Reliable Multicasting in High-speed LANs

Abstract

There is an increasing number of distributed applications, some of them fault-tolerant, and it has been recognized that its construction may benefit from the existence of reliable broadcast protocols. Reliable broadcasting has deserved considerable attention recently. Some systems are clock-driven, exhibiting tight synchrony: they rely on clock synchronization and space redundancy. Others, like the AMp, an atomic multicast protocol for local area networks, are clock-less.

In application-independent systems, most of the time domain requirements are not of the hard real-time kind — which clock-driven protocols are most suited for — but rather of the on-line, or soft real-time kind. To encourage utilization of reliable broadcast protocols in such applications, it is mandatory that benefits in quality of service are not considered too costly in performance. Clock-less protocols provide an answer to this requirement, since they can trade synchronism for fast termination.

This paper deals with the performance implications of supporting distributed applications, with clock-less reliable broadcast protocols. The question of the network influence in protocol performance is analyzed. We predict the performance of AMp on two target LANs: the l0Mb/s token-bus, and the 100 Mb/s FDDI. We observe that the figures achieved with a LAN based approach are good enough to match the requirements of high performance distributed kernels. Additionally, we show that, under these conditions, using a high-speed LAN does present an advantage: not only throughput increases, a natural consequence, but also speed, measured in duration of single AMp executions, for small messages.

This work has been supported in part by the CEC, through Esprit Project 1226 — DELTA-4, and by JNICT, through Project 87634 — ESTIMULO. This paper contains some material previously published in the proceedings of the ACM SIG-COMM’89 Conference, Austin-USA, September 1989.