Distributed Software Router Management

With the stunning success of the Internet, information and communication technologies diffused increasingly attracting more uses to join the the Internet arsenal which in turn accelerates the traffic growth. This growth rate does not seem to slow down in near future. Networking devices support these traffic growth by offering an ever increasing transmission and switching speed, mostly due to the technological advancement of microelectronics granted by Moore’s Law. However, the comparable growth rate of the Internet and electronic devices suggest that capacity of systems will become a crucial factor in the years ahead. Besides the growth rate challenge that electronic devices face with respect to traffic growth, networking devices have always been characterized by the development of proprietary architectures. This means that incompatible equipment and architectures, especially in terms of configuration and management procedures. The major drawback of such industrial practice, however, is that the devices lack flexibility and programmability which is one of the source of ossification for today’s Internet. Thus scaling or modifying networking devices, particularly routers, for a desired function requires a flexible and programmable devices. Software routers (SRs) based on personal computers (PCs) are among these devices that satisfy the flexibility and programmability criteria. Furthermore, the availability of large number of open-source software for networking applications both for data as well as control plane and the low cost PCs driven by PC-market economy scale make software routers appealing alternative to expensive proprietary networking devices. That is, while software routers have the advantage of being flexible, programmable and low cost, proprietary networking equipments are usually expensive, difficult to extend, program, or otherwise experiment with because they rely on specialized and closed hardware and software. Despite their advantages, however, software routers are not without limitation. The objections to software routers include limited performance, scalability problems and lack of advanced functionality. These limitations arose from the fact that a single server limited by PCI bus width and CPU is given a responsibility to process large amount of packets. Offloading some packet processing tasks performed by the CPU to other processors, such as GPUs of the same PC or external CPUs, is a viable approach to overcome some of these limitations. In line with this, a distributed Multi-Stage Software Router (MSSR) architecture has been proposed in order to overcome both the performance and scalability issues of single PC based software routers. The architecture has three stages: i) a front-end layer-2 load balancers (LBs), open-software or open-hardware based, that act as interfaces to the external networks and distribute IP packets to ii) back-end personal computers (BEPCs), also named back-end routers in this thesis, that provide IP routing functionality, and iii) an interconnection network, based on Ethernet switches, that connects the two stages. Performance scaling of the architecture is achieved by increasing the redundancy of the routing functionality stage where multiple servers are given a coordinated task of routing packets. The scalability problem related to number of interfaces per PC is also tackled in MSSR by bundling two or more PCs’ interfaces through a switch at the front-end stage. The overall architecture is controlled and managed by a control entity named Virtual Control Processor (virtualCP), which runs on a selected back-end router, through a DIST protocol. This entity is also responsible to hide the internal details of the multistage software router architecture such that the whole architecture appear to external network devices as a single device. However, building a flexible and scalable high-performance MSSR architecture requires large number of independently, but coordinately, running internal components. As the number of internal devices increase so does the architecture control and management complexity. In addition, redundant components to scale performance means power wastage at low loads. These challenges have to be addressed in making the multistage software router a functional and competent network device. Consequently, the contribution of this thesis is to develop an MSSR centralized management system that deals with these challenges. The management system has two broadly classified sub-systems: I) power management: a module responsible to address the energy inefficiency in multistage software router architecture II) unified information management: a module responsible to create a unified management information base such that the distributed multistage router architecture appears as a single device to external network from management information perspective. The distributed multistage router power management module tries to minimize the energy consumption of the architecture by resizing the architecture to the traffic demand. During low load periods only few components, especially that of routing functionality stage, are required to readily give a service. Thus it is wise to device a mechanism that puts idle components to low power mode to save energy during low load periods. In this thesis an optimal and two heuristic algorithms, namely on-line and off-line, are proposed to adapt the architecture to an input load demand. We demonstrate that the optimal algorithm, besides having scalability issue, is an off-line approach that introduce service disruption and delay during the architecture reconfiguration period. In solving these issues, heuristic solutions are proposed and their performance is measured against the optimal solution. Results show that the algorithms fairly approximate the optimal solution and use of these algorithms save up to 57.44% of the total architecture energy consumption during low load periods. The on-line algorithms are superior among the heuristic solutions as it has the advantage of being less disruptive and has minimal service delay. Furthermore, the thesis shows that the proposed algorithms will be more efficient if the architecture is designed keeping in mind energy as one of the design parameter. In achieving this goal three different approaches to design an MSSR architecture are proposed and their energy saving efficient is evaluated both with respect to the optimal solution and other similar cluster design approaches. The multistage software router is unique from a single device as it is composed of independently running components. This means that the MSSR management information is distributed in the architecture since individual components register their own management information. It is said, however, that the MSSR internal devices work cooperatively to appear as a single network device to the external network. The MSSR architecture, as a single device, therefore requires its own management information base which is built from the management information bases dispersed among internal components. This thesis proposes a mechanism to collect and organize this distributed management information and create a single management information base representing the whole architecture. Accordingly existing SNMP management communication model has been modified to fit to distributed multi-stage router architecture and a possible management architecture is proposed. In compiling the management information, different schemes has been adopted to deal with different SNMP management information variables. Scalability analysis shows that proposed management system scales well and does not pose a threat to the overall architecture scalability.