Network utilization optimizer for SD-WAN

摘要

the question how to use the maximum of network possibilities is still open. ISPs and large distributed companies use only 40-50% of total network bandwidth. Technology that helps to increase network bandwidth utilization and redundancy is crucial and there is still no generic and simple solution for this problem. The SDN architecture advocates the separation of data and control plane, and helps to simplify the network management and maintenance due to logically centralized software. Basing on this approach, our team has implemented a simple solution solving network utilization issue. Remote SDN controller runs on high performance server and this enables to apply relatively complex per-flow global routing algorithms. An application tracks network state. In case of link fault, the flows affected by outage are re-routed over alternative path. A whole network acts as the single distributed L2 switch from external connections perspective, but solution architecture allows to change a whole network representation from L2 switch to distributed L3 router. Application was developed by using OpenFlow technologies at data plane devices. The application uses modified Dijkstra algorithm. The algorithm searches for the route with the best spare capacity based on actual network utilization. Also the algorithm allows to control route length over per-hop penalty. So the developed application allows to apply per-flow policing in terms of bandwidth and latency. Nowadays OpenFlow controllers don't have a standardized API and it makes it impossible to change a controller for your application. To avoid this issue an OpenFlow independent Controller-Application specific interface has been developed. Interface uses application specific proprietary message format optimized to increase configuration performance. So our application is flexible in choosing OpenFlow controller. Characteristics for our prototype have been defined based on performance characteristics of Yarnet ISP located in Yaro- lavl and it should work with 30 nodes (each node has at least 3 connections per switch) and establish 5000 flows per second and has traffic outage less than 1 second. The characteristics were measured using simulated and target test environment. Developed application will be used as the framework to implement traffic policing features, QoS, bandwidth and latency reservation.