Fastpath Optimizations for Cluster Recovery in Shared-Disk Systems
Proceedings of the 2004 ACM/IEEE conference on Supercomputing
| Hi-index | 0.00 |
To reduce or eliminate the implosion of acknowledgments at the source of a very large multicast group, current reliable multicast protocols either organize the receivers of a group into a tree or ring, or implement NAK-avoidance algorithms based on random timers. All of these approaches have their limitations and advantages. We present a ring-based reliable multicast protocol, called the evenly-loaded ring protocol (ELRP), which removes processing bursts from nodes in the ring. ELRP achieves throughputs as good as those obtained with tree-based protocols and better than those of receiver-initiated protocols with NAK avoidance. We propose a deterministic NAK avoidance (DNKA) algorithm based on ELRP that performs better than the random NAK avoidance scheme (RNKA). Because tree structures are better suited than rings for wide areas involving paths with long delays, we propose the hybrid reliable multicast protocol (HRMP), which organizes receivers into a backbone acknowledgment tree that connects local areas. Any tree-based protocol with positive acknowledgment can be run in the backbone acknowledgment tree, and ELRP is run in the local areas. The maximum achievable throughput in HRMP is analyzed, which shows that it is completely scalable. The delay of DNKA and RNKA schemes are also analyzed which shows that compared with RNKA, DNKA reduces the retransmission delay significantly.