Evaluation of Causal Distributed Shared Memory for Data-race-free Programs

Abstract

Distributed Shared Memory (DSM) is becoming an accepted abstraction for programming distributed systems. Although DSM simplifies the programming of distributed applications, maintaining a consistent view of shared memory operations across processors in a distributed system can be expensive. The causal consistency model of DSM can allow more efficient implementations of DSM because it requires that only causally ordered memory operations be viewed in the same order at different processors. Also, weakly ordered systems have been proposed which advocate the use of synchronization information to reduce the frequency of communication between processors. We have implemented a system that exploits both the weaker consistency of causal memory and the synchronization information used in weakly ordered systems. Consistency is ensured by locally invalidating data that is suspected to be causally overwritten and this is only done when certain synchronization operations complete at a processor. Data-race-free programs can be written in this system assuming that the system provided sequentially consistent memory. By implementing applications that have a variety of data sharing patterns, we show that performance comparable to message passing implementations of the applications can be achieved in the causal DSM system. The improved performance is due to a significant reduction in communication costs compared to the implementation of a strongly consistent memory system. These results show that causal memory can meet the consistency and performance requirements of many distributed applications.