TCP Switching exploits the fact that most of our communications are connection oriented and reliable. Rather than using complex signaling mechanisms to create and destroy circuits, TCP Switching piggybacks on existing end-to-end mechanisms to manage circuits. More specifically, TCP Switching uses the initial handshake of the most common type of flows, TCP connections, to create a circuit in the core of the network. When a circuit request message gets dropped, TCP Switching relies on the TCP retransmission mechanisms to set up the circuit again at some later time.
In addition, TCP Switching tries to exploit some of the statistical multiplexing that exists among flows. Obviously, it does not achieve the statistical multiplexing of packet switching because, if a flow does not fully utilize its circuit capacity, this unused bandwidth is wasted. However, TCP Switching does not reserve resources for inactive flows, and so those resources can be employed by other active flows. In this way, TCP Switching achieves some statistical multiplexing gain.
TCP Switching is indeed an extreme technology. In what follows, I will discuss some of the concerns that arise when this approach is described; namely, I will discuss the impact of single-packet flows, bandwidth inefficiencies and denial of service.
The longer flows are, the more efficient TCP Switching because the circuit setup cost is amortized over the longer data transfer. It is unclear how flow sizes will evolve in the future. On one hand, there is a trend for longer flows from downloads and streaming of songs and video. On the other hand, traffic from sensors is likely to consist of very short exchanges, perhaps consisting of single-packet flows. Even though most probably those single-packet flows will be aggregated before being sent through the Internet backbone [2], it is worth asking what would happen if single-packet flows became a large fraction of Internet traffic.
As mentioned in Section 4.3.3, packet switching can be considered to be a special case of circuit switching in which all flows are 1-packet long. The processing and forwarding of a circuit request in the control plane of a circuit switch is similar to the forwarding of a packet in the data plane of a router. When a circuit arrives, a next-hop lookup has to be performed and resources have to be checked. If these resources are available, the crossconnect needs to be scheduled; otherwise, the request has to be buffered or dropped. The only difference with packet switching is that state is maintained so that next time data arrives for that circuit, the data path can forward the information without consulting the control plane.
In TCP Switching, single-packet flows are forwarded as if using packet switching by the control plane, while long flows are forwarded by the data path of circuit switching at a much higher rate. In order to avoid interactions in the control plane between these two classes of flows, one can create two separate queues and process them differently (e.g., the single-packet flows would not write any state).
In TCP Switching, the wastage of bandwidth is evident because it suffers from acute fragmentation; the bandwidth allocated to a circuit is reserved for its associated flow, and if the flow does not use it, the bandwidth is wasted. Nevertheless, it is not clear to what extent the bandwidth inefficiency is a problem because, as shown in Figure 1.3, optical-link capacity does not have the technology limitations of buffering and electronic processing. One should then ask the question, how much speedup is needed in a circuit switch to compensate for the wasted bandwidth?
In order to quantify how much bandwidth remains unused, we need to look at the time diagram of a typical circuit, as shown in Figure 4.7. As we can see, during the lifetime of a TCP-Switching circuit, there are three phases when bandwidth is wasted by a TCP flow: (1) during the slow start phase, when the source has not yet found the available bandwidth in the circuit, (2) during the congestion avoidance phase, and (3) during the inactivity period that is used to timeout and destroy the circuit. The total amount of bandwidth that is wasted in each phase will depend on the source activity, the flow length, the round-trip time, and the inactivity timeout. For example, the so-called TCP ``mice'' or ``dragonflies'' [23] are so short that they do not enter the congestion avoidance phase.
Given that application flow sizes are typically shorter than 10s, an inactivity timeout of 60s such as the one proposed in [129,74] is extremely wasteful. Better efficiencies can be achieved with a timeout value that is a little larger than the RTT (as proposed in Section 4.3.3) because this timeout value is comparable to the duration of the slow-start phase, which lasts a few RTTs.
In any case, traffic is highly asymmetric. Usually, one end host holds the information, and the other simply downloads it, such as with web browsing. This means that one direction of the connection will be filling up the pipe with large packets (typically of 1500 bytes), whereas the other direction will be sending 40-Byte long acknowledgements. If the two circuits belonging to a bi-directional flow are symmetric, then even if we achieved a bandwidth efficiency of close to 100% in the direction of the download, the reverse direction will get an efficiency of less than 2.7%. The overall efficiency would be only 51%. However, the direction of download is not uniformly distributed, as servers tend to be placed in PoPs and co-location facilities, and so the direction in which the bottleneck occurs will get a bandwidth efficiency closer to 100% than to 2.7%.4.6 If the bandwidth inefficiency of the reverse circuit proved to be critical, one could allocate less bandwidth to the return channel for the acknowledgements.
Denial of service is an important concern for TCP Switching. With only a few well-crafted packets one can reserve a huge amount of bandwidth, preventing others from using it. This problem is not new, and it is common to other systems that do resource reservation. Two solutions are possible here: one is to use external economic incentives and penalties to deter a user from taking more resources than he/she needs. The other is to restrict the maximum number of simultaneous flows that an ingress boundary router may accept from a single user.
On the other hand, one of the advantages of TCP Switching is that circuits are reserved exclusively for one flow, so, contrary to packet-switched networks, it is easy to track a circuit back to its source, and it is virtually impossible for others to spoof a circuit or to hijack it without the cooperation of a switch. This inherent authentication makes the enforcement of policies across domains easier than in the current Internet.