Maintaining and Forecasting SmartX Multi-View Spatio-temporal Visibility for Cloud-native Edge Boxes in Multisite Cloud Playground

Abstract

The decreasing cost and power consumption of intelligent, interconnected, and interactive devices at the edge of the internet are creating massive opportunities for emerging technology paradigms such as cloud, software-defined networking (SDN), and internet of things (IoT) to improve efficiency, reliability and productivity. These ICT infrastructure transformations, however are causing new sets of operational problems for operators, as a consequence of inadequate end-to-end visibility and visualization support about underlay infrastructure. Particularly for multi-tenant and multi-site underlay infrastructure, visibility (measurement data and data for tracing) is essential for the effective operation to easily pinpoint server faults, network bottlenecks, and application performance troubles. In this regards, we propose a concept of efficient methods for persistently measure and trace multi-layer data (underlay, physical, and virtual resource-layers and flow layer) and continuously operated by tiny-sized DevOps-style team. The aim is to leverage the existing SmartX MultiView Framework (MVF) for collecting a limited spatio-temporal (record both time and location) visibility in a defined summarized format. The managed summarised collection is leveraged as input data for preliminary Deep learning (Models) as a prototype for predicting flow-layer visibility in a restrict environment. The proposed solution is deployed on multi-sites cloud-native based edge boxes deployed at OF@TEIN+ as Open and Federated Future Internet (SDN/NFV/Cloud-integrated) playground (i.e., miniaturized testbed). It has centralized center inside the playground tower, to control and monitor distributed hyper-converged box-style site resources. The main objectives of the work is to optimize operational continuity with minimum data loss in order to maintain and sustain services. We present an introductory concept of flow-layer traffic prediction to train a Long Short-Term Memory (LSTM) neural network using multivariate features. This aim of forecasting to fairly evaluate the performance of the proposed LSTM models on our dataset. Finally, a unified storage engine is followed by effective visualization scheme such as real-time multi-layer time-series visualization, Onion ring visualization and unified Data Summary report to provide effective insights to Operators.

Based on verification and measurement results, our proposed showed minimal visibility data loss and recovery during short-term network connectivity outages (i.e. 99.9% visibility collection), and reduce the storage requirements up to 80%while maintaining high accuracy . The recorded accuracy in the testing process for both LSTM and Seq-to-seq LSTM MNN is 91.48% and 92.51%, respectively.