Building Ethernet Connectivity Servicesfor Provider Networks

Eduard Bonada, Dolors SalaNetworking and Communications Research Group

Department of Information and Communications TechnologiesUniversitat Pompeu Fabra

{eduard.bonada, dolors.sala}@upf.edu

I. INTRODUCTION

Ethernet has become a dominant technology and it is nowa universal communication interface more than just an accesstechnology. Its low cost, high data rates, low complexityand simple maintenance offer Network Providers a goodopportunity for using Ethernet data networks at very largescale replacing the existing ATM/SONET or IP networking[1].

Ethernet was originally designed for LANs without verystrict requirements. But using it as a carrier-grade technologyrepresents a new application that leads to new requirements.These networks are larger, with higher redundancy and man-aged (no best-effort any more). Moreover, differentiation ofservices must be offered to allow for capacity limitations,different response time for different flows, or to considerprivacy and security aspects. In addition, a quick recovery timein case of failure is required (current providers SLA state amaximum time of 50ms).

Besides, there are other aspects that are of Providers interest.For instance, a low management complexity is desired in orderto avoid difficult maintaining tasks and to reduce the potentialamount of configuration mistakes. Ethernet’s plug’n’play is akey feature in this aspect.

Providers have already started deploying Ethernet technol-ogy in their networks. The simplest option is to implementEthernet on top of other technologies (SDH, Copper or HFC)but the performance is not optimized [2]. An emergent ap-proach is to deploy a GMPLS-based control plane to managethe Ethernet network [3] but it introduces the managementcomplexity mentioned before.

Another option is to extend Ethernet Bridging in order tosatisfy the needed requirements. Bridging is strongly linkedto Ethernet and its design has always been optimized forthis technology. We think it is the best direction to takebecause it represents the natural evolution of the technology.Moreover, it keeps the original essentials of Ethernet such aslow complexity or plug’n’play.

Ethernet Bridging provides the connectivity extension re-quired to overcome the distance and performance limitations ofa single network segment (a LAN). It is basically the intercon-nection of such segments through a multi-hop network formedby bridges [4]. It is plug’n’play because the configuration

needed to run (principally forwarding tables) is learned fromthe same network operation. However, this learning featurerequires a loop-free network (only a single path between everynode pair is allowed). Since redundancy is common in mostnetwork deployments, loops usually exist. The Spanning TreeProtocol (STP) avoids them by creating a logical tree-shapedtopology on top of the physical one (note a tree has not loops).It elects some of the network links for communication, thosebelonging to the tree, and blocks the rest. In order to avoiddata frames encounter temporary loops, the communication isstopped until the protocol has not configured the tree.

The original IEEE STP standard has suffered several evolu-tions in order to overcome its drawbacks as new applicationsfor Ethernet Bridging appeared. STP has a very slow config-uration time (50sec). The Rapid STP addresses this recoveryissue reducing it to the order of a few seconds (still far fromthe required 50ms). Another disadvantage is that the utilizationof a single tree only allows for communication through thoselinks that belong to the tree. Consequently networks cannot take advantage of its redundancy. This low efficiency issolved by the Multiple STP: a framework to segregate differentVLAN traffic into different trees. Another weak point is theconfiguration of sub-obtimal paths to communicate betweenmost of the network nodes. In a tree, the path between theroot and any other node (a tree-branch) is a shortest path.However, the path between two nodes in different branches isnot the optimal one. The IEEE Shortest Path Bridging (SPB)configures as many trees as nodes in the network so as toobtain a shortest path communication. This last is still underdiscussion in the IEEE 802.1.

All IEEE evolutions follow the trivial steps and are focusedon extending the features of the single tree by includingadditional configuration. There are other approaches that applydifferent bridging techniques like multiple trees managed by acentralized server [5], routed paths [6][7] or turn prohibitionalgorithms [8]. None of them complies with all bridgingfeatures of simplicity and automatic configuration.

Our approach is based on defining an improved type oftopology (improved tree) that provides total connectivity tothe network and meets the stated requirements: recoverytime below 50ms, path optimality, redundancy utilization anddifferent path selection based on node or link capabilities. Therest of this document includes the initial steps of this study.

978-1-4244-3968-3/09/$25.00 ©2009

2

II. ANALYSIS OF THE TREE CONSTRUCTION

This first phase focuses on the study of the STP behaviorand the characterization of the time it takes to configure thetree (convergence time). A rigorous study of the algorithm anda simulation model in ns-3 [9] are the basic methodologicalelements of this part.

The STP and its evolutions use a distance-vector algorithmto compute the paths that form the tree [10]. Note that a treehas a root node that connects to the rest through the treebranches. The tree creation process is based on an exchange ofmessages between network nodes. The nodes send messagesclaiming to be the root of the tree. However, only the nodewith the lowest id will really become the unique root. Themessages that start at the root are forwarded at every hopwhile others’ messages are discarded and hence not forwarded.It can be seen as a set of concentric wavefronts around eachnode advancing through the network where the root wavefrontalways beats all the others. A node realizes about the real rootwhen it receives a message started in the root node. Thus,when the root messages get to all nodes, they all know who isthe root and the path to it (through the port where the messageis received). This allows them to create the tree branches. Atthis point the tree is fully configured and communication canbe activated again. For further details see [11].

The most important conclusion is that the theoretical con-vergence time depends on the length of the path from the rootto the furthest node. The feasibility of this theoretical bounddepends on detecting the final state when the root messageis totally propagated. However, it is not easy to detect thecompletion of the tree to stop the algorithm as all nodesoperate distributed with minimal state. STP avoids additionalstate using timers. These timers need to be set up to verylarge values for proper operation under any network scenarioresulting in delays so large that convergence time is the sameregardless of the physical network.

Results in [11] show that a 4-degree network topology (eachnode has a maximum of 3 neighbors) could increase up to 4-million nodes and still converge in 50ms. On the other hand,STP would be limited to 400 nodes in the same scenario.

III. MULTIPLE TREES SUPERPOSITION STUDY

While the previous phase is focused on the process thatconfigures the tree, the objective of this second one is to devisethe shape of the improved topology.

Our approach is based on the study of how different treescan be superposed and which properties can be derived.The idea is to allow the configuration of multiple trees butminimizing the total amount of state required to maintainthem. Basically, we envision that some branches may coincidein different trees and hence these can be somehow shared.

In order to carry out the first tests the simulation in [9] hasbeen extended to support several tree instances in the samenetwork. At every execution, a tree rooted at each networknode is configured (as many trees as nodes). The superpositionbetween these different trees is analyzed.

TABLE ISUPERPOSITION LEVEL IN A 4-DEGREE GRID TOPOLOGY

number number number of superpositionnodes trees combinations level

42 6 75%3 4 63%

92 36 71%3 84 50%

162 120 70%3 560 60%

252 300 68%3 2300 60%

Table I includes the level of superposition (fourth column)when comparing different number of trees (second column)in a 4-degree topology of different size (first column). Thirdcolumn includes all possible combinations of tree comparisonsfor each case. Note that the superposition level is the averagevalue of all possible combinations.

The level of superposition is obtained as follows. Each treeinstance has a set of links that belong to the tree and a setthat do not (as previously said the STP blocks those links notbelonging to the tree). The number of tree links that coincidein all compared trees over the total number of links is thesuperposition level.

This preliminary results show a very high level of su-perposition that we are currently characterizing. We plan toexploit this high level of superposition to design the necessaryextensions of STP.

REFERENCES

[1] G Chiruvolu, A Ge, D Elie-Dit-Cosaque, M Ali, and J Rouyer. Issues andapproaches on extending Ethernet beyond LANs. IEEE CommunicationsMagazine, 42(3):80–86, 2004.

[2] A Kasim, P Adhikari, N Chen, N Finn, N Ghani, M Hajduczenia,P Havala, G Heron, M Howard, and L Martini. Delivering CarrierEthernet: Extending Ethernet beyond the LAN. McGrawHill, 2007.

[3] A Takacs, H Green, and B Tremblay. GMPLS controlled Ethernet: anemerging packet-oriented transport technology. IEEE CommunicationsMagazine, 46(9):118–124, 2008.

[4] IEEE standard for local and metropolitan area networks Media AccessControl (MAC) bridges. IEEE Std 802.1D-2004 (Revision of IEEE Std802.1D-1998), pages i–269, 2004.

[5] S Sharma, K Gopalan, S Nanda, and T Chiueh. Viking: a multi-spanning-tree Ethernet architecture for metropolitan area and clusternetworks. INFOCOM 2004, 4:2283 – 2294 vol.4, Feb 2004.

[6] R Perlman. Rbridges: Transparent Routing. INFOCOM 2004, 2:1211 –1218 vol.2, Feb 2004.

[7] K Lui, W Lee, and K Nahrstedt. STAR: a transparent spanning treebridge protocol with alternate routing. SIGCOMM 2002, 32(3), Jul 2002.

[8] Thomas Rodeheffer, Chandramohan Thekkath, and Darrell Anderson.SmartBridge: a Scalable Bridge Architecture. SIGCOMM 2000, Aug2000.

[9] E Bonada, D Cavic, and D Sala. Implementation of a layer-2 bridge inns3 (poster). SIMUTools 2008, 2008.

[10] R Perlman. An algorithm for distributed computation of a spanning treein an extended LAN. SIGCOMM 1985, pages 44–53, 1985.

[11] E Bonada and D Sala. On the Theoretical Bounds of the Spanning TreeAlgorithm. Jornadas Telecom I+D Bilbao 2008, 2008.