No abstract

--This paper studies the performance of anInternet that uses circuit switching instead of, or in addition to, packet switching. On the face of it, this would seem a pointless exercise; the Internet is packet switched, and was deliberately built that way to enable the efficiencies afforded by statistical multiplexing, and the robustness of fast re-routing around failures. But link utilization is low, and falling, particularly at the core of the Internet, which means that statistical multiplexing is less important than it once was. And circuit switches today are capable of rapid reconfiguration around failures. There is also renewed interest in circuit switching because of the ease of building very high capacity optical circuit switches. While several proposals have suggested ways in which circuit switching may be introduced into the Internet, this paper is based on TCP Switching, in which a new circuit is created for each application flow. Here, we explore the performance of a network that uses TCP Switching, with particular emphasis on the reponse time experienced by users. We use simple M/G/1 and M/G/N queues to model application flows in both packet switched and circuit switched networks, as well as ns-2 simulations. We conclude that because of high bandwidth long-lived flows, it does not make sense to use circuit switching in shared access or local area networks. But our results suggest that in the core of the network, where high capacity is needed most, and where there are many flows per link, there is little or no difference in performance between circuit switching and packet switching. Given that circuit switches can be built to be much faster than packet switches, this suggests that a circuit switched core warrants further investigation.

Circuit switches have simpler data paths and are potentially much faster than packet switches. Taking advantage of this difference makes very high capacity all-optical circuit switches feasible, whereas all-optical packet switches are a long way from commercial practicality. Peak-allocation, that is eliminating the benefits of statistical multiplexing, is circuit switches' main disadvantage, and what prevents their widespread adoption. However, the rising amount of already abundant link capacity will eliminate this drawback. Our research focuses on how the existing IP infrastructure can incorporate fast, simple (and perhaps optical) circuit switches. Several approaches to this already exist, 1-5 but we propose a technique called transmissioncontrol protocol (TCP) switching in which each application flow (usually an individual TCP connection) triggers its own end-to-end circuit creation across a circuit switched core. Based on IP switching, 6 TCP switching incorporates modified circuit switches that use existing IP routing protocols to establish circuits. Routing occurs hop by hop, and circuit maintenance uses soft state, that is, it is removed through an inactivity timeout.Many believe that only routers, links, and end hosts, all using packet switching, comprise the Internet. In reality, the Internet uses circuit switching, both at its core (Sonet, SDH, dense wave digital multiplexing or DWDM), and in its last mile (modems, DSL). Internet Protocol treats these circuits as static, point-to-point links connecting adjacent nodes; the physical circuits and IP belong to different layers, and they are completely decoupled since they operate autonomously and without cooperation. Decoupling of layers has many advantages. It lets the circuit switched physical layer evolve independently of IP-to both Sonet/SDH, and DWDM. IP runs over a huge variety of physical layers regardless of the underlying technology. However, a lot of repetition exists between the packet-switched IP layer and the circuitswitched physical layer. For example, a network must route both IP datagrams and circuit paths, yet, they use different routing protocols, and their implementations are incompatible. This makes simple and obvious operations infeasible. For example, if traffic increases between two neighboring routers, no simple or standard way exists to automatically increase the capacity between them. The problem is with how circuit switches interact with IP routers. TCP switching presents a method of interaction, promising automatic and dynamic circuit allocation.

Abstract-While it is technically pleasing to believe that IP will dominate all forms of communication, our delight in its elegance is making us overlook its shortcomings. IP is an excellent means to exchange data, which explains its success. It remains ill-suited as a means to provide many other types of service; and is too crude to form the transport infrastructure in its own right. To allow the continued success of IP, we must be open-minded to it living alongside, and merging with, other techniques (such as circuit switching) and protocols that are optimized to different needs. I. INTRODUCTIONWhatever the initial goals of the Internet, there are two main characteristics that seem to account for its success: Reachability and Heterogeneity. IP provides a simple, single, global address to reach every host, enables unfettered access between all hosts, and adapts the topology to restore reachability when links and routers fail. IP hides heterogeneity in the sense that it provides a single, simple service abstraction that is largely independent of the physical links over which it runs. As a result, IP provides service to a huge variety of applications and operates over extremely diverse link technologies.The growth and success of IP has given rise to some widely held assumptions amongst researchers, the networking industry and the public at large. One common assumption is that it is only a matter of time before IP becomes the sole global communication infrastructure, dwarfing and eventually displacing existing communication infrastructures such as telephone, cable and TV networks. IP is already universally used for data networking in wired networks, and is being rapidly adopted for data communications in wireless and mobile networks. IP is increasingly used for both local and long-distance voice communications, and it is technically feasible for packetswitched IP to replace SONET/SDH.A related assumption is that IP Routers (based on packet-switching and datagram routing) will become the most important, or perhaps only, type of switching device inside the network. This is based on our collective belief that packet-switching is inherently superior to circuit switching because of the efficiencies of statistical multiplexing, and the ability of IP to route around failures. It is widely assumed that IP is simpler than circuit switching, and should be more economical to deploy and manage. And with continued advances in the underlying technology, we will no doubt see faster and faster links and routers throughout the Internet infrastructure. It is also widely assumed that IP will become the common convergence layer for all communication infrastructures. All communication services will be built on top of IP technology. In addition to information retrieval, we will stream video and audio, place phone calls, hold video-conferences, teach classes, and perform surgery.On the face of it, these assumptions are quite reasonable. Technically, IP is flexible enough to support all communication needs, from best-effort to real-time. ...

Circuit switches have simpler data paths and are potentially much faster than packet switches. Taking advantage of this difference makes very high capacity all-optical circuit switches feasible, whereas all-optical packet switches are a long way from commercial practicality. Peak-allocation, that is eliminating the benefits of statistical multiplexing, is circuit switches' main disadvantage, and what prevents their widespread adoption. However, the rising amount of already abundant link capacity will eliminate this drawback. Our research focuses on how the existing IP infrastructure can incorporate fast, simple (and perhaps optical) circuit switches. Several approaches to this already exist, 1-5 but we propose a technique called transmissioncontrol protocol (TCP) switching in which each application flow (usually an individual TCP connection) triggers its own end-to-end circuit creation across a circuit switched core. Based on IP switching, 6 TCP switching incorporates modified circuit switches that use existing IP routing protocols to establish circuits. Routing occurs hop by hop, and circuit maintenance uses soft state, that is, it is removed through an inactivity timeout.Many believe that only routers, links, and end hosts, all using packet switching, comprise the Internet. In reality, the Internet uses circuit switching, both at its core (Sonet, SDH, dense wave digital multiplexing or DWDM), and in its last mile (modems, DSL). Internet Protocol treats these circuits as static, point-to-point links connecting adjacent nodes; the physical circuits and IP belong to different layers, and they are completely decoupled since they operate autonomously and without cooperation. Decoupling of layers has many advantages. It lets the circuit switched physical layer evolve independently of IP-to both Sonet/SDH, and DWDM. IP runs over a huge variety of physical layers regardless of the underlying technology. However, a lot of repetition exists between the packet-switched IP layer and the circuitswitched physical layer. For example, a network must route both IP datagrams and circuit paths, yet, they use different routing protocols, and their implementations are incompatible. This makes simple and obvious operations infeasible. For example, if traffic increases between two neighboring routers, no simple or standard way exists to automatically increase the capacity between them. The problem is with how circuit switches interact with IP routers. TCP switching presents a method of interaction, promising automatic and dynamic circuit allocation.