Measuring Consistency Costs for Distributed Shared Data

Distributed Shared Memory (DSM) systems typically support one consistency model [3, 5, 6]. However, recent work [1, 11, 12, 14, 16] proposes the use of adaptive consistency based on a heuristical analysis of recent access patterns. Although heuristic-based approaches can significantly improve runtime, the access pattern alone does not necessarily define the most appropriate consistency model. The size of updates and other factors related to the computing environment, such as heavily loaded links, heavily loaded nodes, bursty traffic patterns, and network latency all affect performance. Multiple access patterns within the application also make it difficult to select the most appropriate consistency model. This paper presents a measurementbased approach to the problem of selecting the most appropriate consistency model for the current application in the current runtime environment. We show that measurementbased analysis provides an accurate estimate of performance and therefore can be used to select the most appropriate consistency model, even in cases where hueristicbased approaches fail to select the appropriate model.