Real-time Data Analytics for Cloud Network Management (REALM)

Purpose and goal

We devised efficient methods for real-time prediction of service metrics from device statistics using machine-learning techniques. These methods are building blocks for novel cloud management functions such as quality assurance. We performed software development and extensive experimentation on cloud testbeds at KTH and at Ericsson Research. We showed that (client-side) service metrics for a cluster-based VoD service and a key-value store can be predicted with high accuracy. We demonstrated a REALM prototype to stakeholders and research audiences.

Results and expected effects

A key achievement of REALM has been the establishment of a strong stakeholder network within Ericsson. REALM has leveraged this network to communicate its vision and technology approach for cloud service assurance through analytics. REALM made contributions to the development of products in the intersection of cloud management and analytics. The concept of service assurance through machine learning translated well to functional requirements. Two patent applications have been filed.

Approach and implementation

We investigated linear and non-linear methods for metrics prediction. The best results were achieved with tree-based methods. We further studied online algorithms for real-time prediction and experimented with automated feature reduction techniques. We built a real-time analytics engine, which reads in live statistics from devices and performs prediction. We developed concepts to predict violations of service-level objectives (SLOs) on a telecom cloud. In order to allow our method to scale, we investigated a distributed learning scheme for predicting violations of SLOs.