Sandhya Dwarkadas, Willy Zwaenepoel
This paper presents an evaluation of three software implementations of release consistency. Release consistent protocols allow data communication to be aggregated, and multiple writers to simultaneously modify a single page. We evaluated an eager invalidate protocol that enforces consistency when synchronization variables are released, a lazy invalidate protocol that enforces consistency when synchronization variables are acquired, and a lazy hybrid protocol that selectively uses update to reduce access misses. Our evaluation is based on implementations running on DECstation-5000/240s connected by an ATM LAN, and an execution driven simulator that allows us to vary network parameters. Our results show that the lazy protocols consistently outperform the eager protocol for all but one application, and that the lazy hybrid performs the best overall. However, the relative performance of the implementations is highly dependent on the relative speeds of the network, processor, and communication software. Lower bandwidths and high per byte software communication costs favor the lazy invalidate protocol, while high bandwidths and low per byte costs favor the hybrid. Performance of the eager protocol approaches that of the lazy protocols only when communication becomes essentially free.
1995
