Service workload patterns for Qos-driven cloud resource management

Abstract

Cloud service providers negotiate SLAs for customer services they offer based on the reliability of performance and availability of their lower-level platform infrastructure. While availability management is more mature, performance management is less reliable. In order to support a continuous approach that supports the initial static infrastructure configuration as well as dynamic reconfiguration and auto-scaling, an accurate and efficient solution is required. We propose a prediction technique that combines a workload pattern mining approach with a traditional collaborative filtering solution to meet the accuracy and efficiency requirements. Service workload patterns abstract common infrastructure workloads from monitoring logs and act as a part of a first-stage high-performant configuration mechanism before more complex traditional methods are considered. This enhances current reactive rule-based scalability approaches and basic prediction techniques by a hybrid prediction solution. Uncertainty and noise are additional challenges that emerge in multi-layered, often federated cloud architectures. We specifically add log smoothing combined with a fuzzy logic approach to make the prediction solution more robust in the context of these challenges.