3. Optimizing a Cloud Contract Portfolio Using Genetic Programming-Based Load Models

Infrastructure-as-a-Service (IaaS) cloud providers offer a number of different tariff structures. The user has to balance the flexibility of the often quoted pay-by-the-hour, fixed price (“on demand”) model against the lower-cost-per-hour rate of a “reserved contract”. These tariff structures offer a significantly reduced cost per server hour (up to 50 %), in exchange for an up-front payment by the consumer. In order to reduce costs using these reserved contracts, a user has to make an estimation of its future compute demands, and purchase reserved contracts accordingly. The key to optimizing these cost benefits is to have an accurate model of the customer’s future compute load – where that load can have a variety of trends and cyclic behaviour on multiple time scales. In this chapter, we use genetic programming to develop load models for a number of large-scale web sites based on real-world data. The predicted future load is subsequently used by a resource manager to optimize the amount of IaaS servers a consumer should allocate at a cloud provider, and the optimal tariff plans (from a cost perspective) for that allocation. Our results illustrate the benefits of load forecasting for cost-efficient IaaS portfolio selection. They also might be of interest for the Genetic Programming (GP) community as a demonstration that GP symbolic regression can be successfully used for modelling discrete time series and has a tremendous potential for time lag identification and model structure discovery.