Reflective Control for an Elastic Cloud Application: An Automated Experiment Workbench

• Main Authors: Azbayar Demberel, Jeff Chase, Shivnath Babu
• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Subjects: Orca
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.149.914; http://www.usenix.org/events/hotcloud09/tech/full_papers/demberel.pdf

This paper addresses “reflective ” control for applications that use server resources from a shared cloud infrastructure opportunistically. In this approach, an external reflective controller launches application functions based on knowledge of what resources are available from the cloud, their cost, and their value to the application through time. As a driving example, we consider reflective control for an important use of elastic computing: a virtual workbench for digital experiments, focusing on automated benchmarking. We report progress on a Workbench Automation/Intelligence Framework (Waif), and show how it can adapt to available cloud resources by planning and launching experiments in parallel. Waif is part of the ongoing Automat project – an open testbed for programmable hosting centers, built on the ORCA resource leasing platform. We designed a prototype Waif, directed at constructing server performance models by mapping server behavior within a multi-dimensional parameter space. The planner estimates the value and cost of candidate experiments based on the results of completed experiments. In this setting, we show the potential of reflective control to accelerate progress toward a benchmarking objective in a way that balances speed, accuracy, and cost. 1