In Quest of the Bottleneck - Monitoring Parallel Database Systems

Monitoring query processing has proven to be an effective technique to detect bottlenecks in sequential query execution systems' components. Monitoring distributed execution in parallel systems, however, is a difficult task. The monitoring data of all nodes must be collected at the same time to value load-balancing and scheduling. Furthermore, we want to extract not only information about resource allocation of single processes but also allow for monitoring of single tuples and attribute value distribution. The monitoring system we present in this paper is a client-server architecture where each process of the parallel database system is a client of the monitoring server. This architecture enables the user to monitor even load-balancing or scheduling effects. The monitoring extensions, each client is attached with, provide inspection of the processes at a fine granularity. Furthermore user-defined interpretation of monitoring data on the client side enables emulation of special parallel hardware by low costly components.