Combined Vertical and Horizontal Autoscaling Through Model Predictive Control

Meeting performance targets of co-located distributed applications in virtualized environments is a challenging goal. In this context, vertical and horizontal scaling are promising techniques; the former varies the resource sharing on each individual machine, whereas the latter deals with choosing the number of virtual machines employed. State-of-the-art approaches mainly apply vertical and horizontal scaling in an isolated fashion, in particular assuming a fixed and symmetric load balancing across replicas. Unfortunately this may result unsatisfactory when replicas execute in environments with different computational resources. To overcome this limitation, we propose a novel combined runtime technique to determine the resource sharing quota and the horizontal load balancing policy in order to fulfill performance goals such as response time and throughput of co-located applications. Starting from a performance model as a multi-class queuing network, we formulate a model-predictive control problem which can be efficiently solved by linear programming. A validation performed on a shared virtualized environment hosting two real applications shows that only a combined vertical and horizontal load balancing adaptation can efficiently achieve desired performance targets in the presence of heterogeneous computational resources.