Grant: Collaborative Research and Development Grants

This project investigates on-line multi objective optimization over a finite moving horizon for a cloud computing infrastructure. Currently, cloud resources are allocated to applications in a static manner, based on the perceived workloads at the time of the deployment. Due to up and down swings in workloads, this static deployment results in over or under provisioning hence yielding high costs, waste of resources and poor quality of services. At the same time, reactive on-line optimization, as suggested by current research, leads to oscillations in provisioning and unprovisioning of resources which in turn lead to application failures and hence poor quality of service. The first goal of the project is to develop an on-line method that combines forward and feed-back optimization by using a cloud predictive model. The method optimizes a multi-objective function that captures different stakeholder's objectives, policy constraints, cost of resources and risk. The multi-objective function is optimized over a look ahead time window. An optimization algorithm computes the optimal system control inputs for each time step of the look ahead time window. To account for errors, at each time step, a feedback control loop re-computes the optimization steps and update the predictive model. To scale for large clouds that run a large number of applications, the optimization techniques will be implemented in a decentralized architecture. A second goal of the project is to investigate design techniques for large adaptive systems, that is, systems that change their behaviour and structure to cope with changes in environmental conditions.