Extensible, Scalable Monitoring for Clusters of Computers
Eric J. Anderson, David A. Patterson · 1997
Introduction Monitoring a large cluster of cooperating computers requires extensibility, fault tolerance, and scalability. We handle the evolution of software and hardware in our cluster by using relational tables to make CARD extensible. We detect and recover from node and network failures by using timestamps to resynchronize out system. We improve data scalability by using a hierarchy of databases and a hybrid push/pull protocol for efficiently delivering data from sources to sinks. Finally, we improve visualization scalability by statistical aggregation and using color to reduce information loss. We have designed and implemented a system for flexible, online gathering and visualization of statistics and textual information from hundreds of data sources. CARD gathers node statistics such as CPU and disk usage, and node information such as executing processes. We will describe the system we developed to monitor our cluster, explain how we solved the four problems described a