Satellite Switching with Neural Network Techniques for DiffServ Policy

Due to wide coverage and asymmetric bandwidth, satellite communications are becoming an important part of global multimedia networks. Actual trend in global communications is to integrate satellite systems in the development of Internet [1]. A lot of services are being created exploiting wide diffusion of Internet, however results are not always acceptable due to different requirements of each service. In particular each network-based application has different requirements in terms of delay, jitter, bandwidth and packet loss. Nowadays, Internet is a Best Effort network and each packet is processed as quickly as possible, without any guarantee on delivery time. The delay introduced by each network node has also a high variability. The expansion of Internet to a commercial infrastructure requires the development of new protocols, architectures and network systems in order to guarantee different Quality of Service (QoS) levels to various applications according to their needs. The introduction of QoS-driven policy in a satellite system requires the development of advanced solutions in terms of complexity and cost due to reduced on-board capabilities. The Internet Engineering Task Force (IETF) has proposed many service models and mechanisms in order to meet the demand for QoS; in this paper we consider the Differentiated Services (DiffServ) approach [2] that is scalable and keep low complexity core nodes, like satellites, moving complexity to the edge nodes. This approach uses the DS Code Point field of the IPvX packet header in order to identify the payload type. Each node manages the packets according to this field; in DiffServ systems this operation is known as PHB (Per Hop Behaviour). The satellite system considered in this paper is geostationary (GEO) and multibeam. For our purpose, the orbit type modify the average delay and jitter without any change to the structure of the switch. The presence of more than one input beam (uplink) and one output beam (downlink) requires procedures to avoid packet collision; at the same time we have implemented a mechanism to supply QoS. Previous techniques are able to have a complete control over switching matrices hardware-based without introducing QoS requirements in solving the switching problem [3]. Our approach is to use neural networks, for their flexibility and easy adaptive structure that let introduce QoS constrains without any change to the basic structure of the neural controller.