies are compared to measurements taken on real hardware, thereby showing where each may be safely used. Use of