A hierarchical and distributed monitoring system in interprocess communications

A hierarchical and distributed performance monitoring tool for a rapid prototyping system for distributed algorithms is presented. The monitoring system consists of a hierarchy of monitoring units distributed over the processor nodes. The units are partitioned into several conditionally independent groups, each of which has a group leader unit. Monitoring data are recorded by each individual unit and are collected and processed by the group leaders. The global picture of program behaviour is formed by the top-level unit of the hierarchy, resulting from a sequence of integrations of lower-level, partially processed data. This distributed, hierarchical nature of the monitoring system architecture offers substantial improvements in the efficiency of processing monitoring data. Furthermore, the architecture lends itself to a parallel implementation. Extensive experiments have been undertaken to make a comparison in performance between the proposed monitoring system architecture on different network topologies and a conventional, centralised monitoring system. Results show that the distributed monitoring system is scalable and that there is a significant performance improvement compared with a centralised system.