The Evolution of Layered Protocol Stacks Leads to an Hourglass-Shaped Architecture��� Saamer Akhshabi College of Computing Georgia Institute of Technology sakhshab@cc.gatech.edu Constantine Dovrolis College of Computing Georgia Institute of Technology dovrolis@cc.gatech.edu ABSTRACT The Internet protocol stack has a layered architecture that resem- bles an hourglass. The lower and higher layers tend to see frequent innovations, while the protocols at the waist of the hourglass appear to be ���ossified���. We propose EvoArch, an abstract model for study- ing protocol stacks and their evolution. EvoArch is based on a few principles about layered network architectures and their evolution in a competitive environment where protocols acquire value based on their higher layer applications and compete with other protocols at the same layer. EvoArch produces an hourglass structure that is similar to the Internet architecture from general initial conditions and in a robust manner. It also suggests a plausible explanation why some protocols, such as TCP or IP, managed to survive much longer than most other protocols at the same layers. Furthermore, it suggests ways to design more competitive new protocols and more evolvable future Internet architectures. Categories and Subject Descriptors: C.2.5 [Computer Commu- nication Networks]: Internet General Terms: Theory Keywords: Internet Architecture, Future Internet, Layering, Net- work Science, Evolutionary Kernels, Evolution. 1. INTRODUCTION Why does the Internet protocol stack resemble an hourglass? Is it a coincidence, intentional design, or the result of an evolutionary process in which new protocols compete with existing protocols that offer similar functionality and services? The protocol stack was not always shaped in this way. For instance, until the early nineties there were several other network-layer protocols compet- ing with IPv4, including Novell���s IPX, the X.25 network protocol used in Frame Relay, the ATM network layer signaling protocol, and several others. It was through a long process that IPv4 eventu- ally prevailed as practically the only surviving protocol at layer-3, creating a very narrow waist at the Internet architecture hourglass (see Figure 1). ���This research was supported by the NSF award 0831848 (���To- wards a Theory of Network Evolution���). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGCOMM���11, August 15���19, 2011, Toronto, Ontario, Canada. Copyright 2011 ACM 978-1-4503-0797-0/11/08 ...$10.00. TCP HTTP Skype/Kazaa Kazaa Twisted Pair Fiber Optical CDMA TDMA 802.11 Ethernet PPP IPv4 UDP RTP MPlayer Skype FireFox Silverlight Thunderbird Coaxial Cable DOCSIS P2P Protocol POP SMTP Figure 1: An (incomplete) illustration of the hourglass Internet architecture. Another important question is: why do we tend to see more fre- quent innovations at the lower or higher layers of the protocol hourglass, while the protocols at the waist of the hourglass appear to be ���ossified��� and difficult to replace? During the last 30���40 years we have seen many new physical and data link layer protocols created and surviving. And of course the same can be said about applications and application-layer protocols. On the other hand, the protocols at the waist of the hourglass (mostly IPv4, TCP and UDP) have been extremely stable and they have managed to out- compete any protocols that offer the same or similar functionality. How can a new protocol manage to survive the intense competition with those core protocols at the waist of the Internet hourglass? In fact, the ossification of the hourglass waist has been a major moti- vation for ���clean-slate��� efforts to design a novel future Internet ar- chitecture [16]. There are two important questions in that context. First, how can we make it more likely that a new (and potentially better) protocol replaces an existing and widely used incumbent protocol? And second, how can we make sure that a new archi- tecture we design today will not be ossified 10���20 years later? In other words, what makes a protocol stack or network architecture evolvable? The previous questions have generated an interesting debate [9, 10, 19]. In this paper, we attempt a first effort to study protocol stacks (and layered architectures, more generally) as well as their evolu- tion in a rigorous and quantitative manner. Instead of only con- sidering a specific protocol stack, we propose an abstract model in which protocols are represented by nodes, services are repre- sented by directed links, and so a protocol stack becomes a layered directed acyclic graph (or network). Further, the topology of this graph changes with time as new nodes are created at different lay- 206

ers, and existing nodes are removed as a result of competition with other nodes at the same layer. The proposed evolutionary model, referred to as EvoArch, is based on few principles about layered network architectures in which an ���item��� (or service) at layer-X is constructed (or composed) us- ing items at layer-(X-1). These principles capture the following: (a) the source of evolutionary value for an item, (b) the generality of items as we move to higher layers, (c) the condition under which two items compete, (d) the condition under which one item causes the death or removal of a competing item. Perhaps surprisingly, these few principles are sufficient to produce hourglass-shaped layered networks in relatively short evolutionary periods. As with any other model, EvoArch is only an abstraction of re- ality focusing on specific observed phenomena, in this case the hourglass structure of the Internet protocol stack, and attempting to identify a parsimonious set of principles or mechanisms that are sufficient to reproduce the observed phenomena. As such, EvoArch is an explanatory model (as opposed to black-box models that aim to only describe statistically some observations). EvoArch deliber- ately ignores many aspects of protocol architectures, such as the functionality of each layer, technological constraints, debates in standardization committees, and others.1 The fact that these practi- cal aspects are not considered by EvoArch does not mean that they are insignificant it means, however, that if the evolution of network architectures follows the principles that EvoArch is based on, then those aspects are neither necessary nor sufficient for the emergence of the hourglass structure. EvoArch is certainly not going to be the only model, or ���the cor- rect model���, for the emergence of hourglass-shaped network archi- tectures. It is likely that there are other models that can produce the same hourglass structure, based on different principles and param- eters. Additionally, EvoArch does not aim to capture every aspect of the Internet architecture it only focuses on the emergence of the hourglass structure, and so it may be the wrong model to use for other purposes (e.g., to study the economics of new protocol deployment). G.Box wrote that ���all models are wrong but some models are useful��� [3]. We believe that EvoArch is a useful model for (at least) the following ten reasons: 1- It gives us a new way to think about protocol stacks and net- work architectures and to study their evolutionary properties based on few fundamental principles (��2). 2- EvoArch provides a plausible explanation (but certainly not the only explanation) for the emergence of hourglass-like architectures in a bottom-up manner (��3). 3- EvoArch shows how the location and width of the hourglass waist can follow from certain key parameters of the underlying evo- lutionary process (��5). 4- EvoArch can be parameterized to produce a structure that is sim- ilar to the TCP/IP protocol stack, and it suggests an intriguing ex- planation for the survival of these protocols in the early days of the Internet (��5.4). 5- EvoArch suggests how to make a new protocol more likely to survive in a competitive environment, when there is a strong in- cumbent (��5.5). 6- EvoArch provides recommendations to designers of future Inter- net architectures that aim to make the latter more evolvable (��5.5). 7- EvoArch predicts that few protocols at the waist (or close to it) become ossified, surviving much longer than most other protocols 1The reader can see some of the criticism raised by anonymous reviewers in Section 9. at the same layer, and it shows how such ossified protocols can be eventually replaced (��6). 8- When we extend EvoArch to capture the effect of different pro- tocol qualities, we find that the lower part of the hourglass is sig- nificantly smaller than the upper part (��7.1). 9- The most stable protocols at the waist of the architecture are of- ten not those with the highest quality (��7.2). 10- Finally, EvoArch offers a new way to think about the compe- tition between IPv4 and IPv6 and to understand why the latter has not managed to replace the former (��7.3). The rest of the paper is structured as follows. In Section 2, we describe EvoArch and explain how the model relates to protocol stacks and evolving network architectures. In Section 3, we present basic results to illustrate the behavior of the model and introduce some key metrics. Section 4 is a robustness study showing that the model produces hourglass structures for a wide range of parameter values. The effect of those parameters is studied in Section 5 fo- cusing on the location and width of the waist. Section 6 examines the evolutionary kernels of the architecture, i.e., those few nodes at the waist that survive much longer than other nodes. Section 7 generalizes EvoArch in an important and realistic manner: what if different protocols at the same layer have different qualities (such as performance or extent of deployment)? We review related work in Section 8, present some criticism in Section 9, and conclude in Section 10. 2. MODEL DESCRIPTION In EvoArch, a protocol stack is modeled as a directed and acyclic network with L layers (see Figure 2). Protocols are represented by nodes, and protocol dependencies are represented by directed edges. If a protocol u at layer l uses the service provided by a pro- tocol w at layer l���1, the network includes an ���upwards��� edge from w to u.2 The layer of a node u is denoted by l(u). The incoming edges to a node u originate at the substrates of u, represented by the set of nodes S(u). Every node has at least one substrate, ex- cept the nodes at the bottom layer. The outgoing edges of a node u terminate at the products of u, represented by the set of nodes P (u). Every node has at least one product, except the nodes at the top layer. The substrates of a node are determined probabilistically when that node is created.3 Specifically, each layer l is associated with a probability s(l): a node u at layer l + 1 selects independently each node of layer l as substrate with probability s(l). We refer to s(l) as the generality of layer l. s(l) decreases as we move to higher layers, i.e., s(i) s(j) for i j. The decreasing generality prob- abilities capture that protocols at lower layers are more general in terms of their function or provided service than protocols at higher layers. For instance, in the case of the Internet protocol stack, a protocol at layer-1 offers a very general bit transfer service between two directly connected points this is a service or function that al- most any higher layer protocol would need. On the other extreme, an application-layer protocol, such as SMTP, offers a very special- ized service and it is only used by applications that are related to 2In practice, the principle of strict layering is occasionally violated through tunnels or other forms of virtual networks. For the most part, however, layering is the norm in protocol architectures rather than the exception. Considering architectures without strict layer- ing is outside the scope of this paper and an interesting subject for future research. 3Of course in practice substrates are never chosen randomly. The use of randomness in the model implies that a realistic mechanism of substrate selection is not necessary for the emergence of the hourglass structure. 207