Automatic Distribution of Shared Data Objects

Abstract

Object-based distributed shared memory is a form of shared virtual memory in which shared data are encapsulated in objects. One of the main research issues related to shared objects is how to implement them efficiently on a distributed-memory system. Our Orca system [1, 3] uses two distribution strategies. In the simplest case, an object is stored on one processor (single-copy mode), and other processors access it through remote invocations. To reduce communication overhead, each object is migrated to the processor that accesses it most frequently. The other strategy we use is to replicate shared objects that have a high read/write ratio. With a replicated object, read-only operations can be executed locally. Operations that change the object update all copies in a consistent way, using totally-ordered group communication [2]. For reasons of transparency and portability we want to decide automatically which of these strategies to use for a given object.

This research is supported by a PIONIER grant from the Netherlands Organization for Scientific Research (N.W.O.).