Auto-3P: An autonomous VNF performance prediction & placement framework based on machine learning

Abstract

We propose Auto-3P, an Autonomous module for Virtual Network Functions Performance Prediction and Placement at network cloud and edge facilities based on Machine Learning (ML). Auto-3P augments the autonomous placement capabilities of MANagement and Orchestration frameworks (MANOs) by considering both resource availability at hosting nodes and the implied impact of a VNF node placement decisions on the whole service level end-to-end performance. Unlike that, most existing placement methods take a rather myopic approach after manual rule-based decisions, and/or based exclusively on a host-centric view that focuses merely on node-local resource availability and network metrics. We evaluate and validate Auto-3P with real-field trials in the context of a well-defined Smart City Safety use case using a real end-to-end application over a real city-based testbed. We meticulously conduct repeated tests to assess (i) the accuracy of our adopted prediction models; and (ii) their placement performance against three other existing MANO approaches, namely, a “Traditional”, a “Latency-aware” and a “Random” one, as well as against a collection of well-known Time Series Forecasting (TSF) methods. Our results show that the accuracy of our ML models outperforms the one by TSF models, with the most prominent accuracy performances being exhibited by models such as K-Nearest Neighbors Regression (K-NNR), Decision Tree (DT), and Support Vector Regression (SVR). What is more, the resulted end-to-end service level performance of our approach outperforms “Traditional”, “Latency-aware”, and Random MANO placement. Last, Auto-3P achieves load balancing at selected VNF hosts without degrading end-to-end service level delay, and without a need for a (fixed) overload threshold check, unlike what is suggested by other works in the literature for coping with heavy system-wide load conditions.

Introduction

Nowadays networks are increasingly becoming largely programmable and software-driven, following the trend for adopting the concepts of Network Function Virtualization (NFV) and Software Defined Networking (SDN). NFV and SDN demonstrate their capability to satisfy the much-needed reliability, low latency, and availability in many application fields such as industrial automation [1] and even beyond that. This signifies an important paradigm shift towards Network Softwarization (NS) and leveraging autonomous network and service management, for which both NFV and SDN are key enablers.

But in order to leverage the evident benefits of a fully automated networking paradigm, there is a series of main challenges that have been raised and need to be addressed. Important examples include autonomous self-deploying of Virtual Network Functions (VNFs), VNF self-destruction (i.e., automatic termination), and self-healing (i.e., automatic troubleshooting), just to name a few. In this context, identifying the appropriate network locations for VNFs is a key optimization problem that seeks for a solution. Existing Management and Orchestration frameworks (MANO), such as the OSM MANO [2], apply mechanisms to instantiate VNFs on hosts with sufficient resources (e.g., number of Central Processing Unit (CPU) cores, amount of memory and storage), which nonetheless disregard important performance factors like latency and host load.

This problem is well-known and gained further attention due to the 5G technology expectations stated by the International Telecommunication Union (ITU)¹ with respect to Ultra Reliable Low Latency Communications (URLLC). As a result, a significant amount of effort has been put lately on optimizing VNF placement at the network’s edge with the goal of reducing application and service delay. A key enabler concept for this is Multi-access Edge Computing (MEC), which allows to move much of cloud computing functionality to the edge of the network, hence processing traffic closer to users instead of forwarding traffic and/or traffic information to cloud. This not only mitigates cloud congestion but also causes significant latency reduction for end-to-end services.

Nonetheless, as there are no information suggesting the optimum location for delay-sensitive VNFs, corresponding placement decisions to either the cloud or MEC facilities are still dominated by manual operations. However, the outcome of manual actions is questionable because they come without any satisfactory placement optimization guarantee. Furthermore, achieving accurate VNF performance predictions at different MEC nodes poses a significant contribution to realising the Zero-touch network and Service Management (ZSM)² [3] concept.

The prediction of Predicting post-placement VNF performance constitutes a complex task that comes with challenges. For end-to-end delay-critical services, predictions depend on multiple stochastic variables that refer to devices and cloud/edge components such as load and latency. To the best of our knowledge, there is little progress in the related literature [4] on forecasting (i) VNF performance in the broader context of service instantiation as well as on (ii) autonomous placement of stand-alone VNFs., hence giving rise to the following questions:

• “How can one predict end-to-end application-level VNF performance prior to appointing a VNF’s host location?

• Assuming such a performance prediction, how can a VNF then be autonomously deployed?”

To address these questions, we propose “Auto-3P”, a solution based on Machine Learning (ML) models that make accurate predictions of the Total Response Time (T_total) for a VNF running an end-to-end service, thus augmenting existing MANO frameworks by autonomously placing VNFs at hosting locations. As we elaborate further in Section 2, Auto-3P (i) provides an autonomous placement mechanism to any MANO frameworks by (ii) applying ML models that suggest the best VNF location.

We assess the performance of Auto-3P through an extensive experimental campaign showing that our solution can deliver better end-to-end service performance based on autonomous VNF placement decisions. We also compare Auto-3P against existing approaches, showing that in most of the cases considered, Auto-3P places VNFs at nodes that can yield optimum performance.

The contributions of this paper are summarized as follows.

• A novel approach to VNF performance prediction and to autonomous VNF placement:The novelty of our work lies in augmenting MANO VNF placement decisions with accurate performance predictions based on specially trained ML models. What is more, our approach considers the impact of VNF placement decisions at a service level, end-to-end. Moreover, we try to address an existing gap in well-established orchestrators like OSM or Openstack [5]. These orchestrators may pose a powerful placement mechanism, but they do not provide native support for ML modules. Last, only a few ML research efforts aside Auto-3P have focused on the prediction of service level end-to-end network or processing delay. Exploring this direction has enough potential for improving VNF placement, especially for delay-sensitive use cases.

• Real testbed and use case based experimentation:Unlike many previous studies that are simulation-based, the current one is experimental-based, hence demonstrating the capability of deploying a solution like Auto-3P in real-life networks under realistic conditions. To do so, we train and test our prediction models using (i) real data from a (ii) real VNF video transcoder in the context of a (iii) real Smart City Safety use case and finally, in a (iv) real deployment environment. As a result, we bridge an important gap in a predominantly simulation-based literature on VNF placement [6], [7], [8], [9], [10], [11].

• Comprehensive comparative study of VNF prediction models and corresponding placement decisions:Our work integrates VNF performance predictions and placement decisions. As such, we study and compare different prediction performances by different ML algorithms as well as by non-ML models. Moreover, we study the placement efficiency of MANO when augmented with Auto-3P against a Traditional, a Latency-aware, and a Random MANO placement method. Our evaluation results denote that Auto-3P significantly improves MANO placement decisions concerning both balancing system-wide load and decreasing end-to-end service level delay. These findings contribute towards the ultimate goal of ZSM in 5G regarding delay-sensitive services and applications.

This article is structured as follows. Section 2 provides the reader with a necessary background on MANO systems, defines the problem targeted by this article and analyses the related work on ML-based VNF management. Section 3 describes the prediction model including a detailed discussion on the autonomous VNF placement module. Section 4 describes in detail our experimental setup and scenarios. Section 5 presents our evaluation results. Finally, we draw our conclusions and discuss our future work plans in Section 6.