IaaS Reservation Optimization for Multi-QoS Level based SaaS Provisioning Hai Dong [0000-0002-7033-5688] School of Science, RMIT University, Melbourne, Australia hai.dong@rmit.edu.au Abstract. We design an evolutionary computation based solution for SaaS providers to optimize the cost of reserved IaaS instance subscription. The services provisioned by SaaS providers have multiple options of Quality of Service (QoS) levels. A Bayesian network model is designed to map between the QoS levels of SaaS instances and their actual resource consumption on IaaS instances. A genetic algorithm with heuristics is designed to optimize the reserved IaaS instance subscription. The proposed solution is evaluated via a set of simulated experiments. Keywords: Evolutionary computation · IaaS instance subscription · Multi-QoS level based SaaS provisioning. 1 Introduction Software as a Service (SaaS) is becoming a popular software delivery model for increasing business applications [1]. In the SaaS model, software applications are installed in IaaS instances for customers to access. Considering the factor of cost, increasing SaaS providers prefer to subscribe IaaS instances instead of building own clouds 1 . IaaS instances can be offered in flexible ways. For example, Amazon EC2 provides three purchasing options to customers – on-demand, reserved, and spot 2 . Among these three options, the reserved option has unique advantages: compared to the on-demand option, the reserved option has up to 75% discount on pricing; compared to the spot option, the reserved option is interruption-free in relative long terms. Those advantages make the reserved instance a better option for SaaS provision that needs a relatively long-term and stable cloud infrastructure support. Most SaaS providers price their products using a monthly or annually subscription fee. An emerging trend is SaaS provisioning is that a SaaS can be provisioned in different QoS levels to fulfil the diverse needs of customers on QoS [2]. The QoS is usually guaranteed in Service Level Agreements (SLAs), by which breaching the promised QoS may occur a penalty and damage the providers’ reputation [15]. As certain QoS parameters (e.g., response time) of SaaS provisioning have strong correlations with the capacity of its underneath IaaS instances [3], SaaS provisioning with different QoS levels simultaneously implies different capacities of IaaS instances. Therefore, to guarantee the QoS of a SaaS provisioning, the appropriate amount of IaaS resources, namely the right types and number of IaaS instances, need to be allocated to the SaaS demands. For SaaS providers, if they can learn about the demands of customers in the next period, a better strategy to minimize their cost is that they only need to subscribe the right type and number of IaaS instances at a proper time before the demands arriving, considering the time delay for virtual machine initialization and application installation. Existing research studying customer-centric cloud resource provisioning focuses on subscription cost minimiza- tion based on customers’ functional (resource) requirements and ignore their non-functional (QoS) requirements. The difficulty lies on mapping between customers’ QoS requirements and provisioned cloud resources. In this pa- per, our goal is to design an optimization strategy enabling SaaS providers to subscribe the appropriate types and number of IaaS instances to meet SaaS customers’ requirements with multiple QoS levels, by exploring the map- ping between the QoS parameters (e.g., response time) of SaaS provisioning and the capacity of subscribed IaaS instances, with the goals of cost minimization and SLA compliance. Evolutionary algorithms, in contrast to traditional optimization algorithms, are more efficient in nonlinear constraint based optimization and more tolerant to noisy observations [4]. Some typical evolutionary algorithms, such as genetic algorithms (GA), particle swarm optimization, art colony optimization, etc., have been widely applied in multiple areas of cloud computing, such as cloud service composition, cloud resource allocation, cloud workload distribution, etc. In this paper, we design a GA based solution to achieve our goal. The major contributions of this research are summarized as follows: 1 http://www.gartner.com/newsroom/id/2923217 2 http://aws.amazon.com/ec2/ ICONIP2019 Proceedings 65 Volume 16, No.1 Australian Journal of Intelligent Information Processing Systems