Peridot: Towards Automated Runtime Detection of Performance Bottlenecks

摘要

Performance analysis of parallel applications can be a time-consuming and daunting task, as it requires detailed a understanding of the interactions of the system's components. We present the design and the prototypical implementation of a system for the automation of the performance analysis process developed within the Peridot project. Our system is based on the notion of cooperating agents that detect performance problems automatically at runtime and in a distributed fashion, avoiding several problems of classical performance analysis techniques such as overwhelmingly large trace files.