|
Abstract : |
Recent analyses of end-to-end Internet communication have revealed widely-varying application perceived performance. This suggests that Internet-wide distributed databases with global serializability requirements may perform poorly. Nevertheless, there exist some distributed databases that can be designed for acceptable performance on the Internet. A key characteristic of these database is a tolerance for convergent, non-transactional replication. This paper describes the design of a non-transactional replication scheme that uses IP multicast. The IP multicast service is unreliable; we use a stochastic wait with suppression technique to scalably recover lost updates. This technique reduces duplicate retransmissions of lost updates. Detailed simulations demonstrate that, even in large sparsely-distributed databases, less than 10 % of the losses result in duplicate retransmissions. For the expected update traOEc patterns, our approach outperforms other multicast loss recovery mechanisms. We have implemented a distributed Internet routing policy database using this replication scheme; a small experiment demonstrates the scheme's feasibility. 1, |
