MAPLE: A Machine learning Approach for Efficient Placement and Adjustment of Virtual Network Functions

摘要

As one of the many advantages of cloud computing, Network Function Virtualization (NFV) has revolutionized the network and telecommunication industry through enabling the migration of network functions from expensive dedicated hardware to software-defined components that run in the form of Virtual Network Functions (VNFs). However, with NFV comes numerous challenges related mainly to the complexity of deploying and adjusting VNFs in the physical networks, owing to the huge number of nodes and links in today's datacenters, and the inter-dependency among VNFs forming a certain network service. Several contributions have been made in an attempt to answer these challenges, where most of the existing solutions focus on the static placement of VNFs and overlook the dynamic aspect of the problem, which arises mainly due to the ever-changing resource availability in the cloud datacenters and the continuous mobility of the users. Few attempts have been lately made to incorporate the dynamic aspect to the VNF deployment solutions. The main problem of these approaches lies in their reactive readjustment scheme which determines the placement/migration strategy upon the receipt of a new request or the happening of a certain event, thus resulting in high setup latencies. In this paper, we take advantage of machine learning to reduce the complexity of the placement and readjustment processes through designing a cluster-based proactive solution. The solution consists of (1) an Integer Linear Programming (ILP) model that considers a tradeoff between the minimization of the latency, Service-Level Objective (SLO) violation cost, hardware utilization, and VNF readjustment cost, (2) an optimized k-medoids clustering approach which proactively partitions the substrate network into a set of disjoint on-demand clusters and (3) data-driven cluster-based placement and readjustment algorithms that capitalize on machine learning to intelligently eliminate some cost functions from the optimization problem to boost its feasibility in large-scale networks. Simulation results show that the proposed solution considerably reduces the readjustment time and decrease the hardware utilization compared to the K-means, original k-medoids and migration without clustering approaches.