Check out the new USENIX Web site. next up previous
Next: Optimistic replica coordination Up: Introduction Previous: Introduction

Graph-based replica management

Pangaea's replica management must satisfy three goals. First, it must support a large number of replicas, to maximize availability. Second, it needs to manage the replicas of each file independently, since it is difficult to predict file-access patterns accurately in a wide area. Third, it needs to support dynamic addition and removal of replicas even when some nodes are not available. Pangaea addresses these challenges by maintaining a sparse, yet strongly connected and randomized graph of replicas for each file. The graph is used both to propagate updates and to discover other replicas during replica addition and removal. This design offers three important benefits:

Available and inexpensive membership management:
A replica can be added by connecting to a few live replicas that it discovers, no matter how many other replicas are unavailable. Since the graph is sparse, adding or removing a replica involves only a constant cost, regardless of the total number of replicas.

Available update distribution:
Pangaea can distribute updates to all live replicas of a file as far as its graph is connected. The redundant and flexible nature of graphs makes them extremely unlikely to be disconnected even after multiple node or link failures.
Network economy:
The random-graph design facilitates the efficient use of wide-area network bandwidth, for a system with an aggressive replication policy. Pangaea achieves this by clustering replicas in physical proximity tightly in the graph, and by creating a spanning tree along faster edges dynamically during update propagation.


next up previous
Next: Optimistic replica coordination Up: Introduction Previous: Introduction
Yasushi Saito 2002-10-08