... buffers^1.1

In practice, routers are built to handle congestion for a much lesser period of time, needing fewer buffers, but increasing the packet loss probability during overload.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 1997.^1.2

This trend could be broken by the sudden adoption of a new bandwidth-intensive application, such as video streaming, similar to the period of 1995-6 when the massive adoption of the web made traffic double every 3-4 months.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... years^1.3

Experts believe that this trend will slow down as microprocessor and ASIC technologies gradually move from the current two-year cycle to a three-year node cycle after 2004.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... space.^1.4

Actually they will occupy more than five times the space, as many of the routers linecards will be used to connect to other routers within the same central office, rather than to routers in other locations.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... linecard.^1.5

Needless to say, this vision of the future could completely change if a breakthrough in technology made fast, high-density optical memories possible.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... information^1.6

today, over 65% of the traffic is web browsing and peer-to-peer file sharing [31].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... infrastructure^2.1

In this chapter, transport is used in the sense of the infrastructure over which many service networks run, not in the sense of the OSI protocol layer.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... (FDM),^2.2

(Dense) Wavelength Division Multiplexing, (D)WDM, is a subclass of FDM that uses optical wavelengths as channels.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... flow.^2.3

Note that the source and the destination need not be edge nodes. They can be aggregation nodes in the middle of the network that combine several user flows into one big information flow.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... resources.^2.4

Resources have contention when they have more arrivals/requests that what they can process. Two examples are the outgoing links and the router interconnect.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 2017.^2.5

It is interesting to note that for IP revenues to surpass those of long-distance telephony the Internet revenue per household would have to multiply by 358%.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... VPNs.^2.6

Interestingly, these features are added to provide traffic isolation and engineering, features that are intrinsic to circuit switching.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... time,^2.7

In April 1995, commercial Internet was born after the decommissioning of the NSFnet. In March 1994, Sprint first announced its deployment of directional SONET rings. The first deployments of WDM were from June 1996.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... network.^2.8

Stanford University (with a population of about 15,000 people) employs 80 full-time telephone engineers, 25 full-time IP network engineers, and 350 part-time local IP network administrators.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... O'Dell^2.9

Former Senior Vice President of UUNET, responsible for technical strategic direction and architecture of the network.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... edge.^2.10

Chapter 4 and Chapter 5 describe two ways of integrating IP and circuit switching in the core.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... files,^3.1

Web browsing and file sharing represent over 65% of Internet transferred bytes today [31].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... file^3.2

For purposes of this chapter, I will define ``1 Mbit'' to be equal to 10⁶ bits, not 2¹⁰ bits, in order to simplify our examples.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... circuit^4.1

It is worth noting that virtual circuits are just a connection-oriented packet-switching technique.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... network^4.2

This data was made available through the National Science Foundation Cooperative Agreement No. ANI-9807479, and the National Laboratory of Applied Network Research.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... size^4.3

The flow size is the number of bytes transported by the flow, including the header overhead, control messages and retransmissions.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... lifetime.^4.4

If the circuit were timed out during the lifetime of the flow, it can be reestablished rapidly. However, there is a risk that a new request gets rejected because the old resources have been claimed by another flow.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... router.^4.5

This prototype was built by Byung-Gon Chun, an M.S. student at Stanford, for a 10-week class project under my supervision [39].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 2.7\%.^4.6

Since links in the core are usually symmetrical, the bandwidth inefficiencies caused by this traffic unbalance also affect packet switching.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...chap:TCPSwitching.^5.1

Typically, electronic SONET circuit switches use circuit granularities of STS-1 (51 Mbit/s) or higher, whereas DWDM switches have granularities of OC-48 (2.5 Gbit/s) or higher.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... peer.^5.2

This is the main reason for the advance notice in Figure 5.3. Unfortunately, it was impossible to know what application caused this behavior because the trace was ``sanitized''.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... bandwidth^5.3

The only requirement for the flow peak bandwidth is that it has to be larger than the flow average bandwidth.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... bandwidths,^5.4

In the analysis, the peak flow rate is defined as the minimum number of 56-Kbit/s circuits that are needed to carry the average flow rate. Over 97.5% of the flows fitted within a single 56-Kbit/s circuit. A tighter bound could have been used, as well.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... bandwidth.^5.5

One could use the number of bytes transferred by the flow instead of the flow average bandwidth, since the latter is equal to the former divided by the flow duration.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... -0.299,^5.6

For TCP traffic, the correlation coefficient was between -0.137 and -0.310, whereas for non-TCP traffic, it was between -0.089 and -0.391.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... label^6.1

If the LSP is composed of several parallel channels, the downstream node may return one label for each channel.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... restoration^6.2

Protection refers to the extremely fast recovery from a failure (such as the 50 ms recovery time of SONET/SDH rings), whereas restoration is a slower failure recovery that relies on the regular signaling and routing mechanisms to re-establish the service.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... wavelengths^6.3

These wavelengths carry either SONET/SDH channels of 2.5-Gbit/s or 10-Gbit/s, or Ethernet channels of 1 Gbit/s or 10 Gbit/s.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 3R^6.4

Reamplification, Reshaping and Retiming.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... (SOA-based^6.5

Semiconductor Optical Amplifier.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... (O/E^6.6

Electronics-to-Optics.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... time.^6.7

Remember that TCP Switching also had an increase of the RTT because of the transmission times over thin circuits. As the access-link rate increases in TCP Switching, the RTT increase becomes negligible.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... arrivals^6.8

A recommended sampling rate is 1 packet out of 100.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... switches.^6.9

Even if OBS uses circuit switching to forward the bursts, from the point of view of performance it behaves like a packet switching technology that switches very large packets (the bursts) using cut-trough techniques.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...chap:CSvsPS.^7.1

It is interesting to note that despite the almost simultaneous coming of age of IP and SONET/SDH in the late 80's, current IP routers have not been able to displace electronic circuit switching in the core, as shown in Table 2.1. This could be an indication of what can happen with OBS/OPS.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.