A dynamic pricing algorithm for a network of virtual resources. Bram Naudts * , Mario Flores † , Rashid Mijumbi ‡ , Sofie Verbrugge * , Joan Serrat † and Didier Colle * * Department of Information Technology Ghent University, Ghent, Belgium Email: {bram.naudts,sofie.verbrugge,didier.colle}@intec.ugent.be † Network Engineering Department Universitat Polit` ecnica de Catalunya, Barcelona, Spain Email: {mario.flores@entel,serrat@tsc}.upc.edu ‡ Telecommunications Software and Systems Group Waterford Institute of Technology, Waterford, Ireland Email: rmijumbi@tssg.org Abstract—A service chain is a combination of network ser- vices (e.g. a firewall) that are interconnected to support an application (e.g. a video-on-demand service). Building a service chain requires a set of specialized hardware devices each of which need to be configured with its own command syntax. By moving management functions out of the hardware into controller software, software-defined networking (SDN) simplifies provisioning and reconfiguration of service chains. By moving the network functions out of dedicated hardware devices into software running on standard x86 servers, network function virtualization (NFV) turns the deployment of a service chain into a more (cost)-efficient and flexible process. In an SDN/NFV- based architecture, those service chains are composed of vir- tual network functions (VNFs) that need to be mapped to physical network components. In literature, several algorithmic approaches exist to do so efficiently and cost-effective. However, once mapped, a simple revenue model is used for pricing the requested substrate resources. This often lead to a loss of revenue for the infrastructure provider. In this paper, we propose a more advanced, dynamic pricing algorithm for pricing the requested substrate resources. The proposed algorithm increases the revenue of the infrastructure provider based on historic data, current infrastructure utilization levels and the pricing of competitors. Our experimental evalu- ation shows that the proposed algorithm increases the revenue of the infrastructure provider significantly, independent of the average network utilization. I. I NTRODUCTION Today, building a service chain to support a new application requires specialized hardware devices (middleboxes). Each of which need to be configured individually via a vendor specific syntax leading to complex configuration, high change for error and high operational expenditures. As application loads vary over the day and often increase over time, building a service chain means overprovisioning of the devices to support the maximum level of demand leading to extra capital expenditures. As service chains are often built to support mul- tiple applications, data sometimes passes through unnecessary network devices or servers. By using SDN and NFV concepts, the effort and time needed to build a service chain to support a new application can be reduced. In a SDN/NFV-approach, a service chain is an abstraction to define high-level services in a generic way. The service is described by a chain of high-level Network Functions (NFs) and pre-defined parameters which is referred to as a Service Graph (SG) [1]. These SGs need to be mapped to the physical infrastructure. As such, an embedding algorithm is needed to determine if the NFs and their connections in the SGs can be mapped to the physical infrastructure (virtual network embedding problem, VNEP). In literate, several algorithmic approaches exist to solve the VNEP. We refer the interested reader to [2] for a survey of VNE algorithms. Initial studies on placement of VNFs and VNF chains in both IP and optical networks are presented in [3], [4], [5], [1], [6], [7]. There are however other, related challenges that receive fewer attention. The authors of [8], for example declared that finding out advanced economic models, instead of the simple revenue model used in the existing literature, for VN pricing is an important research topic that needs further attention. In that field, our work is related to the online pricing literature that deals with instantaneous demand dynamics and the adjustment of prices on the spot. Dynamic pricing has become an active field of the revenue management literature, with successful realworld applications in industries such as travel, fashion, and so on [9], [10], [11]. Closely related to our work, revenue management has also been applied to the field of cloud computing, [12], [13], [14], [15], [16]. For cloud providers, unlike other fields, revenue not only depends on the (unknown) number of customers, but also on the (unknown) duration of usage. As such, not only arrival rates but also service times are stochastic. In those works, resources are however consid- ered as interchangeable. When embedding service graphs, the customer will however typically have a set of requirements (e.g. delay, location, etc.) for a network of resources. Re- sources are therefore hardly interchangeable without harming the expected quality-of-service. To our knowledge, work in this field is limited. The most related work to ours are [17] in which the negotiation process in a multi-domain environment