A Reliable Multicast Approach To Replica Management for Grids

In this paper we describe a new proposal for update propagation in the context of multi-master Replica Management Systems. Data replication is a well-known technique aimed at maintaining a number of copies of the same data item at geographically distributed sites. Our approach consists in lazily propagating updates through reliable multicast transmissions, while replica state estimation is accomplished by an original variant of the well-known Timestamp Matrix technique. More in detail, each site multicasts its Times Stamp Vector to all the other sites so that they can main tain their own Timestamp Matrix. The Timestamp Matrix makes it possible for a site to decide whether to reliably multicast the updates it is in possession of, or not, thus supporting epidemic propagation of updates. When compared to previous approaches, our solution reduces the amount of transmitted log update information, decreases propagation time and potentially increases system scalability with respect to the number of sites. Our model captures the requirements of data-intensive replica services in scalable Grid environments. These services require systems to automatically exchange update files in order to reach global temporal consistency. A Java prototype of our update propagation system has been implemented and func tionality tests have been performed showing the effectiveness of our approach.