RealVNF – Improved Coordination of Chained VNFs under Realistic Conditions

Name of the participant: Stefan Schneider

Description of the IT research project: When requesting services like video streaming or live communication, multiple components (e.g., firewalls or video optimizers) are involved and have to process the request. Service function chaining (SFC) connects the involved components into chains that user requests traverse sequentially. Since user demand changes over time, these components are implemented in software as virtual network functions (VNFs) that can be started and stopped on demand to balance the load. To minimize latency and improve user experience, the VNFs of an SFC should run close to the user locations.

The decision of how many VNF instances are required to satisfy the user demand (called “scaling”) and where to run the instances (called “placement”) is typically performed in a reactive manner, i.e., starting/stopping VNF instances after user demand changes. The delay of starting a new VNF instance can be non-negligible (in the order of 10-100 s) such that increased user demand may not be handled properly during this startup time, leading to higher delays. Furthermore, current approaches usually focus on carefully designed models and algorithms by experts, highly tailored to a specific scenario. These models easily fail if the scenario changes and the underlying assumptions are no longer met.

Therefore, this project focuses on predictive and self-driving scaling and placement of VNFs in an SFC, starting new VNF instances shortly before they are actually needed. Such proactive behavior could improve the overall service quality by reducing delays and preventing VNF overload while waiting for new VNF instances to start. To enable such proactive SFC, the project focuses on machine-learning and deep-learning techniques.

The project also considers other aspects of improving service coordination in practice, e.g., highly scalable solutions that work in realistic large-scale networks with rapidly arriving and fluctuating user demand.

To read more about this project, click here.

Software Campus partner: University Paderborn, Huawei
Implementation period: 01.11.2018 – 31.01.2021