Grid networks enabling grids with advanced communication technology phần 9 doc

14.2 Creating Next-generation Network Services and Infrastructure 279After 35 years of increasing Internet success, communications standards organiza-tions have now almost universally en

Routing Decision Node (RDN) is assumed The first three schemes basic standalonemechanisms, and the last scheme attempts to find the best combination of the abovetwo schemes The primary route from node I to E is depicted as a thick black lineand the link 2→ 3 is assumed to be broken during the operation.

When the headend 2 (or tailend 3) of link 2→ 3 detects the link failure, it informsthe RDN via the control plane The RDN conducts the routing recomputation andupdates the forwarding tables for all nodes, and new bursts will subsequently followthe new routes For example, new bursts will follow the route I→ 4 → 5 → 6 E Thissolution is optimal and the existing routing protocol can handle it well However,global routing table updating is a slow process (in seconds or even minutes) due

to the long round-trip time for the signal transmission and processing between theOBS nodes and the routing entity As a result, a large amount of bursts will be lostbefore the forwarding tables are updated

This is similar to the traditional deflection routing usually seen in a congestionresolution scheme When the headend 2 detects the link failure, it will automaticallypick up the alternative next hop in the forwarding table for every new burst whosenext hop on its primary route passes the faulty link In the example, new burstsfrom I to E will follow the alternative route I→ 1 → 2 → 5 → 6 → E This would bethe fastest restoration scheme since new bursts will be deflected to an alternativegood link right after the link failure is detected locally Therefore, it will incur thesmallest restoration loss However, because all the affected bursts are deflected toone alternative path, this scheme would increase the congestion loss

In this scheme, the headend 2 will also send a different fault notification message

to all its adjacent nodes in addition to the one to the RDN This fault notificationmessage contains the destination information for all the primary routes passing thefaulty link After receiving this message, each of the adjacent nodes will pick up

an alternative next hop for the affected bursts that are heading to the faulty linkaccording to their primary route In the example, bursts from I to E will take the newroute I→ 1 → 4 → 5 → 6 → E Compared with the local deflection scheme, neighbordeflection has the potential to make the rerouted traffic more distributed instead

of being totally deflected to one alternative path In this way, less congestion andtherefore less burst loss may occur However, this scheme requires extra one-hopfault notification One possible problem is that, if the network traffic load is alreadyvery heavy, distributed deflection may have a negative impact as it may deterioratethe congestion condition all over the network

While the restoration based on the neighbor deflection may make the deflected loadmore balanced, it suffers from an extra one-hop fault notification delay that will result

13.6 Recovery and Restoration 271

in greater overall burst loss because of bigger restoration loss Therefore, a more

efficient algorithm is a combination of local deflection and neighbor deflection, i.e.,

the affected bursts are deflected locally until the adjacent nodes receive the fault

notification At that time the affected bursts will be deflected in a distributive way

We name this scheme distributed deflection

One interesting observation from scheme 2 is that the capacity of the links between

the headend (node 2) of the faulty link and its adjacent nodes (node 1) will not

be utilized if affected bursts are deflected at adjacent nodes Therefore, we define a

distribution ratio, , to determine the portion of affected bursts that will be deflected

at the adjacent nodes That is, after the adjacent nodes receive the fault notification,

 affected bursts will be deflected distributively and 1– affected bursts will be

forwarded to the headend node of the faulty link to be deflected locally

With a different value of ∈ 01, we have a different variance of the distributed

restoration scheme When= 0, it is equivalent to scheme 1, local deflection-based

restoration When = 1, it becomes scheme 3, the distributed deflection-based

restoration We use  distributed deflection to denote the generalized distributed

deflection mechanism We also note that using  introduces only a tiny amount

of management complexity in the adjacent nodes We expect that there exists an

optimal value of that makes the affected bursts deflected in a most balanced way

such that the minimum burst loss can be achieved

Here are defined three restoration classes with different priority values for all bursts

When a burst is generated at the edge node, it will be assigned a priority value in its

control packet according to its restoration QoS requirement Bursts in the highest

class will pick up best from the local and neighbor deflection restoration schemes

during different restoration periods The local deflection will be chosen during the

fault notification period because of its shorter fault notification time And the local

and neighbor deflection scheme with shorter alternative route length (number of

hops in this paper) during the deflection period will be chosen because of its lower

backup capacity requirement (and possible lower average burst loss probability)

Bursts in the middle class will be restored via neighbor deflection, and the bursts in

the lowest class will be dropped until the global forwarding table update is finished

In this way, the lower class bursts do not compete for the bandwidth with the higher

class bursts in the deflected alternative routes

Figure 13.6 depicts the simulation results on a NSF topology The y-axis represents

the overall burst loss probability under medium traffic during the three network

operational phases around a link failure In the x-axis, 1 represents the period before

the link failure, 2 represents the period before a global forwarding table update and

only the proposed fast restoration schemes are in place, and 3 represents the period

after the global forwarding table update We observe that the burst loss probability

is very low for both phases 1 and 3, though it is actually a little bit higher in phase

3 due to the reduction in the network capacity

0.1 0.11 0.12 0.13 0.14 0.15 0.16

Global update Local deflection Neighbour deflection Distributed deflection

Restoration phase

Figure 13.6. Restoration performance

However, the loss probability could increase significantly in phase 2 Relying only

on the forwarding table update would incur very high burst loss in this phase.Nonetheless, the loss probability increases only moderately when the proposed fastrestoration schemes are used in this phase Among the four restoration schemes,distributed deflection shows the best performance (almost no extra burst loss)followed by local deflection and neighbor deflection Global routing update incursthe highest burst loss Specifically, the improvements from using the three fastrestoration schemes over the global forwarding table update are 24.3%, 20.1%, and10.5%, respectively

The  distributed deflection and class-based QoS restoration schemes are built

based on the performance differentiation of above four basic restoration schemes,therefore they can be directly built into the fault management module to providedifferentiated restoration service We note that the above schemes are studied for OBSnetworks, but they could be directly extended to other packet-switched networks

13.7 INTEGRATED FAULT MANAGEMENT

Except for the unreliable fault detector, the current Globus toolkit does not provideother fault tolerant mechanisms There exist a number of different fault handlingmechanisms for Grid applications to respond to system/component failures [37]:

• fail-stop: stop the application;

• ignore the failure: continue the application execution;

• fail-over: assign the application to new resources and restart;

• migration:replicationandreliablegroupcommunicationtocontinuetheexecution

References 273

Fail-over and migration are the two choices for fault-tolerant computing and can

be achieved either in a deterministic way, in which the backup compute resourceand network route are precomputed along with the primary, or in a dynamic way,

in which the backup is dynamically discovered after the failure occurrence anddetection We note that this could be coupled with the connection protection andrestoration mechanism in the network context discussed in Section 13.6

At the system level, performance monitoring, data analysis and fault detection,fault recovery will form a feedback loop to guarantee the system performance andavailability The complexity comes from the fact that performance/reliability data can

be collected in different timescales from different layers of the Grid networks (layer1/2/3 and application) Therefore, a Grid platform needs to provide different levels

of fault tolerance capability in terms of timescale and granularity A large portion

of the fault management activities are conducted within the Grid infrastructure andtransparent to the end-user applications

From the point view of applications, we identify the following failure scenariosand fault handling mechanisms to explicitly integrate the Grid service migration withthe network protection or restoration:

• Fail-over or migrate within the same host(s) The application requires dedicated or

shared backup connections for the primary connections (either unicast or cast) The applications can also specify the network restoration time requirement

multi-• Fail-over or mitigate to different host(s) This mechanism is more generic in the

sense that the application requires redundant Grid resource/service In the casethat the primary Grid resource fails, the Grid middleware can reset the connec-tions to the backup Grid resource/service and jobs can be migrated directly fromthe failed primary resource to the backups

Further studies are under way for the aforementioned cases in terms of dynamicconnections setup and network resource allocation

REFERENCES

[1] F Cappello (2005) “Fault Tolerance in Grid and Grid 5000,” IFIP Workshop on GridComputing and Dependability, Hakone, Japan, July 2005

[2] K Birman, “Like it or Not, Web Services are Distributed Objects!”, white paper

[3] V Paxson, G Almes, J Mahdavi, and M Mathis (1998) “Framework for IP PerformanceMetrics.” RFC 2330, May 1998

[4] R Hughes-Jones, P Clarke, and S Dallison (2005) “Performance of Gigabit and 10 GigabitEthernet NICs with Server Quality Motherboards,” grid edition of Future Generation Computer Systems, 21, 469–488.

[5] M Mathis and M Allman (2001) “A Framework for Defining Empirical Bulk TransferCapacity Metrics,” RFC 3148, July 2001

[6] Source code and documentation of iperf is available at http://dast.nlanr.net/Projects/ Iperf/.[7] G Almes, S Kalidindi, and M Zekauskas (1999) “A One-way Delay Metric for IPPM.” RFC

2679, September 1999

[8] G Almes, S Kalidindi, and M Zekauskas (1999) “A Round-trip Delay Metric for IPPM.”RFC2681, September 1999

[9] T Ferrari and F Giacomini (2004) “Network Monitoring for Grid Performance tion.”Computer Communications, 27, 1357–1363.

Optimiza-[10] S Andreozzi, D Antoniades, A Ciuffoletti, A Ghiselli, E.P Markatos, M Polychronakis,and P Trimintzios (2005) “Issues about the Integration of Passive and Active Moni-toring for Grid Networks,” Proceedings of the CoreGRID Integration Workshop,November 2005

[11] Source code and documentation of cricket is available at http://people.ee.ethz.ch/∼oetiker/webtools/mrtg/

[12] Multi Router Traffic Grapher software and documentation, http://people.ee.ethz.ch/∼oetiker/webtools/mrtg/

[13] “Grid network Monitoring: Demonstration of Enhanced Monitoring Tools,” DeliverableD7.2, EU Datagrid Document: WP7-D7.2–0110-4-1, January 2002

[14] “Final Report On Network Infrastructure And Services”, Deliverable D7.4, EU DatagridDocument Datagrid-07-D7-4-0206-2.0.doc, January 26, 2004

[15] EGGE JRA4: “Development of Network Services, Network Performance Monitoring –e2emonit,” http://marianne.in2p3.fr/egee/network/download.shtml

[16] Internet2 end-to-end performance initiative http://e2epi.internet2.edu/web traceroute.Home page for web based traceroute http://www.traceroute.org/

[17] A large collection of traceroute, looking glass, route servers and BGP links traceroutemay be found at http://www.traceroute.org/

[18] J Boote, R Carlson, and I Kim NDT source code and documentation available athttp://sourceforge.net/projects/ndt

[19] R.L Cottrell and Connie Logg (2002) “A new high performance network and applicationmonitoring infrastructure.” Technical Report SLAC-PUB-9202, SLAC

[20] IEPM monitoring page, http://www.slac.stanford.edu/comp/net/iepm-bw.slac ford.edu/slac_wan_bw_tests.html

stan-[21] Source code and documentation of thrulay is available at http://people.internet2.edu/∼shalunov/thrulay/

[22] C Dovrolis, P Ramanathan, and D Moore (2001) “What do Packet Dispersion TechniquesMeasure?,” IEEE INFOCOM

[23] Source code and documentation of pathload is available at http://www.cc.gatech.edu/fac/Constantinos.Dovrolis/pathload.html

[24] Source code and documentation of pathchirp is available at http://www.spin.rice.edu/Software/pathChirp/

[25] Source code and documentation of bbftp is available at http://doc.in2p3.fr/bbftp/.[26] A Hanushevsky, source code and documentation of bbcp is available at http://www.slac.stanford.edu/∼abh/bbcp/

[27] L Cottrell (2002) “IEPM-BW a New Network/Application Throughput PerformanceMeasurement Infrastructure,” Network Measurements Working Group GGF Toronto,February 2002, www.slac.stanford.edu/grp/scs/net/talk/ggf-feb02.html

[28] R Hughes-Jones, source code and documentation of UDPmon, a tool for investigatingnetwork performance, is available at www.hep.man.ac.uk/∼rich/net

[29] Endace Measurement Systems, http://www.endace.com

[30] J Hall, I Pratt, and I Leslie (2001) “Non-Intrusive Estimation of Web Server Delays,”Proceedings of IEEE LCN2001, pp 215–224

[31] Web100 Project team Web100 project, http://www.web100.org

[32] B Tierney, NetLogger source code and documentation available at didc.lbl.gov/NetLogger/

http://www-[33] “Telecommunications: Glossary of Telecommunication Terms”, Federal Standard 1037C,August 7, 1996

Fault-[36] H Jin, D Zou, H Chen, J Sun, and S Wu (2003) “Fault-tolerant Grid Architecture andPractice,”Journal of Computing Science and Technology, 18, 423–433.

[37] P Stelling, C DeMatteis, I Foster, C Kesselman, C Lee, and G von Laszewski (1999)

“A Fault Detection Service for Wide Area Distributed Computations,”Cluster Computing,

2(2), 117–128

[38] O Bonaventure, C Filsfils, and P Francois (2005) “Achieving Sub-50 onds Recovery upon BGP Peering Link Failures,” Co-Next 2005, Toulouse, France,October 2005

Millisec-[39] J Lang, “Link Management Protocol (LMP),” IETF RFC 4204

[40] M Médard and S Lumetta (2003) “Network Reliability and Fault Tolerance”,Wiley clopedia of Engineering (ed by J.G Proakis), John & Wiley & Sons Ltd.

Ency-[41] L Valcarenghi (2004) “On the Advantages of Integrating Service Migration and GMPLSPath Restoration for Recovering Grid Service Connectivity”, 1st International Workshop

on Networks for Grid Applications (gridNets 2004), San Jose, CA, October 29, 2004.[42] D Papadimitriou (ed.) “Analysis of Generalized Multi-Protocol Label Switching (GMPLS)based Recovery Mechanisms (including Protection and Restoration),” Draft-ietf-ccamp-gmpls-recovery-analysis-05

[43] Y Xin and G.N Rouskas (2004) “A Study of Path Protection in Large-Scale OpticalNetworks,”Photonic Network Communications, 7(3), 267–278.

[44] J L Marzo, E Calle, C Scoglio, and T Anjali (2003) “Adding QoS Protection in Order toEnhance MPLS QoS Routing,” Proceedings of IEEE ICC ’03

[45] S Rai and B Mukherjee (2005) “IP Resilience within an Autonomous System: CurrentApproaches, Challenges, and Future Directions”, IEEE Communications Magazine,

Manage-[48] Y Xin, J Teng, G Karmous-Edwards, G.N Rouskas, and D Stevenson (2004) “A NovelFast Restoration Mechanism for Optical Burst Switched Networks”, Proceedings of theThird Workshop on Optical Burst Switching, San Jose, CA, USA, October 2004

Currently, many communities are examining requirements and designs fornext-generation networks, including research communities, standards associations,technology developers, and international networking consortia Some of these newdesigns are implemented in early prototypes, including in next-generation opencommunication exchanges This chapter describes the general concept of such anexchange, and presents the services and functions of one that would be based on afoundation of optical services.

Grid Networks: Enabling Grids with Advanced Communication Technology Franco Travostino, Joe Mambretti,

14.2 CREATING NEXT-GENERATION NETWORK SERVICES AND

in Internet architecture For example, it would be highly distributed

Even without the Grid as a driving force to motivate the creation of communicationsinfrastructure with these characteristics, architectural designs based on these princi-ples are motivated by other dynamics Major changes are required in the methods

by which communications services and infrastructure are designed and provisioned.These changes are motivated by the explosive growth in new data services, by theadditional communities adopting data services, and by the number and type ofcommunication-enabled devices Soon services will be required for 3 billion mobilegeneral communication devices

Today, there is a growing need to provide seamless, ubiquitous access, at anytime from any location and any device In addition to general communicationdevices, many more billions of special-purpose communications-enabled devices areanticipated: sensors, Radio Frequency Identification (RFID) tags, smart dust, nano-embedded materials, and others All of these requirements are motivating a transfor-mation of the design of core network infrastructure The legacy architectural modelsfor services design and provisioning cannot meet the needs of these new require-ments A fundamentally new approach is required The guiding principles for thatnew approach will be based on those that informed Internet architecture The nextsection presents some of these principles, which are key enablers for next-generationcommunication services and facilities

In previous chapters, the important of the end-to-end principle in distributed tructure design was discussed Its basic premise is that the core of such distributedinfrastructure, including networks, should be kept as simple as possible, and theintelligence of the infrastructure should be placed at the edge This principle is nowalmost universally accepted by all networking standards bodies Recently, the ITUformally endorsed the basic principles of the Internet design through its Study Group

infras-13 recommendation for a Next-Generation Network (NGN) [1] Balancing the need

to maintain the end-to-end principle and to create new types of “intelligent” corefacilities is a key challenge

14.2 Creating Next-generation Network Services and Infrastructure 279

After 35 years of increasing Internet success, communications standards

organiza-tions have now almost universally endorsed a architecture founded on the

communi-cation of packet-based data The IETF has demonstrated the utility of this approach

The IEEE has established initiatives to create layer 2 transport enhancements to better

support based services The ITU-T NGN recommendation specifies “A

packet-based network able to provide telecommunication services and able to make use of

multiple broadband, QoS-enabled transport capabilities.” The NGN Release 1

frame-work indicated that all near-term future services are to be based on IP, regardless of

underlying transport

Earlier chapters have stressed the importance of abstracting capabilities from

infras-tructure The design of the Internet has been based on this principle Attaining this

goal is being assisted by the general recognition that legacy vertical infrastructure

stacks, each supporting a separate service, must be replaced by an infrastructure that

is capable of supporting multiple services This design direction implies a

funda-mentally new consideration of basic architectural principals The ITU-T has begun

to formally adopt these concepts One indication of this direction is inherent in

the ITU-T NGN recommendation that “service-related functions” should be

“inde-pendent from underlying transport related technologies.” These recommendations

indicates that the design should enable “unfettered access for users to networks,”

through open interfaces [1]

Also, the ITU-T is beginning to move toward a more flexible architectural model

than the one set forth in the classic 7 layer Open Systems Interconnect (OSI) basic

reference model [2] This trend is reflected in the ITU-T recommendations for the

functional models for NGN services [3] This document notes that the standard

model may not be carried forward into future designs The number of layers may

not be the same, the functions of the layers may not be the same, standard attributes

may not be the same, future protocols may not be defined by X.200, e.g., IP, and

understandings of adherence to standards may not be the same

Another important architectural design principle for any future network is

self-organization

The Internet has been described as a self-organizing network This description has

been used to define its architecture in general, because many of its individual

proto-cols are designed to adapt automatically to changing network conditions without

having to rely on external processes This term has also been used to describe

the behavior of individual protocols and functions, such as those that automatically

adjust to changing conditions Because the Internet was designed for an unreliable

infrastructure, it can, therefore, tolerate uncertain conditions better than networks

that depend on a highly reliable infrastructure

The principle of self-organization for networks is particularly important forscalability in a ubiquitous communication infrastructure containing an extremelylarge number of edge devices, especially mobile edge devices The only method forensuring reliable services in this type of global distributed environment is through anarchitecture that enables edge devices to be self-sufficient, to be context aware, self-configuring, adjustable to fault conditions and highly tolerant of instability Theseattributes enable individual services, objects, and processes to automatically adjust

to changing circumstances Recently, progress has been made in identifying mental issues in requirements and potential solutions for autoconfiguration forcomponents in large-scale dynamic IP networks [4]

funda-Another motivation for developing architecture for self organizing networks, such

as ad hoc networks, is to enable communication services in areas with minimalinfrastructure or unstable infrastructure

Another key principle for next generation communications architecture is ensuringcapabilities for distributed service creation Traditionally, service providers have care-fully defined services, and have then designed the technology and infrastructure tosupport those services, with an understanding that they will remain in place forlong periods of time However, in an environment of rapid change, in which newservices can be conceptualized continually, this approach is no longer feasible Thearchitecture must anticipate and support continuous changes in services

Also, the architecture should be designed to provide support for distributed servicecreation It should provide the tools and functionality for services creation to broadcommunities and allows those communities to create, manage, and change theirown services

14.3 LARGE-SCALE DISTRIBUTED FACILITIES

These design principles are inherent in the basic architecture of the Internet, andthey are reflected in Grid design principles Currently, these principles are used

to design new types of large-scale distributed infrastructure to provide support forGrid services The communications infrastructure for these environments does notresemble a traditional network Within this environment, it is possible either to accept

14.4 Designs for an Open Services Communications Exchange 281

predefined default services or to create required services through dynamic, ad hoc

processes These environments are used to develop and implement customized Grid

services and specialized communication services For example, in some Grid

envi-ronments, communications services are used as a large-scale distributed backbone

to support specialized Grid processes

This architectural approach provides highly distributed communities with direct

access to resources, toolsets, and other components, including complete resource

management environments, that allow them to create new services and to change

existing services Network services, core components, resources such as lightpaths,

layer 2 channels, cross-connections, and even individual elements can be identified

and partitioned into separate domains and provided with secure access mechanisms

that enable organizations, individuals, communities, or applications to discover,

interlink and utilize these resources Accessible resources can even include basic

optical components, which can be managed and controlled by highly distributed

edge processes

14.4 DESIGNS FOR AN OPEN SERVICES COMMUNICATIONS

EXCHANGE

The architecture described here is implemented within many Grid environments As

these environments were developed, it was recognized that an important component

within this type of large scale distributed facility is an open, neutral communications

exchange point The design of such exchanges constitutes a significant departure

from that of traditional carrier communications exchanges

Today, almost all data networks exchange information at Internet exchange

facil-ities Almost all of these facilities have an extremely limited range of services, and

limited peering capabilities – primarily they provide capabilities for provider peering

at layer 3 and, for some under restrictive policies, layer 2 Also, traditional exchanges

are oriented almost exclusively to provider-to-provider traffic exchanges

The international advanced networking research community is developing new

designs for data communication exchange points that provide for many more types

of services, including Grid services and layer 1 peering capabilities These designs are

currently being developed and implemented in prototype by research communities

An open Grid service exchange provides mechanisms for nonprovider services

exchange through subpartitioning of resources For example, it can allow segments of

resources along with selected control and management processes for those resources

to be allocated to specific external entities

A basic design objective for an open Grid services exchange is the creation of a facility

that fully supports interactions among a complete range of advanced Grid services,

within single domains and across multiple domains Because Grid services are defined

as those that can be constructed by integrating multiple resources, advertised by

common methods, this type of exchange provides capabilities for resource discovery,assembly, use, reconfiguration, and discard.

Unlike existing exchanges, this type of exchange will provide methods for ering a complete range of services that can be utilized and manipulated as Gridresources, including options for integration with other distributed Grid resources.This capability provides far more options for fulfilling requirements of services andapplications than are possible with traditional exchanges

discov-This type of exchange fulfills the design principles described earlier, especiallythose related to decentralization Using such an exchange, there is no need torely on a carrier cloud for transport Grid collaborations will be able to managededicated communications resources and services within an environment that theycan customize directly in accordance with their requirements

The infrastructure design used for traditional exchanges is specifically oriented to

a narrow set of limited services As a result, such exchanges require major physicalprovisioning to add new services, enhance existing services, or to remove obsoleteservices The new type of exchange described here will provide a flexible infrastruc-ture with many component resources that can be dynamically combined and shapedinto an almost unlimited number of services

The abstraction layer made possible by the Grid service-oriented approach toarchitecture, combined with new types of network middleware and agile opticalcomponents, allows provisioning to be undertaken through high-level signaling,protocols, and APIs rather than through physical provisioning

These exchanges have the following characteristics

• They are based on a services-oriented network model, and they provide facilities

to support that model, incorporating capabilities that provide direct access tonetwork services Resources are advertised through standard, common servicesmodels

• They are provider and services neutral They allow any legitimate tion service entity to utilize internal resources Because they are based on aservices-oriented architecture, services provisioning, access, and control can beundertaken by multiple service providers, using distributed management andcontrol systems

communica-• They provide a significantly expanded set of services, both a wide range ofstandard, advanced, and novel communication services and other data-orientedservices, including Grid services

• They support ad hoc dynamic services and resource provisioning

• They are highly decentralized They enable multiple capabilities for autonomouspeerings, including ad hoc peerings, among client constituencies, both serviceproviders and other communities such as individual organizations

14.5 Open Grid Optical Exchanges 283

• They enable peering to take place at any service or communications level,

including those for data transport services at layer 1

• They provide capabilities for extensive integration and service concatenation at

multiple levels, e.g., layer 1, layer 2, and layer 3, and among multiple services

and resources

• They enable such integration across multiple independent domains

• They allow for direct peer-to-peer connections from specifically designated sites,

and enable those sites to established unique shared network services

Currently, a new architecture for next-generation open exchange points based

on these concepts is being designed, developed, implemented in prototype, and

in some cases being placed into production The most advanced designs for these

exchange points incorporate innovative methods of providing a services foundation

based on dynamic lightpath provisioning Unlike traditional exchanges points, these

exchanges will provide accessible layer 1 transport services as foundation blocks

for all other services This capability will be a key foundation service within such

exchanges

14.5 OPEN GRID OPTICAL EXCHANGES

Because optical services layers within these exchanges will provide for key

foun-dation capabilities, it is appropriate to term these types of exchanges “open grid

optical exchanges.” As noted, these facilities represent a significant departure from

traditional exchanges

There are currently three types of Internet exchanges [5,6]:

(1) LAN-based Internet exchanges These are the most common exchanges, and they

typically implement layer 2 switches at their core, although some exchanges are

distributed This is the only stateless exchange Blocking is possible if multiple

peers want to send traffic to the same peer at the same time

(2) ATM-based Internet exchanges These are statefull exchanges, with Permanent

Virtual Circuits (PVCs) at their core If Variable Bit Rate (VBR) circuits are used,

there is no guaranteed congestion-free transmission The use of Constant Bit

Rate (CBR), on the other hand, results in very inefficient use of resources and

poor scalability

(3) MPLS-based Internet exchanges MPLS exchanges are both stateful and packet

based Though the service may be congestion free, it requires a lookup at the IP

routing table at the ingress edge Label Switching Router (LSR) This approach

results in a relatively expensive operation for very-high-bandwidth datastreams

Because of the properties of these types of exchanges, they cannot provide many

services required by Grids These services are not sufficient to support even a few very

high-bandwidth flows that do not require routing Either it is technically impossible(for LAN-based Internet exchanges) or it yields unnecessary and costly overhead(for ATM- and MPLS-based Internet exchanges) Also, such exchanges provide nocapabilities for external entities to implement and customize services dynamically.

This section describes the rationale and model for an exchange that can be termed

an “open Grid optical exchange.” There are multiple reasons for developing this type

of exchange, including those that are derived from the design principles described

in earlier sections This section also describes the basic interfaces and services thatsuch an optical exchange may provide

The case for an open Grid optical exchange is derived directly from a number ofkey requirements, including the need for direct access to services at lower layers

A growing requirement has been recognized for transport services that allow datatraffic to remain at lower layers (layer 2, layer 1) and not be converted, for examplefor routing through other layers Also, there are reasons based on requirementsfor service quality, and others that relate to taking advantage of multiple newoptical innovations First, optical connection-oriented solutions can ensure highquality, e.g., they can be guaranteed congestion free This feature eliminates anyneed to impose complex QoS technologies In addition, connection-oriented linksallow users to use more efficient, but TCP-unfriendly, transport protocols Suchflows would be problematic in a common layer 3 exchange Although other tech-nologies also attempt to offer high-quality services with connectionless transport,such as Multi-Protocol Label Switching (MPLS) and Optical Burst Switching (OBS),these services do not match the performance of optical services for many requiredGrid services

Second, there is an economic argument for this type of exchange – oriented technologies will always remain an order of magnitude cheaper becausethere is no need to look into headers at the datastream itself Even as traditionalfacility components are becoming less expensive, optical components are also contin-uing to fall in price, and the relative price ratio is at least remaining constant(although there is an argument that optical component prices are falling faster).Many new technologies are being developed [7], such as Optical Add-Drop Multi-plexers (OADM)s and advanced photonic devices that will provide powerful newcapabilities to these environments Furthermore, the long-term forecast for opticalnetwork components is that they will become increasingly less expensive, includingsuch key devices as optical shift registers [8] If current research projects developoptical routers, this type of facility would be able to take immediate advantage of theircapabilities

connection-Third, it is more efficient to keep traffic that does not need any high-levelfunctionality at the lowest layer possible In particular, this will apply to relatively long(minutes or more) and high-bandwidth (gigabps or more) datastreams between alimited number of destinations Examples of these requirements are data replicationprocesses and instrumentation Grids where huge datastreams must be collected at acentral place (like the eVLBI project [9]) The packets in these streams do not need

14.5 Open Grid Optical Exchanges 285

routing One knows the data has to go from this source to that destination In general,

this is called the service approach: apply only the required services, nothing more

Fourth, even with a limited number of destinations, it is very likely that

data-intensive traffic streams will traverse multiple network domains Considering

scal-ability issues, it is likely that the most cost-effective means for peering among the

domains will be at optical peering locations

The next sections of this chapter describe the concept of an optical exchange, and

they provide an example of how such an exchange may be implemented Different

applications require different transport services, including layer 1 services [10],

ranging from layer 0 to layer 3 There could be many types of optical exchanges

For example, one could be a trivial co-location site that provides no more

function-ality other than rack space and the ability for providers to have bilateral peerings

with other providers at the same co-locations Others can be highly complex with

extensive services Instead of describing four or more different potential exchanges,

a generic architecture for an exchange is described, consisting of a “black box” with

advertised interfaces and services (Figure 14.1)

The “black box” interfaces can implement different services and protocols For

example, one interface may be used to carry undefined traffic over 32 wavelengths

using Dense Wavelength Division Multiplexing (DWDM), while an other interface may

carry only one signal at 1310 nm, which carries SDH-framed traffic A third interface

may be LAN-PHY-based Ethernet, where traffic must reside in the 145.146.100.0/24

subnet

The basic functionality of any exchange is to transport traffic from one

organiza-tional entity to another, usually a provider For an optical exchange, this requirement

implies that the main function is providing cross-connections among interfaces,

forming circuits The different interfaces require the optical exchange to extract a

given signal from a provider circuit and inject it in another provider circuit using

some form of multiplexing

Optical Exchange

Interfaces

Services External

connections

Figure 14.1. Components of an optical exchange

For example, if one provider uses DWDM with 50 GHz spacing, and another uses

100 GHz spacing, it will probably be necessary to first demultiplex (demux) andthen multiplex (mux) the signal again If a signal enters at one layer and leaves at

a different layer, the optical exchange effectively acts as an “elevator,” lowering orraising traffic to different layers For example, if a signal comes in as a wavelengthusing DWDM, and is injected in a vLAN, the signal is elevated from layer 1 tolayer 2

The first type of service can be described as providing cross-connections Thesecond type of service would be providing capabilities for muxing and demuxingbitstreams Other more complex services may also be required, such as aggregation

of traffic using a switch and optical multicast and store-and-forward services Anygiven optical exchange may provide only a subset of all possible services However, atsuch an exchange, if other services are not offered by the exchange itself, they may beoffered by a service provider that is connected at the exchange, as Figure 14.2 shows

An important point is that there is no technical obligation to place one set ofservices in the optical exchange domain (theinternal services) and another set in a

services domain (theexternal services) The decision as to which service should be

an internal and which external service is a no technical one determined by other,external, considerations

As noted, an optical exchange may accept one or more types of interfaces The listbelow describes common interfaces

At layer 0, only single-mode fibers are considered Specifically ignored hereare multimode fiber and electrical (UTP) carriers, because currently within opticalexchanges such as NetherLight [11] and StarLight [12], there is a trend toward single-mode fiber and away from multimode fiber and UTP There are three likely causes forthis trend First, optical cross-connects with MEMS switches absorb light at 800 nm,and thus cannot deal with multimode fiber Secondly, DWDM use is increasing, and

Optical exchange

External connections

External services

Services domain

Internal service

Figure 14.2. Internal and external services

14.5 Open Grid Optical Exchanges 287

DWDM is possible only with single-mode fiber Third, single-mode has a wider range

of operation (a few hundred kilometers) than multimode (about 2 km)

At layer 1, each fiber may either carry a single bitstream within the 1260–1675 nm

range or multiple bitstreams of data using separate DWDM wavelengths (lambdas)

Each bitstream (lambda) can use one of the following sampling and framings:

(1) a bitstream, at a certain wavelength, with up to 10.1 GHz or up to 40 GHz

sampling, without other known properties;

(2) a bitstream, at a certain wavelength, with SONET/SDH framing, either OC-48

(2.5 Gbps), OC-192 (10 Gbps), or OC-768 (40 Gbps);

(3) a bitstream, at a certain wavelength, with 1 Gbps Ethernet;

(4) a bitstream, at a certain wavelength, with LAN PHY Ethernet;

(5) a bitstream, at a certain wavelength, with WAN PHY Ethernet

Ethernet is a major service that scales from the local to the WAN scale Each fiber

typically carries one wavelength, or it carries 32 or 64 wavelengths though DWDM

The speed is typically either 1 Gbps or 10 Gbps LAN-PHY variant

From the regional to global scale, typically a single SONET/SDH or WAN-PHY signal

is used per carrier The SONET speed may be OC-48 (2.5 Gbps), OC-192 (10 Gbps),

or OC-768 (40 Gbps) The reason that TDM (with SONET or SDH) is used on a

world-wide scale is that providers of transoceanic links require customers to comply with

SONET/SDH framing Note that WAN-PHY is compatible with SONET framing and also

with SONET optical equipment [13] Thus, WAN-PHY can be used on transoceanic links

There are many ways to multiplex signals in one carrier using DWDM, because

of the variety of wavelength bands and channel spacings Similarly, not all vendors

encapsulate Ethernet in the same way over a SONET/SDH connection An optical

exchange must have this type of information to determine the correct service to use

Of course, if no demultiplexing is required, then the optical exchange does not need

this information

A major issue when connecting two fibers together is to correctly tune the power

levels Especially with single-mode fibers, a single speck of dust can ruin signal

strength When making and breaking connections between two carriers dynamically,

using an optical cross-connect, the signal strength should have a known, predefined

value when leaving the optical exchange domain

The protocol used on each circuit either is unknown to the optical exchange or

may be specified by:

(1) IP over Ethernet (note that if all interfaces are of this type, the optical exchange

would be reduced to a LAN-based Internet exchange);

(2) Ethernet implemented with 802.1q (vLAN tagging) [14]

Most end-to-end connections are full-duplex, even if the application does not

require dedicated bandwidth in both directions Most applications do not support a

one-way connection, but a relatively easy way to optimize resource use is to support

asynchronous connections At a minimum, an optical exchange must be aware if an

interface is always full-duplex or supports one-way connections

14.5.5 OPTICAL EXCHANGE SERVICES

This section describes possible services for an optical exchange However, not all

of these services have to be offered by the optical exchange itself: some may beoutsourced to a services provider connected to the optical exchange, or may not beoffered at all The abbreviations are used in the service matrix below

(a) Cross (connect) Given two interfaces of equal type, be able to make a

cross-connect between these interfaces Typically, this should be done in a user- ortraffic-initiated way by a software control plane However, here this limitation

is not imposed

(b) Regenerate Amplify or attenuate the power levels, to match a certain output

power level; amplify and reshape; or reamplify, reshape, and retime (3R).(c)  convert Wavelength conversion, by regenerating the signal or by physically

altering the wavelength Regenerating, using tunable transponders, may allownetwork engineers to monitor the Bit Error Rate (BER) of a signal, but requiresthe regenerator to understand the modulation and possible framing of thesignal

(d) WDM mux/demux Multiplex wavelengths of different color into a single carrier,

and demultiplex different signals in a single carrier into many separate fibers.This process does not need to convert the wavelengths itself An advanceddemultiplexer may first demux the signal into sets of wavelengths, called wave-bands, before the wavebands are demuxed into individual wavelengths [15].Also, not all multiplexers are compatible, since different optical bands andchannel spacings may be used

(e) (Optical) multicast The ability to duplicate an optical signal as-is Of course,

this can only be done one-way Possible uses include visualization

(f) TDM mux/demux The ability to extract an Ethernet signal from a SONET or

SDH carrier or to insert one or more Ethernet connections in a SONET or SDHcarrier It should be noted that not all SONET/SDH multiplexers are compatible.(g) SONET (switching): The ability to combine and switch SONET or SDH circuits,

without knowing the contents of the circuits

(h) Aggregate There may be different needs for an Ethernet switch in an optical

exchange First, it allows aggregation of traffic, although this may cause tion

conges-(i) (Ethernet) conversion A second use for an Ethernet switch is the conversion

between different types of framing The most useful conversion is perhaps theability to convert between LAN-PHY and WAN-PHY

(j) vLAN encap/decap Ethernet encapsulation A third use for an Ethernet switch

is the encapsulation of traffic in a vLAN trunk [14] This allows the combining

of different, separable datastreams on a single link

The combination of DWDM support, wavelength conversion, and cross-connectswill effectively provide for an OAD) facility Multiple services may be combined in asingle device