An Adaptive Protocol for Implementing Causally Consistent Distributed Services

Distributed services that are accessed by widely distributed clients are becoming common place (e.g., especially services targeted to the home). Such services cannot be provided at the needed level of performance and availability(1) without replicating the service at multiple nodes, and (2) without allowing a \relatively" weak level of consistency among replicated copies of the state of a service. The latter is especially important in large scale systems where communication latencies can be high and communication may not be possible between service instances at all times. In this paper we explore causally consistent distributed services when multiple related services are replicated to meet performance and availability requirements. This consistency criterion is particularly well suited for some distributed services (e.g., cooperative document sharing), and it is attractive because of the eecient implementations that are allowed by it. We develop a new protocol for implementing causally consistent services that has several novel features. First, service instances can be created and deleted dynamically according to service access patterns in the distributed system. Also, when diierent services are related and consistency is needed across replicated copies of such services, our protocol handles the case where diierent but related services are replicated independently. Another novel aspect of our protocol lies in its ability to use both push and pull mechanisms for disseminating updates to objects that encapsulate service state.