Modeling prediction errors
(June 2016 - present)
Software deployments must be carefully provisioned to meet their performance requirements without wasting resources. Most resource provisioning solutions today employ predictions to estimate future demand and provision accordingly. However, naively employing predictions can negate its benefits. For example, provisioning resources based only on the mean of predictions can result in severe performance violations due to uncertainty in the predictor. On the other hand, while additionally provisioning for 2 standard deviations can eliminate performance violations, it can substantially increase resource wastage. The goal of this project is to develop and leverage error models to fully realize the potential of predictors.
The key intellectual contribution of this project is to bridge the gap between predictors and resource
provisioning solutions by investigating the prediction error model. This will be accomplished via three main thrusts: (i) constructing error models that capture the structure of prediction errors, including correlations and prediction quality over time; (ii) developing an algorithmic framework to incorporate the prediction error models and account for switching costs and penalty functions; and (iii) designing systems to exploit the new prediction error-aware algorithms, including multi-resource provisioning and resource placement solutions. The solutions will be experimentally evaluated using available application traces.
© Copyright 2014-2017 PACE Lab, Stony Brook University