This chapter considers circuits between boundary routers. If we need to increase the capacity of an existing circuit, it might be that the current circuit path cannot accommodate this increase, while an alternate path can. In this case, one option is to reroute the whole circuit through a path that has the required capacity (if there is one), but this option might be too costly in terms of signaling and resource consumption. One alternative is to create a separate circuit with a capacity equal to the additional capacity that is needed. However, one problem is that this parallel circuit will have a propagation latency that is different from the original path. If data is injected into the combined circuit, it may happen that a packet is split into two parts that travel through different paths, and so a complex realignment buffer will be required at the egress point to realign the two parts of the packet. Such a mechanism has already been proposed for SONET/SDH, and it is known as virtual concatenation of channels [46,166].
One way of eliminating this realignment is to avoid splitting packets over parallel paths. Packets can then be recovered integrally at the tail end of each backbone circuit and injected directly into the packet-switched part of the network. This method can create some packet reordering within a user flow, which TCP may interpret as packet drops due to congestion. Yet, reordering would be rare if the difference in propagation delays between the parallel paths is smaller than the interarrival time imposed by the access link to consecutive packets of the same flow (for 1500-byte packets, it is 214 ms for 56-Kbit/s access links, and 8 ms for 1.5-Mbit/s access links). One possible solution is to equalize the delay using a fixed-size buffer at the end of one of the sub-circuits. However, this buffer may not be necessary because, as reported recently [17], TCP is not significantly affected by occasional packet reordering in a network.
It should be pointed out that the definition used here for circuit overflow is rather strict, and it represents an upper bound on the packet drop rate. In general, the ingress boundary router will have buffers at the head end of each backbone circuit, which will absorb short fluctuations in the flow rate between boundary routers. For this reason, in the measurements in Figures 5.5 and 5.10, all single-packet flows were ignored. The buffer at the head end will also allow the system to achieve some statistical multiplexing between active flows; something that TCP Switching in Chapter 4 could not achieve. However, as mentioned in Chapter 3, this statistical multiplexing will not necessarily lead to a smaller response time because the flow peak rate will still be capped by the access link.
The approach presented in this chapter does not specify any signaling mechanism and does not impose any requirements on it. One could use existing mechanisms such as the ones envisioned by GMPLS [7] or OIF [13], which will be described in Chapter 6. This method can also be used in conjunction with TCP Switching to control an optical backbone with an electronic outer core and an optical inner core. TCP Switching would control the outer fine-grain electronic circuit switches and would provide the information that is used to control the inner coarse-grain optical circuit switches.