Chorus: an interactive approach to incremental modeling and validation in clouds.

Performance modeling is an emerging approach towards automating management in Clouds. The need for fast model adaptation to dynamic changes, such as workload mix changes and hardware upgrades, has, however, not been previously addressed. Towards this we introduce Chorus, an interactive framework for fast refinement of old models in new contexts and building application end-to-end latency models, incrementally, on-the-fly. Chorus consists of (i) a declarative high-level language for expressing expert hypotheses, and system inquiries (ii) a runtime system for collecting experimental performance samples, learning and refining models for parts of the end-to-end configuration space, on-the-fly. We present our experience with building the Chorus infrastructure, and the corresponding model evolution for two industry-standard applications, running on a multi-tier dynamic content server platform. We show that the Chorus on-the-fly modeling framework provides accurate, fast and flexible performance modeling by reusing old approximate models, while adapting them to new situations.