RConf(PD): Automated resource configuration of complex services in the cloud

Abstract

Optimal deployment of complex services in a virtualized environment is still an open problem. These services typically consist of a set of connected components, and each component may consist of multiple instances. Each instance can in turn be run in different virtual flavors, while the service constructed by the combination of these instances must satisfy a customer Service Level Objective (SLO). While there have been efforts to answer the questions of when to provision additional resources in a running service, and how many resources are needed, the question of what (i.e., which combination of instances) should be provisioned has not been investigated yet. In this work, we offer to service providers the first system that automatically deploys component instances for complex services such that the resource utilization at the providers premises is maximized in the presence of customer constraints. Our system consists of two key technologies (RConf and RConfPD), both of which build on an analytical model based on robust queuing theory to accurately model arbitrary components. With the help of this model, RConf proposes an algorithm to ultimately find the optimal combination of component instances. Our real-world experiments show that, compared to greedy approaches, RConf provisions 20% less resources in the first place, and can reduce resource wastage on live resources by up to 50%. At the same time, RConfPD trades-off some of the optimality of RConf for a computational expense 1–2 orders of magnitude below that of RConf to provision time-sensitive services. Based on a primal–dual algorithm framework RConfPD relaxes the optimality constraints of RConf and removes dominated combinations to determine an approximation for the optimal solution. Our evaluation shows that RConfPD allows for fast decisions (in many cases <1ms), while maintaining 80%–99% of the solution quality of RConf.