Performance of the ISIS Distributed Computing Toolkit

Abstract

The ISIS Toolkit is a programming environment for building process-group structured distributed software. The system is widely used in settings requiring high reliability, strong distributed consistency guarantees, and highspeed communication. In this paper, we describe experimental studies of ISIS performance. Our work explores the impact of hardware support for multicast performance, with a focus on flow control mechanisms. The use of hardware multicast in ISIS has not been discussed elsewhere. One conclusion of the paper is that although ISIS performance is limited primarily by flow control considerations, this type of hardware support can lead to significant performance improvements for certain communication patterns. A second conclusion was that the ISIS flow-control problem is surprisingly difficult. More work in this area, and on the underlying operating system communications layer (UDP), could have significant impact on the system. Keywords and phrases: Distributed computing, performance, process groups, atomic broadcast, causal and total message ordering, cbcast, abcast, multiple process groups, hardware multicast, IP multicast, virtual synchrony, fault- tolerance.