Performance of the Horus Asynchronous Group Communication System under HighLoad

Abstract

Horus is an asynchronous group communications system designed to support a variety of distributed and parallel applications. This paper reports the performance of continuous $n \times n$ multicasts for a Horus protocol stack that provides fully reliable communication over Ethernet supporting IP multicast. This communication pattern not only is highly demanding from the communications point of view, but also is common for an important class of distributed applications. Specific findings reported in this paper are: providing reliability is expensive under high loss; the cost of providing full reliability is modest compared to that of providing simple reliability; the lack of receivers' resources contributes to high loss; the lack of synchronization among group members also contributes to high loss; and controlling the send rate based on receivers' resource availability is a promising method for maintaining high throughput under high load.