Black-box performance control for high-volume non-interactive systems

This paper studies performance control for high-volume non-interactive systems, and uses IBM Tivoli Netcool/Impact--a software product in the IT monitoring and management domain--as a concrete example. High-volume non-interactive systems include a large class of applications where requests or processing tasks are generated automatically in high volume by software tools rather than by interactive users, e.g., data stream processing and search engine index update. These systems are becoming increasingly popular and their performance characteristics are radically different from those of typical online Web applications. Most notably, Web applications are response time sensitive, whereas these systems are throughput centric.

This paper presents a performance controller, TCC, for throughput-centric systems. It takes a black-box approach to probe the achievable maximum throughput that does not saturate any bottleneck resource in a distributed system. Experiments show that TCC performs robustly under different system topologies, handles different types of bottleneck resources (e.g., CPU, memory, disk, and network), and is reactive to resource contentions caused by an uncontrolled external program.