Download Protocol Layering and Internet Policy and more Essays (university) Computer Networks in PDF only on Docsity!
ARTICLE
PROTOCOL LAYERING AND INTERNET POLICY
C HRISTOPHER S. YOO †
An architectural principle known as protocol layering is widely recognized as
one of the foundations of the Internet’s success. In addition, some scholars and
industry participants have urged using the layers model as a central organizing
principle for regulatory policy. Despite its importance as a concept, a comprehensive
analysis of protocol layering and its implications for Internet policy has yet to appear
in the literature. This Article attempts to correct this omission. It begins with a
detailed description of the way the five-layer model developed, introducing protocol
layering’s central features, such as the division of functions across layers, infor-
mation hiding, peer communication, and encapsulation. It then discusses the
model’s implications for whether particular functions are performed at the edge or in
the core of the network, contrasts the model with the way that layering has been
depicted in the legal commentary, and analyzes attempts to use layering as a basis
for competition policy. Next the Article identifies certain emerging features of the
Internet that are placing pressure on the layered model, including WiFi routers,
network-based security, modern routing protocols, and wireless broadband. These
developments illustrate how every architecture inevitably limits functionality as
well as the architecture’s ability to evolve over time in response to changes in the
† (^) John H. Chestnut Professor of Law, Communication, and Computer & Information Sci-
ence and Founding Director of the Center for Technology, Innovation and Competition (CTIC), University of Pennsylvania. I would like to thank the Milton and Miriam Handler Foundation, the New York Bar Foundation, and CTIC for their financial support for this project. I would also like to thank the participants in a workshop conducted at the Center for Law and Information Policy at Fordham Law School; the Duke Law Review ’s 41st Annual Administrative Law Symposium; the workshop on Technology Policy, Law and Economics co-sponsored by the Swiss Federal Institute of Technology (ETH Zürich) and the University of Zürich; and the University of Pennsylvania Law Review ’s symposium, “The Evolving Internet,” held at the University of Pennsylvania Law School, as well as Adam Aviv, Steve Bellovin, Matthew Blaze, Vint Cerf, Bob Kahn, Howard Shelanski, Larry Solum, Jonathan Smith, Konstantinos Stylianou, Rick Whitt, and Jonathan Zittrain for their comments on earlier drafts. All errors are the responsibility of the author.
1708 University of Pennsylvania Law Review [Vol. 161: 1707
technological and economic environment. Together these considerations support
adopting a more dynamic perspective on layering and caution against using layers
as a basis for a regulatory mandate for fear of cementing the existing technology into
place in a way that prevents the network from innovating and evolving in response
to shifts in the underlying technology and consumer demand.
I NTRODUCTION ............................................................................ 1709
I. T HE C ONCEPTUAL U NDERPINNINGS OF
P ROTOCOL L AYERING .............................................................. 1716
A. Modularity Theory ..................................................................... 1718
B. Peer Communication and Encapsulation ........................................
C. The Tradeoffs Inherent in Protocol Layering .................................. 1724
II. T HE I NTERNET AS AN E XAMPLE OF A
L AYERED A RCHITECTURE ...................................................... 1730
A. Connecting Heterogeneous Hosts .................................................. 1730
B. Interconnecting Heterogeneous Transmission Technologies ............... 1735
C. The TCP/IP Reference Model ..................................................... 1742
1. The Application Layer ...................................................... 1742
2. The Transport Layer ........................................................ 1743
3. The Network Layer .......................................................... 1745
4. The Data-Link Layer ........................................................ 1746
5. The Physical Layer ........................................................... 1747
D. Layering’s Implications for Where
Functions Are Performed ............................................................ 1748
III. C HARACTERIZATIONS OF THE L AYERED M ODEL
A PPEARING IN THE L EGAL L ITERATURE ................................. 1748
A. Combining the Transport and Network Layers
into a Single Layer ................................................................... 1749
B. Dumb Pipes vs. the Hourglass Model ........................................... 1750
C. Layering and Competition Policy ................................................. 1752
IV. T HE I MPACT OF T ECHNOLOGICAL C HANGE
ON THE L AYERED M ODEL ...................................................... 1754
A. Reliability ................................................................................ 1755
B. Congestion ................................................................................ 1758
C. Distributed Optimization ............................................................ 1762
1. Aggressive TCP Implementations ..................................... 1762
a. Refusal to Back Off in the Face of Congestion .................... 1763
b. Multiple TCP Sessions .................................................. 1763
c. Autotuning .................................................................. 1765
2. Simultaneous Optimization............................................... 1766
1710 University of Pennsylvania Law Review [Vol. 161: 1707
Indeed, belief in the layered model has become so strong that it is often
widely regarded as the “proper” way to modularize a network.^5
There is widespread agreement that the incorporation of protocol layer-
ing into the Internet’s architecture has yielded substantial benefits. Layer-
ing allows those working on one layer to ignore most of the inner workings
of the other layers, which reduces coordination costs and accelerates product
development times by permitting parallel testing and innovation. Layered
architectures also provide a stable configuration of network resources and
interfaces around which actors can focus their efforts. The current architec-
ture has also proven incredibly resilient. Despite originally being designed
for a much smaller scale and a more limited technological and economic
context, the Internet now integrates a larger number and greater variety of
uses and technologies than its designers ever imagined.^6
Protocol layering has also found its way into discussions of Internet
policy.^7 Early commentators offered it as a technologically agnostic alter-
native to the regime established by the Communications Act of 1934,^8 which
subjected communications to distinct regulatory regimes based on whether
APPROACH 48-56 (5th ed. 2010); ANDREW S. T ANENBAUM , COMPUTER N ETWORKS (4th ed. 2003). Even textbooks organized along different lines still mention protocol layering prominently. See, e.g. , LARRY L. P ETERSON & B RUCE S. D AVIE , C OMPUTER N ETWORKS : A SYSTEMS APPROACH 19-30 (4th ed. 2007). (^5) See David Clark, Foreword to the First Edition of PETERSON & D AVIE , supra note 4, at ix, ix
(“All good computer scientists worship the god of modularity.... The field of network protocols is perhaps unique in that the ‘proper’ modularity has been handed down to us in the form of an international standard: the seven-layer reference model of network protocols from the ISO.”); David D. Clark, Modularity and Efficiency in Protocol Implementation 24 (IETF Network Working Grp. RFC No. 817, 1982) [hereinafter RFC 817], available at http://tools.ietf.org/pdf/ rfc817 (noting the “tempt[ation] to think that a layer boundary... is in fact the proper boundary to use in modularizing the implementation”). (^6) See STUART MINOR B ENJAMIN ET AL., TELECOMMUNICATIONS LAW AND POLICY 713-
(3d ed. 2012). (^7) For leading discussions of layering in the legal literature, see L AWRENCE LESSIG , THE
FUTURE OF I DEAS : T HE FATE OF THE C OMMONS IN A CONNECTED WORLD 23-25 (2001); B ARBARA VAN SCHEWICK, INTERNET ARCHITECTURE AND I NNOVATION 46-57, 83-105 (2010); J ONATHAN ZITTRAIN, T HE FUTURE OF THE I NTERNET AND HOW TO STOP I T 67-69 (2008); Lawrence B. Solum & Minn Chung, The Layers Principle: Internet Architecture and the Law , 79 N OTRE DAME L. REV. 815, 823-849 (2004); Adam Thierer, Are “Dumb Pipe” Mandates Smart Policy? Vertical Integration, Net Neutrality, and the Network Layers Model , 3 J. ON T ELECOMM. & HIGH TECH. L. 275, 279-83 (2005); Kevin Werbach, A Layered Model for Internet Policy , 1 J. ON T ELECOMM. & HIGH T ECH. L. 37, 57-64 (2002); Richard S. Whitt, A Horizontal Leap Forward: Formulating a New Communications Public Policy Framework Based on the Network Layers Model , 56 FED. C OMM. L.J. 587, 601-09, 614-36 (2004); and Timothy Wu, Essay, Application-Centered Internet Analysis , 85 V A. L. REV. 1163, 1164, 1189-92 (1999). (^8) Pub. L. No. 416, 48 Stat. 1064 (codified as amended in 47 U.S.C.).
2013] Protocol Layering and Internet Policy 1711
they were transmitted over telephone wires, coaxial cables, or spectrum.^9
Others argued that the layered model remained properly agnostic about the
content of the rules, but argued that problematic practices arising in one
layer be addressed only through regulations directly targeted at that layer—
rather than through regulations designed to curb that behavior by targeting
another layer or the system as a whole.^10
Other analyses have drawn stronger policy inferences from the layered
model. For example, some commentators argued that the layered model
could support competition policy by providing “natural boundaries” for
defining markets,^11 noting that each layer is subject to different sources of
market power.^12 Others went further, suggesting that the economics of the
lower layers made them particularly susceptible to market power, although
they acknowledged the possibility of deregulating the lower layers once
they became more competitive.^13 Others argued that layering promotes “fair
(^9) See Douglas C. Sicker & Joshua L. Mindel, Refinements of a Layered Model for Telecommuni-
cations Policy , 1 J. ON TELECOMM. & HIGH TECH. L. 69, 72 (2002) (explaining the “existing policy” of regulating different services under different titles of the Communications Act); Werbach, supra note 7, at 64-65 (discussing how the distinct regulation of telephone and cable led to inconsistent treatment between similar technologies, such as DSL and cable modem services); Whitt, supra note 7, at 615-17 (surveying critiques of the current silo approach that uses the underlying technology as the basis for regulation rather than concepts of layering). (^10) See Solum & Chung, supra note 7, at 849, 853 (defining the “layers principle” as the need
to “respect the integrity of the layers” and advocating “minimization of layer-crossing regula- tion”); Douglas C. Sicker, Further Defining a Layered Model for Telecommunications Policy 13 (Oct. 3, 2002) (paper presented at the 30th Telecomms. Policy Research Conf.), available at http:// www.learningace.com/doc/1675669/875a17ec13593859fdd613067974f72b/tprc_layered_model (“[P]olicy issues at one layer should be recognized as different from policy issues at another layer.”). (^11) Robert Cannon, The Legacy of the Federal Communications Commission’s Computer Inquiries ,
55 FED. C OMM. L.J. 167, 195 (2003). (^12) See Robert M. Entman, Transition to an IP Environment (“Higher degrees of competition
may be more feasible and desirable at some layers than others. Therefore, policymakers should recognize that a pro-competitive policy may need to treat different layers differently.”), in ASPEN I NST., T RANSITION TO AN IP E NVIRONMENT: A REPORT OF THE FIFTEENTH ANNUAL ASPEN I NSTITUTE C ONFERENCE ON TELECOMMUNICATIONS P OLICY 1, 2-3 (2001), available at http:// www.aspeninstitute.org/sites/default/files/content/docs/cands/TRANSITION_BK.PDF; Michael L. Katz, Thoughts on the Implications of Technological Change for Telecommunications Policy (noting that “the assessment of market power should largely take place at each layer separately” and discussing how the sources of market power at the transport layer differ from the sources of market power at the applications layer), in ASPEN I NST., supra , at 25, 37-38. (^13) See, e.g. , Werbach, supra note 7, at 66 (“If the physical and logical infrastructure layers in
the relevant markets were sufficiently competitive, ILECs would not be able to gain unfair advantage over competitors at the application and content layers.”); Whitt, supra note 7, at 592, 649, 653, 667 (“[W]hen applied in the telecommunications industry context, the Network Layers Model targets the lower network layers for discrete regulation based on the existence of significant market power, rather than legacy service or industry labels.”).
2013] Protocol Layering and Internet Policy 1713
some observations about implicit tradeoffs in layering that, despite being
made with respect to an earlier architecture and concerns that did not fully
mature, still reflect some basic insights. Clark recognized that the centrality
of layering in the engineering literature “tends to suggest that layering is a
fundamentally wonderful idea which should be a part of every consideration
of protocols.”^20 Such a perspective overlooks the fact that layering provides
“both a benefit and a penalty.”^21 While “[a] visible layer boundary, with a
well specified interface, provides a form of isolation between two layers”
that permits modifications to one layer without interfering with other
layers, “a firm layer boundary almost inevitably leads to inefficient oper-
ation.”^22 Hiding much of the technical complexity behind layer boundaries
prevents other layers from taking advantage of the full functionality of the
underlying technology, which in turn increases the resources needed to
perform the desired task.^23 Thus, the “tempt[ation] to think that a layer
boundary... is in fact the proper boundary to use in modularizing the
implementation” is “a potential snare.”^24 The tradeoff between generality
and efficiency is “rarely acknowledged in the computing literature,” how-
ever.^25 A small but important body of work exists in the engineering
literature exploring how protocol layering can harm network performance.^26
(^20) RFC 817, supra note 5, at 24. (^21) Id. at 16. (^22) Id. (^23) Id. at 17. Clark continues:
In fact, layering is a mixed blessing. Clearly, a layer interface is necessary whenever more than one client of a particular layer is to be allowed to use that same layer. But an interface, precisely because it is fixed, inevitably leads to a lack of complete un- derstanding as to what one layer wishes to obtain from another. This has to lead to inefficiency.
Id. at 24. (^24) Id. ; see also C OMER , supra note 4, at 169 (observing that “strict layering can be extremely
inefficient” by sometimes preventing a layer “from optimizing transfers”); RFC 871, supra note 3, at 11 (listing efficiency concerns arising from layering). (^25) Jean-François Blanchette, A Material History of Bits , 62 J. AM. SOC’Y FOR I NFO. SCI. &
T ECH. 1042, 1047 (2011). (^26) For classic analyses of the potential downsides of protocol layering, see Greg Chesson,
Protocol Engine Design , 1987 PROC. USENIX SUMMER CONF. 209; Jon Crowcroft et al., Is Layering Harmful? , IEEE NETWORK, Jan. 1992, at 20, 23-24; David Tennenhouse, Layered Multiplexing Considered Harmful , in P ROTOCOLS FOR HIGH-SPEED NETWORKS 143, 144-45 (H. Rudin & R. Williamson eds., 1989); David D. Clark & David L. Tennenhouse, Architectural Considerations for a New Generation of Protocols , C OMPUTER C OMM. REV., Sept. 1990, at 200, 205- 07; and Randy Bush & David Meyer, Some Internet Architectural Guidelines and Philosophy 7- (IETF Network Working Grp. RFC No. 3439, 2002) [hereinafter RFC 3439], available at http:// tools.ietf.org/pdf/rfc3439.
1714 University of Pennsylvania Law Review [Vol. 161: 1707
In addition to its impact on efficiency, protocol layering can also have an
adverse impact on innovation that is often overlooked. Although protocol
layering promotes innovations that are consistent with the architecture, at
the same time it impedes innovations that are inconsistent with the design
hierarchy.^27 Moreover, any changes that require a reconfiguration of the
design hierarchy require coordinating with actors operating at the layers
both above and below the locus of the innovation, which makes such
innovations all the more difficult to implement.
The existing policy debate based on protocol layering largely ignores the
extent to which it is something of a mixed blessing from the standpoint of
innovation. On the one hand, to yield any benefits, an architecture must be
relatively stable and change only rarely.^28 Indeed, the natural temptation for
computer scientists to optimize for particular applications^29 or to redesign
the entire system from scratch means that any calls for a fundamental
redesign of the entire architecture should be greeted with a healthy amount
of skepticism.^30
On the other hand, to say architectural changes should be infrequent is
not to say that they should never occur. Even the strongest proponents of
the layered model recognize that the architecture can and should evolve
over time.^31 Major changes transforming the Internet environment—
including the growing heterogeneity of end users, the advent of Internet-
based video and cloud computing, and the emergence of wireless broadband
and the smartphone operating system as the relevant platforms—raise the
possibility that circumstances may have changed sufficiently to justify a
change in the architecture.^32 Indeed, the emergence of wireless broadband
as an important medium of transmission has spawned a growing literature
on cross-layer design exploring new architectures that deviate from the
(^27) See infra notes 76-78 and accompanying text. (^28) See Blanchette, supra note 25, at 1054. (^29) D.L. Parnas, Information Distribution Aspects of Design Methodology , 1 I NFO. PROCESSING 71:
P ROC. IFIP CONG. 71, at 339, 342 (1972). (^30) See Blanchette, supra note 25, at 1055 (noting that developers “often fantasize” about a
“clean slate” but that effective “infrastructural change proceeds just as much through improvisa- tion, bricolage, and drift”). (^31) See Solum & Chung, supra note 7, at 865 (“As the Internet evolves, it is possible that supe-
rior architectures may be conceived.”); Werbach, supra note 7, at 66 (“If the physical and logical infrastructure layers in the relevant markets were sufficiently competitive, ILECs would not be able to gain unfair advantage over competitors at the application and content layers.”); Whitt, supra note 7, at 619 (“At its core, the layers principle is a pragmatic tool... [and] policymakers should take care not to enshrine it as either definitive or dispositive in each and every situation.”). (^32) See C HRISTOPHER S. YOO, T HE D YNAMIC INTERNET 4 (2012) (“The dramatic shift in
Internet usage suggests that its founding architectural principles... may no longer be appropriate today.”).
1716 University of Pennsylvania Law Review [Vol. 161: 1707
of engineering principles will obscure rather than promote sound policy
analysis.
I. THE C ONCEPTUAL U NDERPINNINGS
OF PROTOCOL LAYERING
As described above, understanding the relative merits of a layered archi-
tecture as well as the circumstances under which it can and should change
requires understanding the theory underlying the principle. Because
layering is widely recognized as a particular form of modularity,^37 Section A
offers a basic introduction to modularity theory. Section B moves past
modularity in general to discuss protocol layering in particular. Section C
analyzes the advantages and disadvantages to layering suggested by the theory.
To deal first with some preliminary matters of nomenclature, the com-
puters with which end users connect to the Internet are called hosts , and the
various programs running on any particular host comprise a number of
processes. Because the Internet is a network of networks, some of these
computers are located within a network, while others serve as gateways
between networks. Nodes that route traffic within a network are typically
called switches , while nodes that route traffic between networks are called
routers.^38
A particular convention for formatting, interpreting, and reacting to a
communication is called a protocol.^39 The functions of a protocol are well
illustrated by the protocol used in the traditional postal system. Effective
transmission of the mail requires agreement between those sending and
carrying mail as to where to locate the relevant information. By convention,
the return address for letters is located in the upper left-hand corner, the
postage in the upper right-hand corner, and the destination address in the
middle. The convention for postcards is somewhat different. For many
postcards, the return address, the destination address, and the postage are
all located on the right-hand side of the card.
Mail systems must also agree on how to interpret the content of the in-
formation conveyed in these locations, such as the significance of particular
ZIP codes, street addresses, and bar codes. Standardizing where important
(^37) For discussions of layering as a unique form of modularity, see VAN SCHEWICK, supra note
7, at 46, 379; Blanchette, supra note 25, at 1046; Crowcroft et al., supra note 26, at 23; RFC 871, supra note 3, at 7; Douglas C. Sicker & Lisa Blumensaadt, Misunderstanding the Layered Model(s) , 4 J. ON TELECOMM. & HIGH T ECH. L. 299, 305 (2006); Philip J. Weiser, Law and Information Platforms , 1 J. ON T ELECOMM. & HIGH T ECH. L. 1, 4 (2002); Werbach, supra note 7, at 59 n.85; Wu, supra note 7, at 1190. (^38) P ETERSON & DAVIE , supra note 4, at 253. (^39) KUROSE & ROSS , supra note 4, at 9.
2013] Protocol Layering and Internet Policy 1717
information is located greatly facilitates the mail system’s ability to process
the mail while simultaneously flagging for the carrier what information may
safely be ignored, such as the personal message written on the left-hand side
of post cards. American conventions are by no means the only feasible
formats. Indeed, many mail systems in Asia format addresses in the reverse
order, with the state being listed first, followed by the city, and then the
street address.^40 Despite these differences, these systems will remain
interoperable so long as each mail system is able to interpret and convert
addresses in the other format.
In addition, mail systems must share an understanding of how to handle
particular situations. Some of these features control the tasks internal to one
actor, such as how they should treat hold orders and change-of-address
notices. The actors must also agree on what happens if the post office
attempting to deliver a piece of mail cannot locate the destination address.
For first-class mail, post offices return undeliverable mail to the location
listed in the return address. For lower classes of mail, however, the post
office simply discards the mail.
Understanding how other actors are expected to behave under particular
circumstances provides each actor with a guide to interpreting and reacting
to what has happened. To use the example described above, if a piece of
first-class mail is not returned to sender, the person sending it may assume
that it was successfully delivered.^41 The different treatment of lower classes
of mail means that senders cannot infer anything from the fact that an
article sent via a lower class was not returned. And at a far end of the
spectrum, some classes of service require the post office to send a confirma-
tion of delivery back to the sender once mail is delivered, whereas most
classes of mail do not. If the sender knows that the post office is supposed
to send a delivery confirmation, it may regard the failure to receive a
confirmation within a reasonable amount of time as an indication that the
letter never arrived. Based on this inference, the sender can take whatever
action it deems appropriate, whether that means resending the letter,
choosing a different mode of communication, or abandoning attempts to
convey the information altogether.
(^40) Appendix V International Address Formats , M ICROSOFT D EVELOPER N ETWORK, http://
msdn.microsoft.com/en-us/library/cc195167.aspx (last visited Apr. 4, 2013). (^41) This inference is not conclusive, as it is always possible that the mail was lost or destroyed.
Whether the sender should take additional means to verify delivery depends on the likelihood of an adverse event as well as the value of what was sent.
2013] Protocol Layering and Internet Policy 1719
Consider, for example, the advent of printers using USB ports. Certain
aspects of printer design are intimately tied to whether the printer is a laser
or an inkjet printer. The creation of a modular interface allows computers
connected to those printers to ignore almost all of the details about how any
particular printer operates. As long as the computer provides the data in the
correct format, the printer should operate without any problems. Conversely,
as long as the printer remains ready to process any data submitted in the
correct format, the printer’s design can be changed without having any
impact on the overall system.
Despite the design architect’s best efforts, modular systems can rarely be
defined a priori. Many aspects of how tasks interact with one another can be
understood only after the architect experiments with different solutions.^50
Consequently, modular systems more often result from “improvisation,
bricolage, and drift” than from some theoretical conception of the ideal
architecture.^51
B. Peer Communication and Encapsulation
Layering represents a very particular form of modularity, in which
different parts of the overall system are arranged into parallel hierarchies. In
the typical Internet transaction, a process generates a message and transfers
it to the operating system running on the host. The operating system
divides the message into packets configured for the Internet and hands
them off to the first-hop router of a communications network. The sending
communications network will convey these packets to the receiving com-
munications network, which in turn passes them to the receiving host’s
operating system. The operating system then passes them to the process
running on the receiving host.
The type of modularity enforced by layering has several distinct charac-
teristics. First, the modules are arranged into a series of client-server
relationships, where “each layer is a server to the layer above, and a client to
the layer below.”^52 Second, this arrangement is strictly hierarchical; every
(^50) See BALDWIN & C LARK, supra note 48, at 254 (“Given such a high degree of complexity, it
simply is not possible for designers to know enough about the system to eliminate all uncertainty. Thus each new design is fundamentally an experiment.”); Sendil K. Ethiraj & Daniel Levinthal, Modularity and Innovation in Complex Systems , 50 MGMT. SCI. 159, 172 (2004) (noting that designers “lack omniscience”). (^51) Blanchette, supra note 25, at 1055. (^52) Id. at 1046.
1720 University of Pennsylvania Law Review [Vol. 161: 1707
layer interacts exclusively with the layers above and below with respect to
every communication without bypassing either.^53
Third, layering differs from other modular schemes in its focus on estab-
lishing communication between peers.^54 Unlike the example of the USB
port given above, in which a personal computer could establish a connection
directly with a printer, layered architectures require that connections be
established only between parallel elements in the hierarchy. For example,
applications communicate with other applications; operating systems commu-
nicate with other operating systems; routers communicate with other routers.
Figure 1: Layering as Peer Communication
55
Layering ensures that peers communicate only with peers operating at
the same level through a practice known as encapsulation. Under this ap-
proach, each layer takes the data packet provided by the layer above,
extracts all of the information that the next layer will need to perform its
functions, places that information into a new packet’s header, and then
places the entirety of the packet it received from the higher layer in the
payload of the new packet.^56 Since layering requires that each layer examine
only the information contained in the header and prohibits it from examining
(^53) See RFC 871, supra note 3, at 9 (explaining that protocols operate in a hierarchy with a
strict “chain of command”). (^54) See T ANENBAUM , supra note 4, at 26-30 (explaining how layering protocols facilitate
communication between peers). (^55) See id. at 27 fig.1-13. (^56) KUROSE & ROSS , supra note 4, at 55-56.
1722 University of Pennsylvania Law Review [Vol. 161: 1707
of the message from Vint—a message no one but the two of them had any
occasion to see.
Note the key features of this system. Each of the peers receives the exact
same communication. After the memo is removed from the interoffice mail
envelope, Bob receives the exact message sent by Vint. Once the U.S. mail
envelope is opened, the receiving mail room receives precisely the same
interoffice mail envelope and contents as the one sent by the sending mail
room. After the shipping container is unpacked, the two post offices ex-
change precisely the same envelope containing the same message as well.
The fact that each actor encapsulates the entirety of the communication in a
larger envelope ensures that at each step the original message can be
recovered unaltered. In addition, refusing to look inside the envelope until
it is de-encapsulated ensures that lower-layer protocols cannot make any
assumptions about objects handed to them by the upper-layer protocols.
Using an example more closely related to the Internet, consider what
occurs when an end user sends an email message. The end user’s email
client, such as Microsoft Outlook, first extracts the source and destination
addresses and places that information in the Simple Mail Transfer Protocol
(SMTP) header, which is the general format used for email. Then the email
client encapsulates the email in an SMTP message by placing the entire
email in the payload of the packet and passes it to the next layer, the
transport layer. The transport layer, in turn, reads the information it needs
from the SMTP header and places that information in the Transmission
Control Protocol (TCP) header, places the segment in the payload, and
passes the segment to the network layer. The network layer reads the source
and destination addresses from the TCP header, places the necessary
information in the Internet Protocol (IP) header, encapsulates the segment
into an IP datagram, and passes the datagram to the data-link layer. The
data-link layer reads the information it needs from the IP header, places the
relevant information in the frame header, and encapsulates the datagram
into a data-link layer frame. The data-link layer will use the header infor-
mation to move the frame hop by hop.
As long as each hop remains within both the same network and the same
data-link technology, there is no need to de-encapsulate the frame to reveal
information. But when the hop reaches a gateway to another network, the
technology may change. Because moving to another network may involve
shifting to a different data-link technology, the gateway de-encapsulates the
data-link layer frame and passes it to the next network as an IP datagram.
The next network will re-encapsulate it in a frame appropriate for its data-
link technology and will pass it along in this manner until it reaches the
border of another network, when the de-encapsulation process begins again.
2013] Protocol Layering and Internet Policy 1723
Eventually, the packet will reach the receiving host. The host will de-
encapsulate the data-link header and pass the IP datagram to the network
layer. The network layer will remove the IP header and pass the segment to
the transport layer. The transport layer will strip off the TCP header and
pass the SMTP message to the application layer. Finally, the application
layer will strip off the SMTP header and pass the email to the program that
the receiving end user is using to read the message.
Although it is the email clients that are exchanging messages, no direct
transfers of data occur between them. Instead, each layer passes the data
and control information down through the stack until it reaches the physical
layer, which provides the only direct connection. After the communication
arrives at the receiving host, the data passes up through the layered stack
until it reaches the same layer as the peer sending the communication, at
which point it can act on the same information. For this reason, all connec-
tions above the physical layer are simply virtual.^60
It is often said that layered architectures ensure that every entity in the
receiving hierarchy receives the exact same object as the one sent by its peer
in the sending hierarchy.^61 As is the case with almost every generalized
network engineering principle, there are important exceptions. To return to
the postal example, one post office may use a postmark to cancel a stamp, in
which case the object received will differ in a small way from the object
sent. Similar changes occur in the email example. For example, intermediate
mail transfer agents will include information in email messages noting that
they were received. In addition, the IP header contains a counter that is
reduced by one every time a packet traverses a hop, with routers ceasing to
route a packet further when the counter hits zero to ensure that packets do
not wander around the Internet forever. Nonetheless, the generalization
remains a useful concept as a reference model.
A related principle of layering is that lower layers can make no assump-
tions about the nature of the communications in the upper layers.^62 Because
all of the information that any particular layer needs is supposed to be
included in the packet’s header, no layer is expected to look inside the
payload of any packet that it processes. Any attempt to do so is regarded as
(^60) T ANENBAUM , supra note 4, at 27, 30. (^61) See, e.g. , C OMER, supra note 4, at 164 (“Layered protocols are designed so that layer n at
the destination receives exactly the same object sent by layer n at the source.”); id. at 165 (“Thus, the layering principle states that the packet received by the transport layer at the ultimate destination is identical to the packet sent by the transport layer at the original source.”). (^62) T ANENBAUM , supra note 4, at 448; RFC 871, supra note 3, at 16.
2013] Protocol Layering and Internet Policy 1725
Layering can also hasten innovation by allowing experiments with
different solutions in different layers to proceed in parallel.^73 In addition,
protocol layering accommodates uncertainty by making it easy to incorpo-
rate subsequent improvements into the existing system.^74 As real-option
theory indicates, the ability to postpone such choices can be an important
source of value, provided that the interfaces are clearly defined and remain
stable.^75
Like any modular system, protocol layering embodies a precommitment
about the types of information permitted to pass between modules. Prevent-
ing adjacent modules from acting on certain types of information reduces
the complexity of the system by constraining the number of interdependen-
cies that must be taken into account. At the same time, prohibiting modules
from taking into account all of the possible information inevitably limits
both the efficiency and functionality of the resulting system.
In terms of efficiency, the generality of the layers necessarily means that
certain customized solutions tailored to particular situations must be fore-
gone.^76 The layers and the interfaces connecting them predefine and limit
the way those layers interact.^77
(^71) See Carliss Y. Baldwin & Kim B. Clark, Managing in an Age of Modularity , HARV. B US.
REV., Sept.–Oct. 1997, at 84, 85 (“Different companies can take responsibility for separate modules and be confident that a reliable product will arise from their collective efforts.”). (^72) Langlois & Robertson, supra note 70, at 301. (^73) Baldwin & Clark, supra note 71, at 91; see also Karl Ulrich, The Role of Product Architecture in
the Manufacturing Firm , 24 RES. POL’ Y 419, 435 (1995) (“For the modular architecture, detailed design of each component can proceed almost independently and in parallel.”). (^74) Baldwin & Clark, supra note 71, at 91. (^75) See BALDWIN & C LARK, supra note 48, at 234-37 (applying real option theory to show
how modularity permits industries to postpone having to commit to any particular technological solution); id. at 284-93 (discussing the option value of hidden modules). (^76) See RFC 817, supra note 5, at 16, 24 (“[A]n interface, precisely because it is fixed, inevitably
leads to a lack of complete understanding as to what one layer wishes to obtain from another. This has to lead to inefficiency.”); RFC 871, supra note 3, at 20-22 (discussing the tradeoff inherent in the fact that fixed layers lead to less design flexibility). Blanchette discusses the loss of efficiency resulting from generalization:
[T]he most efficient programs are hand-tailored, providing no generalization whatso- ever; conversely, highly general abstractions will result in significant loss in efficiency. This is because the specification of an abstraction (the interface) general enough to accommodate a wide range of implementations necessarily involves trade-offs, between the freedom that the abstraction provides and the efficiency of possible implemen- tation.
Blanchette, supra note 25, at 1046-47 (internal quotation marks omitted). (^77) See T ANENBAUM , supra note 4, at 27 (“The interface defines which primitive operations
and services the lower layer makes available to the upper one.”); Parnas, supra note 29, at 339 (“The connections between modules are the assumptions which the modules make about each other.”).
1726 University of Pennsylvania Law Review [Vol. 161: 1707
Engineers have long recognized that hiding information can both harm
and promote innovation. Specifically, innovation that depends upon the
sharing of particular information cannot proceed if that information is held
in another layer and if the particular form of modularity imposed by the
architecture does not permit that information to pass through the protocol
stack. The aforementioned information hiding means that layering “hide[s]
vital information that lower layers may need to optimize their performance”
and requires “that the optimization of each layer... be done separately,”
which can “conflict with efficient implementation of data manipulation
functions.”^78
In other words, design hierarchies represent something of a mixed bless-
ing from the standpoint of innovation.^79 On the one hand, they facilitate
innovation that is consistent with the hierarchy. Indeed, the existing layered
architecture has proven incredibly robust. On the other hand, predetermin-
ing the locus of the interfaces and the information that can pass between
layers discourages innovations that are inconsistent with the hierarchy.^80
Protocol layering also limits the network’s ability to evolve. Any system
of modularity necessarily envisions that change to the basic architecture will
occur relatively slowly.^81 For the most part, the stability of the architecture
yields benefits. Predefining the interactions between particular components
makes it easier to make changes to individual components without disturbing
the system as a whole. Without a high degree of stability, actors could not
innovate in individual layers with any confidence.^82
At the same time, however, this stability can impede the network’s abil-
ity to evolve into a fundamentally different architecture. Economic theory
has long recognized that having an installed base in a network industry can
(^78) RFC 3439, supra note 26, at 7-8; see also Crowcroft et al., supra note 26, at 23 (“[T]he flip
side to modularization and data-hiding is that tuning the efficiency of the data path for transfer of data becomes difficult.... Vertical partitioning emphasises the discontinuities in the data path, which then obstruct the application from receiving the quality of service it requires.”). (^79) Christopher S. Yoo, Product Life Cycle Theory and the Maturation of the Internet , 104 NW. U.
L. REV. 641, 655-56 (2010). (^80) See Kim B. Clark, The Interaction of Design Hierarchies and Market Concepts in Technological
Evolution , 14 RES. P OL’ Y 235, 246 (1985) (using automobiles as an example and noting that “[o]nce choices about core concepts in engines were established, innovative effort moved down into subsidiary parameters”). (^81) See C OMM. ON THE I NTERNET IN THE E VOLVING I NFO. INFRASTRUCTURE , supra note
2, at 38 (arguing that innovation at the “center of the network” is slow because building new features into the existing network requires coordinating the actions of multiple developers); Parnas et al., supra note 66, at 409 (explaining that changes to modular interfaces should be limited to changes that are unlikely to be needed). (^82) See Blanchette, supra note 25, at 1054.
