Organic Network Control: Turning Standard Protocols into Evolving Systems

Introduction

The development of networking during the past decades shows that the number of protocols being proposed for different applications on all layers has increased largely. This is partly due to the ubiquitous availability of networked devices for a wide spectrum of applications and ranges from classical desktops and servers in wired networks to small handheld and embedded devices in wireless networks, such as mobile phones and sensor nodes. This development is accompanied by the users’ requirements of a better convergence of all these networks and applications.

Most protocols offer a large number of parameters that allow for adapting them to different usage scenarios, e.g. they allow for changing settings for timeouts, number of nodes to connect with, and retransmission counters. However, these parameters are seldom changed at runtime. Instead, they are mostly investigated and set at design time or – at best – changed manually at runtime. This leads to a rather static configuration even though the situation in the network is constantly changing. These dynamics are not only characterized by changes in, for example, available bandwidth, network topology, and channel quality over time. Additionally, new applications – and protocols respectively – are introduced. The class of Peer-to-Peer applications is probably one of the most dynamic classes of applications contributing to the changes in the protocol landscape during the past years. In essence, well established protocols, e.g., for Web traffic, have to co-exist with protocols that no one would have thought of some years back in the first place.

During the past years, the Organic Computing (OC) Initiative (Schmeck, 2005) has proposed a number of techniques, architectures, and algorithms that support the development of complex systems. One of the key ideas is that the complexity of (current and) future systems, such as the Internet, does not allow for a design-time-only approach when it comes to, for example, testing and optimization. Instead, the OC approach provides means for building systems that adapt and improve at runtime, using for example machine learning techniques. This also includes the seamless integration of new components (like protocols) into an existing system and is frequently referred to as Self-Organization.

In this chapter we present the Organic Network Control (ONC) system (Tomforde et al., 2009b), a three-layered Observer/Controller architecture that allows for “wrapping” existing protocols into a framework which enables a large degree of Self-Organization in existing networks. The architecture has a generic character: it has also been applied to other scenarios like, e.g., vehicular traffic control (Prothmann et al., 2009). The process to apply this to other domains is similar to adapting the ONC framework to network protocols.

The “wrapping” is achieved by adapting the numerous parameters of existing protocols at runtime and in dependence of the current status of the network node, e.g. a router or a user’s PC, running the ONC system. As a basis, the so-called Observer collects information locally available at the network node, like available bandwidth and CPU resources. In turn, the so-called Controller uses this information and a performance measure for the protocol “under observation and control” to evaluate its current performance. If knowledge about similar situations is available from the past or if the performance falls below a threshold, the ONC system applies changes to the configuration of the protocol. In the case of observing a drop of performance below a given threshold, the ONC system makes use of Learning Classifier Systems (LCS) (Wilson, 1995) and Evolutionary Algorithms (EA) (Eiben & Smith, 2003) to find and evaluate better parameter sets. This process does not require any knowledge about the internal behavior of the protocol and takes place at runtime (although it may be supplemented with knowledge acquired during design time).