| In this thesis, a family of strategies for object pre-fetching in a distributed shared virtual memory (DSVM) system is proposed. All the strategies are based on using a per-object “reference predictor” that is associated with an object's entry in the Global Directory of Objects (GDO) thereby providing object-specific pre-fetching ability. Each such reference predictor records some history of object access patterns based on recent method invocations made by the corresponding object. This information is then used to predict which objects should be pre-fetched when a given object is active. After defining a basic predictor, the concepts of pre-fetching “threshold”, pre-fetching “depth” and “path-based pre-fetching” are introduced to improve both pre-fetching efficiency and accuracy. Finally, the data structures and algorithms required to implement six specific pre-fetching strategies are presented and their applicability is discussed. (Abstract shortened by UMI.)... |
