> Manuscript ID: 117376<

1

Abstract—Technological improvements, like EDFAs and

WDM, have revolutionized optical transmission, boosting the development and worldwide deployment of the Internet. Similar progress has not yet impacted the routing layer, as current IP routing technology struggles to deliver the necessary bandwidth at competitive costs. Hybrid electro-optical architectures, where dynamical optical circuit switching is combined with legacy packet routing, have been introduced in the past few years as promising solutions to reduce costs at the IP layer, and to deliver new revenue-generating services and applications. Most of the architectures currently proposed have focused on end-to-end lightpath provisioning, coordinated through a centralized management plane. In this paper we propose Optical IP Switching (OIS), a hybrid electro-optical network architecture that combines IP routing and wavelength switching, using a distributed decision-making process. We report technical and economic analysis, based on simulations obtained through real traffic traces and routing tables, that compare our distributed model to an opaque IP-over-WDM and centralized, transparent overlay models. We also report testbed results that analyze the effect of dynamic transparent switching operation on the TCP and UDP transport protocols. The results obtained show that distributed provisioning based on local traffic analysis can be feasible, although more work is required to evaluate the efficacy of optimization techniques on distributed network models.

Index Terms— cross-layer design, flow-routing, distributed optical provisioning, IP-optical architecture, wavelength switching.

I. INTRODUCTION N the past fifteen years we have seen the Internet evolving from a niche network connecting academics and

research institutes to the main global information and entertainment network. Such exponential evolution was made

Manuscript received September 29, 2009. This material is based upon

work supported by Science Foundation Ireland under Grant no. 03/CE3/I405 as part of the CTVR Telecommunications Research Centre at Trinity College Dublin, Ireland.

Marco Ruffini is with the CTVR Telecommunications Research Centre at Trinity College Dublin, Ireland. ( Phone: 00353-1-896-8441; e-mail: ruffinm@tcd.ie).

Donal O’Mahony is with the CTVR Telecommunications Research Centre at Trinity College Dublin, Ireland. (e-mail: Donal.OMahony@cs.tcd.ie).

Linda E. Doyle O’Mahony is with the CTVR Telecommunications Research Centre at Trinity College Dublin, Ireland. (e-mail: ledoyle@tcd.ie).

possible through the development of Wavelength Division Multiplexing (WDM) and Erbium-Doped Fiber Amplifier (EDFA) technologies, which, over ten years ago, started a deep revolution in optical networking. The former allows transmission of several channels in a single fiber, while the latter amplifies all the channels in a fiber, without performing electrical conversions. Combined together, they increased the overall bandwidth availability and drastically reduced the cost of data transfer, creating an ideal environment for the mass proliferation of the Internet. As a result, existing applications that have conventionally been delivered over different networks are being migrated to the Internet (this is the case for voice calls and TV), while new applications continue to emerge (peer-to-peer networking, interactive gaming, high definition IPTV, e-science applications, to mention only a few).

Although the WDM/EDFA evolution has made the optical transport layer capable of supporting the amount of data generated by these applications, equivalent dramatic progress has not yet occurred at the routing layer. As a result the network bottleneck has moved from the optical transport to the routing layer, as conventional electronic routers do not seem capable of offering a cost-effective solution to the increasing bandwidth demand. In the past few years the research community has challenged this issue by developing hybrid electro-optical architectures aiming at effectively bridging the gap between optical transport and electronic routing. The hybrid electro-optical solutions proposed over the years generally fall into two major categories: Optical Circuit Switching (OCS) and Optical Packet Switching (OPS). In the OCS approach, optical lightpaths are dynamically setup when needed, to transport data transparently between any two nodes in the network. In the OPS approach, data is routed packet-by-packet; the router processes the packet header electronically, determining the next hop node and activating the optical switch to route the packet payload in the optical domain. Other solutions (like for example Optical Burst Switching) fall in between these two categories. The work we present in this paper relates to optical circuit switching. We propose Optical IP Switching (OIS), a hybrid electro-optical routing architecture, originally introduced in [1], [2], which combines an enhanced electronic router and a transparent optical cross-connect, to dynamically create optical paths, redirecting part of the IP traffic load to the optical layer. The main novelty, compared to existing architectures, is that the distributed operations do not refer only to the control plane, but also to the decision-making mechanism. Optical paths are

Marco Ruffini, Donal O'Mahony and Linda Doyle, Member, IEEE

Optical IP Switching: a Flow-based Approach to Distributed Cross-layer Provisioning

I

> Manuscript ID: 117376<

2

automatically engineered following the analysis of local traffic at each node. This distributed approach, besides having advantages in terms of scalability and fault-tolerant design, is highly beneficial in the inter-domain, where each autonomous system makes decisions following proprietary network policies. We believe that the current inter-domain structure, which has provided the Internet with the heterogeneity and ubiquity that have contributed most to its exponential deployment, is the cornerstone that will persist in its near and far-future evolution. Therefore we think that a distributed approach, similar to that currently operated at the IP layer could also be applied at the optical level. The paper is structured as follows: after providing an overview of current approaches to dynamic provisioning of optical circuits, we describe, in section III, the concept of Optical IP Switching. In section IV we report simulation results of technical and economical studies comparing our technique to opaque IP-over-WDM and transparent end-to-end provisioning based on overlay model. In section V we describe the testbed implementation of the OIS architecture and report on the effects of dynamic switching of lightpaths on the Internet transport protocols (i.e., UDP and TCP). Finally, before concluding the paper, we summarize the main future implications of our work.

II. PREVIOUS WORK Dynamic provisioning at the optical layer is a research topic

that attracted the interest of many researchers and Internet service providers over the past few years. The new business model based on the information and service sectors brought about by the new economy has highly increased the dynamics and volatility of the global market. New Internet-based business and applications (for example Skype, Youtube, Facebook, high definition IPTV, to mention only a few) can emerge and grow up very quickly, at unpredictable rates (a recent example being the BBC iPlayer service, which has overloaded British service providers). Such highly dynamic activity directly translates into very dynamic bandwidth requests, for which current network capacity management methods, which generally operate over six to eighteen month periods, have become unsuitable. In addition, as average Internet traffic keeps growing at sustained rate, legacy hop-by-hop routing becomes a non cost-effective solution compared to the economical benefits that transparent bypass of the IP layer can bring [3], [4].

Aware of these issues, vendor and network operators have worked together to produce a new standard for the optical control plane. Their aim is to increase the automation and control that the operators have on their networks, developing one integrated control suite capable of operating at different layers of the protocol stack, from the IP down to the WDM physical layer. Three main international bodies were involved in the standardization process [5]. The ITU-T has worked on the ASON architecture, which describes the basic functional requirements of the optical network and defines the principles for the UNI (User-Network Interface) and NNI (Network-Network Interface). The IETF is developing the GMPLS

architecture which addresses link management, topology discovery, connection provisioning and the protection and restoration issues. Finally the OIF focuses its attention on the actual implementation of the control interfaces (UNI and NNI) defined in the ASON: their efforts culminated in the OIF World Interoperability Demonstrations, which over the years brought different vendors and carriers together for a practical demonstration of the interoperability between their equipments (the 2004 event was reported in [6]). Once completed and fully implemented these standards should accelerate and facilitate the provisioning of optical paths, by automating tasks like topology discovery and connection provisioning.

Different examples of optical control plane implementation are already widespread in experimental networks and research laboratories around the globe. In [7] for example a group of researchers from NTT describes their hybrid electro-optical architecture (dubbed the Hikari router), where electrical routers use GMPLS to create new optical paths, when the existing ones become congested. A similar approach was presented in [8], where hybrid electro-optical nodes set up and delete end-to-end optical paths dynamically, using distributed signaling protocols. A more futuristic model, called Optical Flow Switching (OFS) [9], is currently being researched at MIT. The idea is to create highly dynamic end-to-end lightpaths, requested by the users, to transport traffic flows from source to destination LANs. Such paths transparently cross different MANs and WANs, and are released when the transaction is completed. As optical paths are created end-to-end, no buffering facilities are required in the network; all the traffic is queued at the end-user while it waits for the requested path to become available.

Dynamic optical circuit switching is also used extensively in grid network architectures. The idea of grid networks was developed to support the interaction of high-end applications distributed around the globe that need to exchange information at ultra-high data rates: distributed computing, e-VLBI, High-energy physics, e-Health applications, to mention only a few. Reconfigurable optical networks seem to suit this concept very well, as dedicated optical lightpaths can be established on request to satisfy the large bandwidth demand of such applications. DRAGON [10] and MUPBED [11] are examples of optical architectures suitable for supporting large bandwidth requests, typical of grid networks. It is only with the OptIPuter [12] however that the grid concept is fully embedded into the architecture, with the idea of the Distributed Virtual Computer, where applications automatically negotiate the end-to-end bandwidth with the network, through operations which are completely transparent to the end user.

All projects above are based on dynamic reconfiguration of wavelength channels. Although the physical layer is not fully flexible, much research is currently focusing on increasing dynamic capability of wavelength switched networks. Once faultless wavelength reconfiguration will be achieved and its operation fully trusted by network operators more research activity will need to focus on how such flexibility can be exploited [13]. One of the drivers of automatic lightpath

> Manuscript ID: 117376<

3

triggering will be lightpath-on-demand, typical of business bandwidth services, where a customer that wants to inter-connect two distant premises (e.g., different branches of the same bank of company), can be directly billed for the service provided. Recent studies on global IP traffic however (reported in Fig. 1) show that the growth of consumer traffic will totally dominate over business traffic. IP flows generated by consumer applications are usually in the order of the Mbps to tens of Mbps, a granularity that is too fine for current technology to offer cost-effective wavelength switching (where the minimum unit of bandwidth is in the order of 10 Gbps and evolving towards 40 and 100 Gbps in the immediate future). Following such considerations, automatic lightpath triggering for traffic engineering purposes seems a much stronger driver for reconfigurable optical networking.

In this approach an automatic decision mechanism (for example at the IP layer), creates and deletes lightpaths based on the aggregated traffic generated by millions of consumer users. If we examine current traffic engineering operations however, we see that they are mainly human driven and therefore cumbersome, lengthy and only suitable for topology reconfiguration over relatively long time scales (weeks to months). This approach does not seem suitable to address the dynamics and unpredictability of traffic generated by novel Internet applications.

The OIS architecture we propose in this paper runs continuous adaptation of the wavelength topology by creating distributed and local bypasses of the IP layer. Such an approach is lightweight and scalable and allows following traffic variations over shorter time scales (tens of seconds to minutes).

Fig. 1. Global IP forecast for business and consumer traffic for the years 2005-2011 (source Cisco Systems Inc., 2007)

III. OPTICAL IP SWITCHING ARCHITECTURE Optical IP Switching is a technique we have developed that

creates and deletes optical cut-through paths in a distributed fashion in response to local analysis of IP traffic. The idea of flows bypassing the IP layer is inherited from IP switching, which was first introduced in the electrical domain as a method of combining Asynchronous Transfer Mode (ATM) and IP technologies [14]. This idea, originally conceived by Ipsilon Networks Inc., was then further developed by Cisco's Tag Switching, and finally standardized by the IETF as Multi-Protocol Label Switching (MPLS). In the OIS approach the

dedicated electronic circuits are substituted by wavelength switched paths. Switching data directly in the optical domain has important consequences. On one hand it can allow cost saving, as optical switch ports are data rate independent, and cost tens of times less than IP ports. On the other hand however, optical switching is operated at the wavelength granularity, which can become quite inefficient compared to the packet granularity offered by electronic routers and switches. In the next sections we explain how such issues are targeted by the OIS architecture.

A. The OIS optical router Fig. 2 gives a logical overview of the Optical IP

Switching architecture. The upper part of the diagram specifies the IP/OIS functional elements together with their logical interconnections. The OIS functions are closely integrated with the IP layer processes, with read/write access to the routing table and ability to avail of the routing engine to forward signaling messages. Both the OIS and IP protocols can influence the routing mechanism by modifying the entries in the routing table independently from each other. The advantage brought about by the close integration of IP routing and optical switching is twofold. Firstly, it correlates the routing at the IP and optical layers, introducing a mechanism to engineer together the two layers. Secondly, it guarantees full backward compatibility with the IP protocol, an important characteristic for the practical implementation of Internet architectures.

Fig. 2. Overall Optical IP Switching architecture: IP and OIS functions are closely integrated in the network layer

The optical interfaces allow the router to create and

terminate optical links, over which packets are transmitted and received. Some of these are used to create static default links to neighbors, through which normal IP traffic and control packets are exchanged. The remaining interfaces are used to accommodate the dynamic optical paths. The optical switch is

> Manuscript ID: 117376<

4

the element that physically links the node to the external world. On one hand it allows each interface to connect to any incoming or outgoing fiber; on the other hand it allows the transparent switch of an incoming to an outgoing fiber, creating optical bypasses of the IP layer. The switch is directly controlled by the dynamic allocation engine through a dedicated interface. In the figure, we have illustrated a fiber switch to clarify the interconnections of the optical elements. For a real implementation however, wavelength-selective switches (incorporating the WDM multiplexers) could be employed.

B. Traffic analysis and aggregation Traffic analysis is the first functional step of the Optical

IP Switching mechanism. While in the “Observation” state, every OIS node performs constant analysis of the IP traffic transiting the router, for example using one of the packet sampling mechanisms described in [15]. The information that needs to be collected includes: the interface from which the packet arrived, the output interface, which is selected by the router through the longest prefix matching algorithm, the payload size and the arrival time (time information, for example, can be processed to identify and estimate long lived IP flows [16]).

The basic idea, following the original electronic IP Switching approach, would be to create an optical path as soon as an IP flow of suitable size (dubbed “elephant flows” in the literature) is observed at the IP layer. However, this approach does not scale for Optical IP Switching, as the average elephant flow rate and optical link bandwidth differ by three to four orders of magnitude, making this method inefficient. Although this approach would be suitable for grid networks, where high end applications might easily occupy entire wavelengths, we aim to design an optical architecture suitable for more general network models.

Wavelength utilization is increased by aggregating multiple flows into the same lightpath. Although typical flow-routing techniques identify individual TCP flows based on the 5-tuple (transport protocol, source port, destination port, source IP address and destination IP address), such high granularity would require that routers keep track of several millions of flows at the same time.

In OIS we exploit the fact that heavy-tail traffic distribution (characteristic of the elephants/mice phenomenon) still applies when IP flows are aggregated into prefixes [1]. We have implemented an “aggregation matrix”, illustrated in Fig. 3, with a number of columns and rows equal, respectively, to the number of upstream and downstream neighbors (we assume for simplicity that each neighbor is connected through a single default link). The generic matrix cell (i,j) stores information on traffic incoming from interface “i”, relayed through interface “j”. Each cell of the matrix collects information about all the flows that can be potentially aggregated into a single cut-through path, while, within each cell, sampled packets are sub-categorized depending on the destination prefix they are directed to. For each sampled packet the router checks its destination address and determines the output

interface “j”, using the longest prefix matching algorithm. The size of the payload contributes to the total amount of data carried by the matching prefix within the cell (i,j) (“i” being the interface from which the packet arrived). More advanced algorithms might be developed here that also use packet timing information to predict the traffic behavior, improving the traffic characterization. The design of higher performance algorithms will be addressed in future work.

Fig. 3. Prefix-based aggregation mechanism: traffic statistics are categorized, first depending on source and destination interfaces, and within each cell, on the destination prefix

The advantages introduced by aggregating flows through destination prefixes are many. The most important is that the size of the routing table at the upstream node is not unduly increased, since each prefix summarizes a large amount of IP flows. Moreover, since most of the prefix entries in the routing tables of peering nodes are similar (what changes is instead the next hop value), the upstream node rarely needs to add new entries, and most of the time it will only modify the next-hop value. Prefix summarization also diminishes the signaling overhead, as only IP prefixes are signaled to the upstream neighbor (each prefix accounting for thousands of IP flows). The traffic analysis phase is also simplified, as traffic information is processed at the granularity of the routing prefixes. In particular, packet sampling can be operated at low rates, as characterization of individual IP flows is not required.

The main disadvantage of prefix summarization is that information about each flow is disregarded, therefore it is not possible to guarantee quality of service to individual flows. This means that the OIS architecture we propose in this paper only delivers best effort service. If higher QoS is required, then additional mechanisms, like for example DiffServ or the flow-routing [17] developed by Roberts at Anagran Inc., should be implemented electronically. In this paper we do not investigate the implications of using such mechanisms in the OIS architecture, and a more thorough investigation of this topic is left for future work.

C. Optical path creation After collecting traffic information in the "Observation"

state, the node passes to the "Provisioning" state. At decision time the router analyzes the statistics collected, summing up the amount of data carried by the different prefixes within each cell and evaluates the feasibility of switching such flows into a dedicated optical cut-through path. This path is then established between the selected upstream and downstream neighbors. A path creation algorithm is used for deciding whether the traffic flow associated with a number of network prefixes, should be re-directed with a new cut-through path to locally bypass the IP layer. The algorithm we have used is based on a static threshold. Its value depends on the traffic

> Manuscript ID: 117376<

5

characteristics and should be fine-tuned to optimize cost-effectiveness of the IP bypass operations. In our work we have not applied such optimizations. In the simulations presented in section IV, we vary the amount of traffic, keeping thresholds at the same value. It is possible to extend the algorithms to allow for dynamic thresholds or to use more cooperative decision making strategies which will support clustering of nodes that are affected by the selected flows. This will feature in future work.

The node that proposes the cut-through path passes to its upstream neighbor the list of IP addresses that should be redirected over the new path. Based on this information the upstream node updates its IP routing table and starts injecting the appropriate packets into the cut-through path. Fig. 4 shows the path creation process initiated by node “A” to set up a transparent path from “D” to “C”. Clearly, the advantage of the cut-through path is that node “A” can switch the flows at the optical layer, without consuming expensive router resources. At the bottom of Fig. 4 we see how the routing table of node “D” is modified by the OIS protocol. In order to have the legacy IP and the OIS protocols coexist without interfering with each other, we have added a field beside the “Next hop”, named “Dynamic link”, which indicates the outgoing interface used to inject packets into the cut-through path. If the “Dynamic link” is not null, the packet will be forwarded through the interface indicated by this field, otherwise the “Next hop” value will be used. When a cut-through path is canceled, the “Dynamic Link” entry of the prefixes previously switched through the path is deleted, and the node can forward the incoming packets following the original IP routing table.

Fig. 4. Path creation process: the routing table of the path source is updated to route selected prefixes into dynamically allocated optical paths

A relevant issue in dynamically reconfigurable optical networks is that frequent link reconfiguration can cause stability problems for the routing protocols, as every new cut-through path generates a new link adjacency at the IP layer, causing the generation of routing discovery messages (for example, the link state advertisements generated by the OSPF protocol). In our implementation we have tackled this issue by hiding the dynamic cut-through paths to the link discovery protocol of the IP layer. No link state information is forwarded over the dynamic links and the routing over optical cut-through paths is handled completely by the OIS protocol.

D. Extension of an existing dynamic path Since Optical IP Switching bases the optical provisioning

on local decisions, all the newly generated paths involve only three nodes: a source, a transparent switching node and a destination (where source and destination nodes need to process the packets electronically). The optical path extension process allows the paths to be extended to more nodes, increasing the number of transparent hops on each path.

Path extension, similar to the path creation, is based on local analysis, hence each node can only extend the path by one hop, either upstream or downstream. At decision time the node recognizes that the interface from which the flow aggregate enters the router is already part of an existing cut-through path, and triggers the path extension mechanism. The extension procedure, shown in Fig. 5, consists of node “C” reconfiguring the optical switch to transparently connect the incoming cut-through path to an outgoing port directed towards node G. When node C decides to extend the existing path, it sends a message to the downstream node requesting its availability to extend the path. Just before physically triggering the extension, node C sends the new list of switched prefixes to source node D, which will update its routing table accordingly.

Fig. 5. Path extension process, which allows adding progressively more nodes to the transparent cut-through paths

The main issue with path-extension is that statistically only a subset of the original traffic flow aggregate can be extended towards an adjacent node (additional details can be found in [18]). Consider a flow aggregate being routed through the optical cut-through path D-A-C, in Fig. 5. Only a subset of the initial aggregate is routed to G (the remaining are forwarded to H). As the path is extended towards G (into the D-A-C-G path), the subset of flows directed to H is removed from the cut-through path and routed through the default IP links. The extension algorithm tackles this issue by sending a “path-restoration” message upstream, if the traffic re-directed to the default IP link is above the path creation threshold. Such mechanism triggers the upstream node to promptly reconsider the creation of a new optical path for D-A-C, without waiting

> Manuscript ID: 117376<

6

for the next decision cycle. The extension mechanism introduced in this section, allows

the OIS architecture to build optical paths spanning several hops. Hence, rather than using centralized decision mechanisms, based on end-to-end path allocation, OIS achieves the same goal through distributed decision-making.

E. Cancellation of an existing dynamic path Existing cut-through paths are deleted in order to free

resources like interfaces, optical ports and wavelength channels that have become under-exploited due to changes in traffic. The path cancellation process identifies cut-through paths carrying data rates below a pre-established “path cancellation threshold”. This threshold needs to be lower than the path creation and extension thresholds. We have used a fixed threshold. Since in our architecture all provisioning operations are triggered by traffic observation periods in the order of several seconds to minutes, brief traffic spikes do not compromise the effectiveness of the decision mechanism. OIS cut-through paths can only be canceled by the path source or destination, all the other nodes being transparently bypassed. Every time a dynamic path is deleted, the traffic that was being switched returns to being routed hop-by-hop through the default links. Although such traffic increases the occupancy of routing resources, the freed optical resources can be reused to allocate new cut-through paths, bypassing additional IP traffic.

In the implementation we introduce in this paper, the

threshold values for path creation, extension and cancellation are statically assigned (Table I reports the values we have used for the simulations). In future implementations however the thresholds could be dynamically adapted to the optical and IP resources available at the nodes. For example lower values could be used when a higher number of optical resources is available at the node (with respect to the IP resources), to increase the optical bypass of the IP layer.

IV. ARCHITECTURE SIMULATIONS This section describes the simulations we have performed to

analyze the performance of the Optical IP Switching architecture. The reference topology we have used in our simulations is that of GÉANT, the pan-European data communication network, interconnecting different National Research and Education Networks in Europe, serving over 3500 research and education institutions. The network logical topology (which we have reconstructed in Fig. 6) includes 23 nodes, connected by links with average distance of 797 km. The transit traffic on the network nodes is 36% of the total.

We have taken the GÉANT network as a model for our simulation study because it provides real traffic traces together with BGP routing tables, which allowed us to reconstruct (using the C-BGP simulator [19]) the exact paths of the traffic flows and analyze their aggregation into high bandwidth optical paths.

The simulator we have built is written in PERL. In the simulations, each node operates independently and can only process local information (such as data rate, previous and next

hop) for each routed traffic flow. In contrast to real networks however, all nodes in the simulation start from an initial blank state, with no optical cut-through path setup, and receive full traffic from the traces (which were recorded from GÉANT at peak time). This implies that initially the default IP links need to be provisioned to handle the entire traffic. As the simulation advances and OIS nodes start operating cut-through paths, traffic is redirected from default to dedicated optical paths. At this stage the network enters in a “steady state”, where paths are created and cancelled following the traffic encountered. As the network has reached an average balance between traffic on default IP links and traffic on dynamic cut-through paths, a large amount of the default IP capacity that was used during the initial “transient” at the beginning of the simulation remains unused. Since this transient is only due to the fact that the simulation starts with no optical cut-through path and with full traffic speed, which does not realistically reflect the behavior of an operational network, we have disregarded the over-provisioning due to this transient effect.

Fig. 6. Logical layout topology of the pan-European GÉANT network

Table I reports the default setup parameters we have selected to run the simulations. The observation time determines the time interval during which each node examines the traces, and is set equal to the duration of the traffic traces. The channel rate indicates the maximum rate that each wavelength channel can support. The path creation threshold is the minimum aggregate rate needed to trigger the creation of a new cut-through path. The path extension threshold indicates the minimum amount of data needed to trigger the extension mechanism.

TABLE I DAFAULT SIMULATION PARAMETERS

OIS parameter Value Observation time Channel rate Path creation threshold Path extension Threshold

15 min 1 Gbps 100 Mbps 100 Mbps

A. Comparison between OIS and a centralized architecture In this section we compare our distributed OIS model to a

> Manuscript ID: 117376<

7

reference architecture that creates end-to-end dynamic optical paths with a centralized approach. A practical example of such a model could be a transparent overlay network model, where a network administrator centrally provisions optical paths, following the analysis of traffic demand matrices. For the end-to-end architecture we have assumed that the traffic demand generated by all nodes is available in a central database. A dedicated optical path is then provisioned for each end-to-end demand with average data rate above the threshold of 100 Mbps (for comparison with the OIS threshold). In the analysis we have carried out we have not operated any optimization either in the distributed or centralized approach, as protocol optimization is outside the scope of this work.

The comparisons between the two architectures is illustrated in Fig. 7, plotting both the ratio between optically switched and overall routed traffic (on solid lines) and the average optical channel occupancy (on the dashed lines). The higher bound, equal to the percentage of transit traffic in the network, is also shown on a dashed line. The x-axis reports different levels of traffic, which were obtained by multiplying the original traces (labeled "x1") by progressively increasing factors, reflecting the situation where traffic varies in volume and the distribution of the demand is unchanged. Although this simple method might not correctly represent the future evolution of traffic in the network, it allows us to analyze the impact of higher traffic volumes on the network performance.

By analyzing the percentage of switched data, we see that, as traffic increases, both architectures reach the value of 36%, which represents the transit traffic in the network. This is the maximum amount of traffic that can avail of optical bypass of the IP layer. The OIS approach however presents higher switching capability for lower levels of traffic. This occurs because OIS, by building the optical paths in a distributed fashion, on a hop-by-hop basis, and analyzing the traffic locally, allows better traffic aggregation compared to an architecture that only considers end-to-end demands. As traffic increases however, more and more end-to-end demands grow above the provisioning threshold and the amount of switched traffic becomes similar for the two models. The channel occupancy is instead in favor of the end-to-end provisioning architecture because of the lower number of paths created.

Fig. 7. Comparison between the OIS versus an overlay architecture, showing the switching ability and channel occupancy when transparent switching is operated within a single domain

If we modify the GÉANT topology to simulate networks

with a lower degree of connectivity, the transit traffic increases substantially, up to 64% for a degree-2 network. In our tests [18] the OIS architecture was always able to reach the transit traffic threshold with a saturation curve similar to that shown in Fig. 7.

Fig. 8 reports an interesting case, where we have considered the possibility of extending transparent cut-through paths across the boundaries of the GÉANT domain. The advantage of this approach is that the GÉANT access nodes acting as source or destinations of the optical paths, can extend the paths transparently to the external networks, gaining the benefit of optical bypass of the routing layer. In this case, GÉANT becomes a highly transparent core network, where traffic is mostly switched at the optical level while routing is mainly operated by the external domains. We have calculated that the overall transit traffic, computed among the core GÉANT nodes, is over 98%. Fig. 8 shows that under such circumstances both protocols can reach switching ratios of over 93%.

Although transparency over multiple domains could increase the network performance, the feasibility of this approach in a real implementation might be compromised by the fact that transparent switching across domain boundaries does not allow network operators to control and filter the traffic entering and leaving the domain. The possibility of inter-domain switching will thus depend on the development of new business models and updated protocols for next generation networks.

Fig. 8. Comparison between the OIS versus an overlay architecture when transparent switching is operated across multiple domains

The simulations we have carried out show that the distributed approach performs better than the centralized approach in terms of bypass of IP traffic, while the situation is reversed for the channel occupancy. Additional results can be found in [18]. In the next section we will see that from a cost perspective these two performance metrics tend to balance each other.

The real advantage of OIS is that it uses distributed analysis and decision-making, which brings extensive scalability advantages both in terms of network size and provisioning time. Such characteristics are fundamental for an IP-optical network capable of following traffic variation over short time-scales (i.e., tens of seconds to minutes).

> Manuscript ID: 117376<

8

B. Cost evaluation of the presented architectures In this section we analyze the performance of the OIS

architecture from an economic point of view, focusing on the possible savings in capital expenditures allowed by the transparent bypass of the IP layer. The cost saving introduced by the optical bypass of the IP layer is the lower cost associated with optical switching compared to electronic IP routing, with a cost difference per port that can reach two orders of magnitude (e.g., considering 10 Gbps core router ports). However optical circuit switching operates on a much coarser granularity, which makes the economical advantage in a real environment unclear. An analysis is therefore necessary to evaluate the cost-effectiveness of the OIS architecture.

The cost model we have developed considers the reduction in routing equipment allowed by optical bypass, together with the increase in optical devices needed to implement the OIS concept. Besides the cost of optical switches, the model considers the increase of transport costs in terms of additional optical regenerators, longer reach transmitters and links, and higher capacity WDM systems. Longer reach transmitters are needed because transparent switching increases the distance covered by the optical signals and must be combined with the use of more expensive optical amplifiers offering better signal-to-noise ratio. When the distance covered by the optical path is above the optical reach of the transmitter (which we consider to be 2,500 km), the signal needs to be regenerated, adding a cost that increases proportionally with the number of channels. The increase in the number of WDM channels reflects the increase in the number of lightpaths used for optical bypass of the IP layer.

Fig. 9 shows the basic equipment that we have considered in our cost model. Although not explicitly illustrated, all devices are bidirectional.

Fig. 9. Abstract node models of the opaque and transparent architectures used for the cost analysis

TABLE I I COSTS CONSIDERED FOR THE NETWORK DEVICES

Device Cost (units) IP card (10G, core router) IP router (per 160G traffic) 800Km-reach transceiver 2,500Km-reach transceiver Signal regenerator WDM multiplexer, 40\lambda WDM multiplexer, 80\lambda Link (short reach) Link (long reach) Transparent OXC port WSS 9-ports 80\lambda WSS 20-ports 80 \lambda WSS 40-ports 80\lambda

120 50 1.3 2.2 2.7 4.5 6.7 50 70 2 10 12* 15*

*future estimation - effects of higher costs are evaluated through the sensitivity analysis.

The costs we have considered for the network equipment, illustrated in table II, were taken from [9] and [20]. When we considered high levels of traffic, for which systems with more than 80 wavelengths were needed, we extrapolated the cost, keeping the same relative cost as with the existing systems (for example, assuming that the cost ratio between a 160 and 80 wavelength system is equal to that between an 80 and 40 wavelength system). These calculations are based on the assumption that the cost of WDM and WSS technologies will increase linearly in the foreseeable future. Such assumption is relaxed later in this section, where we report the results of the sensitivity analysis of our model to cost changes, considering the effects of an additional increase in the cost of WSS devices. The link costs were calculated considering an average length of 800 Km, and that in the OIS model every link should be engineered to support an optical reach of 2,500 Km (in order to support transparent switching). The link costs are calculated in terms of signal amplification and dispersion compensation, while fiber placement costs are not included as they are common to the models considered. The cost of the Wavelength Selective Switches (WSS) reported in the table was directly obtained by equipment vendors. Systems with higher degrees and number of ports are obtained by interconnecting more WSS, while a MEMs-based transparent OXC is introduced to ensure full steering ability of the optical transceivers. In our analysis, the cost of IP routing is proportional to amount of data routed at each node, rather than to the number of ports. For example, in our calculations a 10G IP port can be split among a number of links. We differentiate between the cost incurred with the routing IP packets and the cost associated to the number of ports needed. Practically this could be implemented by connecting an electronic switch with multiple 1G ports to each of the 10G IP ports. VLAN tagging can be used to forward packets from the IP port to the correct port on the switch. In a prototype implementation this mechanism could be embedded in the OIS router ports. Although OIS cards would have a different layout from legacy IP ports, they would use similar technology and therefore incur similar costs.

Fig. 10. Cost comparison between the opaque IP-over-WDM and the transparent OIS and overlay architectures

Fig. 10 reports the results of the cost analysis we have performed on the following three models: the opaque point-to-

> Manuscript ID: 117376<

9

point IP, which follows the original GÉANT architecture where packets are routed electronically at each hop; the OIS model; and the centralized overlay end-to-end provisioning architecture. The models use the same parameters described in table II, apart from the channel rate, which is equal to 10 Gbps (creation and extension thresholds are both set to 100Mbps).

The first observation we make is that the difference between the overlay and OIS models is negligible, indicating that both are capable of exploiting the advantages offered by transparent switching in a similar way.

The second observation is the difference between the point-to-point and the transparent switched models. For low levels of traffic, the point-to-point IP model shows an advantage over the transparent models. If in fact traffic is too low, the dynamic optical paths cannot exploit the optical bandwidth enough to make the system cost-effective. As traffic increases however, the optical paths are capable of switching a higher amount of traffic, reducing the expenses for costly IP router cards.

It is interesting to give a time dimension to the traffic increase we consider in our simulations. The x1 value of traffic in the plots refers to network traces collected in 2005 from the pan-European GÉANT network. Although it gave us the opportunity of working on a realistic scenario (in terms of traffic traces and IP routes), GÉANT carries relatively low volume of traffic. Our 2005 traces showed an average overall load at peak times of about 10Gbps. If we consider a core UK commercial network, overall traffic (in 2009) reaches about 1 Tbps (i.e., 100 times higher). From the plots in Fig. 10, a 100 times traffic increases show already a cost advantage of about 20%. Following current UK traffic estimates, the 1000-fold volume (i.e., 10 Tbps) will be reached within the next 5 years, while larger international networks are already close to such values.

We have repeated the cost simulation for the inter-domain case, with transparent paths crossing the domain boundaries. In this model we have added the cost for additional WDM links and transceivers in the external domains. The results are illustrated in Fig. 11. The difference between the overlay and OIS models, similarly to the intra-domain case, is negligible.

Fig. 11. Cost comparison considering a multi-domain environment

The cost difference between the opaque and transparent architectures however is more remarkable, as the savings introduced by transparency are well above 80%. The reason

for such a dramatic cost reduction is the reduction of costly IP resources at the interface between the GÉANT core and the edge networks. As already stated in section IV.A however, implementation of transparent inter-domain networking will require new business models and protocol upgrades.

We have carried out a sensitivity analysis, studying the effects of different equipment costs on the overall network cost. Decreasing the cost of IP cards and increasing that of optical switches diminishes the cost savings of the transparent architectures. Considering the inter-domain model, for example, we have varied the costs by a factor of 4, halving the cost of the IP cards and doubling that of the optical switches. This led to a decrease of the maximum cost saving between transparent and opaque architectures from 83% to 67%. This modest decrease, compared to the high variation in cost considered, indicates that the results we have obtained would be moderately affected by possible cost changes in network equipment.

In our cost study we did not take into consideration how OIS would coexist with other architectures in the access and metro areas. Whether the OIS approach could be cost-effective in the metro is unclear. On the one hand OIS is cost-effective where the traffic matrices change dynamically over time [21], which is typical of metro areas. On the other hand however metro networks carry lower amount of traffic (compared to the core), which reduces the efficiency of OIS.

Nonetheless we believe that optical transparency will play an important role in the reduction of capital expenditures as well as power consumption for the metro and access networks in the future. As network technology progresses towards higher transparency [13], spreading deeper into metro and access areas [22], all transparent network architectures will benefit from higher production volume of optical components (e.g., ROADMs and transceivers), which will progressively reduce their cost. It is also likely that a number of different technologies will coexist. As a translucent architecture (i.e., capable of both transparent and opaque operation), OIS can coexist with both opaque and transparent technologies, which puts it in favorable position in the evolution from opaque to transparent networking. In [23], for example, we have successfully carried out coexistence tests of OIS with the User Controlled Light Path project (UCLP).

In this work we have not carried out quantitative calculations of resiliency-associated costs. Qualitatively, it seems reasonable to suggest that the higher flexibility brought about by dynamic optical path provisioning can reduce over-provisioning costs (as the results reported in [21] suggest). Higher flexibility at the optical layer in fact allows more cost-effective redistribution of the traffic load after a failure occurs. A quantitative analysis of resilient OIS operations is left for future work.

V. TESTBED EVALUATION In order to demonstrate the feasibility of the Optical IP

Switching concept and carry out tests on the impact of dynamic wavelength re-routing on the transport layer protocols, we have assembled a testbed using off-the-shelf

> Manuscript ID: 117376<

10

hardware. The testbed setup is shown in Fig. 12. It is made up of two core nodes of an OIS network, each connected to an access node. One of the channels is used as a default connection, carrying signaling and routed traffic, while the others dynamically accommodate the optical cut-through paths.

For core and edge nodes we have used INTEL machines with 3-GHz Pentium 4 processors running Linux Ubuntu. The OIS software for the routing and signaling was implemented in click language [24]. Each machine was provided with multiple Gigabit Ethernet ports, each connected to an optical CWDM transceiver. The optical switch is a Glimmerglass 16x16 port MEMs-based device with nominal switching time of 25 ms. Each server controls an optical switch through a dedicated TCP connection, using commands expressed in the Transaction Language 1 (TL1).

Fig. 12. Network architecture of the OIS testbed

We have used the testbed primarily to analyze the effects of path creation and extension on the TCP transport protocol. Dynamic creation, extension and cancellation of optical paths in fact might cause packet loss, jitter and out of order arrival, which can degrade the quality of applications transported over UDP and TCP protocols, because it can increase jitter and cause loss and out of order arrival of packets at their destination. For live stream applications transported through UDP, packet loss and jitter deteriorate the perceived quality of the signal in a similar way. In dynamic reconfigurable networks however, the jitter or packet loss is not constant over time but only concentrated around the switching time, so that the effect on a streaming application might be noticeable but very limited in time.

The issue is different for the TCP protocol. Although the effect of packet jitter is usually negligible, as long as it is not above the TCP timeout for the acknowledgment (usually of the order of a few seconds), out of order arrival and packet loss can negatively affect the protocol performance. Packet loss causes both the retransmission of packets and the shrinking of the acknowledgment window, significantly slowing down the data transfer process. In OIS, where packet losses are caused by the switching time of the optical devices, the TCP protocol erroneously attributes such anomalies to network congestion, reducing the TCP window. The out of order arrival can instead produce double acknowledgments. Although such problems have been solved with the wide deployment of improved congestion algorithms in network

stacks (e.g., TCP Reno), our experiments shows that there are some cases where switching optical paths can still cause disruption to the TCP protocol.

A. Path creation: effects of switching from default to dedicated link Issues on packet delivery arise in OCS networks when a

packet flow is switched from a default into a transparent optical path that bypasses some of the IP routing nodes. As optical switches have negligible transit time, packets travelling on the new optical path can overtake those on the IP-routed path, causing out of order arrival at the destination node. This situation occurs when the sum of the transit times of the routers bypassed by the optical path is higher than the time gap between two consecutive packets belonging to the same flow. The packet inter-arrival time [s] of an application

sending data at rate R [Bytes/s], using packets of size B [Bytes] (with the simplification that packets are sent at uniform interval time) can be expressed as:

[s] (1)

If indicates the transit time of the i-th router that will be bypassed by the optical path, we could state that out of order arrival occurs when the sum of the transit times is larger than the packet inter-arrival time:

(2)

considering (1), (2) can be expressed as:

(3)

The higher the rate above threshold, the higher the number of packets that will arrive out of order. Packet transit time in IP routers depends on many parameters (hardware/software implementation, type of packet processing, queuing at the output link, etc.). With no queuing delay at the output interface, we can assume a mean transit time of about 100 µs. Considering that the maximum size for a TCP packet allowed by common Ethernet MTU is 1460 Bytes, we find that if the optical path bypasses one router at a time (like in the OIS architecture), the out of order arrival only occurs for rates higher than about 115 Mbps. This simple calculation however does not consider the effect that router queues and links with different bandwidth have. As packets traverse a network the interval time between packets ceases to be uniform and starts assuming stochastic connotations. Therefore as the term at the denominator (3) increases (due either to an increase on the transit time, or to the fact that more hops are bypassed at once by the optical path) the threshold that determines the rate of possibly disrupted flows decreases. The case becomes even more unpredictable with TCP as data is usually sent as a burst, whose length depends on parameters like maximum window size, round trip time (RTT) and type of congestion avoidance algorithm used. Due to the effect of the TCP window size and RTT time, out of order arrival can occur for TCP for much

> Manuscript ID: 117376<

11

lower rates. If, in addition, we consider that congestion at the router output can increase the router transit time by orders of magnitude, the threshold that determines the rate of possibly disrupted flows decreases even further. If we apply a 5ms queuing delay to (3) for example, the threshold for out-of-order arrival is about 2Mbps.

Fig. 13. Packet arrival versus time of the TCP (Reno) packets received at the destination node

The tests on the TCP protocol were carried out using Linux

sockets (Kernel version 2.6.27-11), which implements the Reno algorithm, which adds fast recovery and avoids invoking congestion avoidance when duplicate acknowledgments are received. We have produced a stream of 160 Mbps, which was divided by the TCP protocol in segments each 1460 Bytes long. The results of our tests show that out-of-order arrival and duplicate acknowledgments occur on a very limited scale. Therefore their effect is barely noticeable as the fast retransmit function of the TCP Reno can easily cope with such small disturbances, keeping the average performance constant. Fig. 13.a shows how TCP packets arrive at the receiver, plotting the sequence number of the received TCP packet versus time. We have then added congestion in the router that is bypassed by the optical path. By increasing the transit time at the router (the value in formula 3), we can simulate congestion, which can provoke significant out-of-order arrival during switching time for a given flow rate. Since TCP however adapts the transmission rate to the congestion, before the optical switching is operated the transmission rate of the TCP flow is throttled down, so that the out of order effect is not severe. Fig. 13.b shows the case where congestion only occurs on the path on which data is being transferred. At the beginning we see that the transfer rate is relatively high, but due to congestion it soon throttles down. At switch-over time, the packets still in the router queue arrive after those in the new path, causing out of order arrival. Since the congestion window is already quite small however, this out-of-order event does not slow down data transfer, because the window is not reduced any further. After the switching, the congestion at the IP router is bypassed and the transfer rate quickly rises again to non congested levels.

The case where congestion occurs both on transmission and acknowledgment paths, shown in Figure 13.c, is more problematic. Since congestion also occurs in the reverse path, slowing down TCP acknowledgments, after a transitory effect during switching time, the transfer rate of the application remains similar to the initial one (although operating over a dedicated optical path). Therefore, in this case the creation of a unidirectional cut-through path does not bring any substantial benefit to the application. In addition, since the TCP acknowledgment packets are transmitted over a congested path, the Reno algorithm does not manage to handle efficiently the out of order arrival issue. This causes a transitory effect that slows down the packet transmission by a few seconds. Such results have important consequences on the decision-making mechanisms of dynamic path creation. If congestion affects both direction of a TCP flow, the optical bypass in the direction of the data transmission does not improve the flow rate and can cause additional transitory impairments. Creating an optical bypass also on the opposite direction would solve the problem. However, since in a real network the upstream congestion could occur at a different router along the path, the coordination of multiple bypasses at different locations could constitute quite a complex network management problem.

We have also carried out tests on UDP. Since the UDP

> Manuscript ID: 117376<

12

protocol does not implement any feedback mechanism, the severity of the out-of-order impairment is directly proportional to the flow rate and router latency. Transitory effects can (for both TCP and UDP) be eliminated by introducing a guard time at the upstream router, which delays the sending of packets on the new optical path. Such delay makes sure that packets on the new path reach the destination after those on the default path have arrived. This guard time is variable and can be provided by the bypassed router, through an up-to-date estimation of its packets queuing time.

B. Path extension: modification of an active path The extension process modifies a cut-through path that is

already transporting data traffic; all the packets traversing the switch during the switching time are lost. This causes the TCP protocol to reduce the size of the contention window and retransmit the packets lost, sensibly reducing the average transfer rate of the process.

We have implemented a simple mechanism to avoid packet loss during path extension. The idea is to avoid packets crossing the optical switch while switching is in operation. Following this mechanism, the node proposing the extension, after completing the default extension signaling operations, sends a "STOP_flow" message to the path source before triggering the optical switch (illustrated in Fig. 14). The path source, on receipt, sends back an "Acknowledgment" message and stops sending traffic towards the selected path, buffering these packets. When receiving the "Acknowledgment" the node extending the path actuates the switch to physically extend the path. After the operation is complete, the node sends back to the path source a "RESTART_flow" message, which triggers the path source to start using again the cut-through path.

Fig. 14. STOP_flow signaling initiating after the extension signaling operations

This mechanism eliminates packet loss during path

extension, introducing instead a temporary increase in jitter in the packet flow. We have measured extension times in the

order of 50 ms. Using a network architecture better optimized for performance however, such time could be reduced to values closer to the 25 ms switching time of MEMs-based devices. The drawback of this method is the buffer required at the source node to store all the packets during the "STOP_flow" time interval. If, for example, we consider an aggregate flow rate of 10Gbps and an extension time of 100 ms, this would require a buffer size of 125 Mbytes. As the number of different path extensions that might be operated at the same time is statistically low (considering that provisioning operations are accomplished on a time-scale ranging from several seconds to minutes), the estimated cost for the additional router memory is only a few hundred dollars, which is negligible compared to the overall router cost.

Although this technique eliminates impairments for the TCP protocol, the jitter introduced might create disruption to UDP-based streaming applications. However, only application that cannot use pre-buffering might suffer from this issue (for example video conferencing), and in most cases this impairment would be perceived as a minor and transitory degradation of the service.

VI. FUTURE IMPLICATIONS The OIS distributed path creation method we have

described in this paper introduces a novel mechanism of traffic analysis and optical path provisioning. Rather than sending traffic information to a centralized agent for analysis and topology optimization, the nodes operate distributed decisions.

The most relevant aspect of the architecture we have presented is that it enables a dynamic topology re-configuration that is both fast and scalable. This is in line with the future trend for the Internet traffic model, which we have discussed in section II. The substantial growth of consumer over business traffic implies that typical consumer applications will dominate the market. Therefore, rather than handling bulky, business-related wavelength services (where bandwidth required, source and destination nodes are specifically requested by a customer, and therefore known in advance), we would expect a network dealing with several flows ranging from a few Mbps to tens of Mbps, whose duration and arrival time are unknown. We believe that the distributed provisioning model will provide the ability to adapt the logical topology to such a highly variable traffic demand.

By introducing the OIS mechanism we raise the question whether it makes sense to extend the distributed IP approach also to the optical layer. Such a model could in principle revolutionize the current centralized management approach that was inherited from the telephone monopolistic system (in the US for example the network was only deregulated in 1996, with the Telecommunication Act). The reasons in favor of the distributed approach are: increased scalability, quicker reaction times, and higher support for inter-domain networking, as historically the distributed IP model has proved very suitable to the autonomous organization of independent domains. The drawback is that distributed decision-making mechanisms might fail to convergence to optimal solutions, if

> Manuscript ID: 117376<

13

the nodes operate with local information. Centralized approaches on the other hand can avail of

global information on link and traffic status to reach optimality. However, for network applications, centralized optimization methods do not scale very well as network dimension and number of constraints increase (for example when physical impairments are taken into account). The issue becomes particularly relevant when considering frequent topology reconfigurations that follow short time-scale traffic variations (considering that optical paths can be created within seconds). In that case centralized optimization algorithms are too slow to deliver timely topology updates, and their optimality becomes worthless as the network state can change significantly in the sort time.

The OIS approach we have introduced describes our initial work on network architectures capable of handling sort time scale (orders of tens of seconds to minutes) traffic variation. This is combined with a cross-layer approach that aims at making the best out of both the routing and wavelength switching domains. Ideally this will lead in the future towards cognitive IP-optical networks, where nodes are intelligent agents capable of taking automated and coordinated decisions to deliver a fully adaptive self-configuring network.

VII. CONCLUSIONS In the previous sections we have introduced the concept of

distributed dynamic provisioning, implemented through the Optical IP Switching architecture. We have then reported the results of simulation and testbed characterization of the architecture, which was compared to a transparent overlay centralized provisioning model. The comparison was then extended through a qualitative discussion, which is summarized in Table III.

Our initial study shows that a distributed decision-making approach to dynamic optical networking can be feasible. The advantages are many, including high scalability and fast topology reconfiguration, which enable dynamic response to highly variable traffic conditions. In our future work we will evaluate more in depth the pros and cons of carrying out fast network adaptation through intelligent network agents.

TABLE I I I

SUMMARY OF COMPARISON BETWEEN OVERLAY AND OIS NETWORK MODELS

Issue Centralized overlay OIS Traffic analysis Path provisioning

- Traffic collection requires central coordination - Traffic information is continuously forwarded to a central agent - The central agent processes very large amount of data - Centralized mechanisms can use global knowledge to provide optimal solutions.

- Traffic information is collected independently at each node - Traffic information is processed locally - Each node only processes an amount of data equal (or less if traffic is sampled) to the traffic it routes - Distributed agents use local information that can lead to sub-optimal decisions.

- Centralized network engineering, with global monitoring, and large-scale optimization is not suitable to address short time-scale traffic variation.

- Distributed agents can react very quickly as network adaptations are dealt with locally.

REFERENCES [1] M. Ruffini, D. O’Mahony and L. Doyle, “Optical IP Switching for

dynamic traffic engineering in next-generation optical networks”, in Proceedings of Optical Network Design and Modeling conference, 2007, pp. 309–318.

[2] M. Ruffini, D. O’Mahony and L. Doyle, “Feasibility of flow-based optical provisioning in GEANT”, in Proceedings of OFC, 2007.

[3] G. L. Jones, W. Forysiak and J. H. B. Nijhof, “Economic benefits of all-optical cross connects and multi-haul DWDM systems for European national networks”, in Proceedings of Optical Fiber Communication Conference, 2004.

[4] P. Ferreira, W. Lehr and L. McKnight, “Optical networks and the future of broadband services”, Elsevier, Technological Forecasting and Social Change Journal, vol. 69, no. 7, pp. 741–758, 2002.

[5] D. Saha, B. Rajagopalan and G. Bernstein, “The optical network control plane: State of the standards and deployment”, IEEE Communications Magazine, vol. 41, no. 8, 2003.

[6] J. D. Jones, L. Ong and M. A. Lazer, “Creating an intelligent optical network worldwide interoperability demonstration”, IEEE Communications Magazine, vol. 42, no. 11, pp. S4–S11, 2004.

[7] E. Oki, K. Shiomoto, D. Shimazaki, N. Yamanaka, W. Imajuku and Y. Takigawa, “Dynamic multilayer routing schemes in GMPLS-based IP+Optical networks”, IEEE Communications Magazine, vol. 43, no. 1, pp. 108–114, 2005.

[8] K. Shiomoto, E. Oki, W. Imajuku, S. Okamoto and N. Yamanaka, “Distributed virtual network topology control mechanism in GMPLS-based multiregion networks”, IEEE Journal on Selected Areas in Communications, vol. 21, no. 8, pp. 1254–1262, 2003.

[9] G. Weichenberg, V. Chan, and M. Medard, “On the throughput-cost tradeoff of multi-tiered optical network architectures”, in Proceedings of IEEE Globecomm conference, 2006, pp. 1–6.

[10] T. Lehman, J. Sobieski and B. Jabbari, “DRAGON: A framework for service provisioning in heterogeneous grid networks”, IEEE Communications Magazine, vol. 44, no. 3, pp. 84–90, 2006.

[11] P. Szegedi, Z. Lakatos, and J. Spath, “Signaling architectures and recovery time scaling for grid applications in IST project MUPBED”, IEEE Communications Magazine, vol. 44, no. 3, pp. 74–82, 2006.

[12] N. Taesombut, F. Uyeda, A. A. Chien, L. Smarr, T. A. DeFanti,P. Padopoulos, J. Leigh, M. Ellisman, and J. Orcutt, “The OptIPuter: High-performance, QoS-guaranteed network service for emerging e-science applications,” IEEE Communications Magazine, vol. 44, no. 5, pp. 38–45, 2006.

[13] O. Gerstel, “IP and optical integration in dynamic networks”, in Proceedings of OFC, 2009.

[14] P. Newman, G. Minshall and T. L. Lyon, “IP switching-ATM under IP”, IEEE/ACM Transactions on Networking, vol. 6, no. 2, pp. 117-129, 1998.

[15] T. Dietz, F. Dressler, G. Carle, B. Claise and P. Aitken, “Information model for packet sampling exports”, IETF RFC, March 2009.

[16] K. Papagiannaki, N. Taft, S. Bhattacharya, P. Thiran, K. Salamatian and C. Diot, “On the feasibility of identifying elephants in Internet backbone traffic”, Sprint ATL, Sprint Labs, 2001.

[17] L. G. Roberts, “The next generation of IP - flow routing”, in Proceedings of SSGRR International Conference, 2003.

[18] M. Ruffini, “Optical IP Switching”. PhD thesis, University of Dublin, Trinity College, 2008.

[19] B. Quoitin and S. Uhlig, “Modeling the routing of an autonomous system with C-BGP”, IEEE Networks, vol. 19, no. 6, pp. 12–19, 2005.

[20] M. Gunkel, R. Leppla, M. Wade, A. Lord, D. Schupke, G. Lehmann, C. Furst, S. Bodamer, B. Bollenz, H. Haustein, H. Nakajima and J. Martensson, “A cost model for the WDM layer”, in Proceedings of Photonics in Switching conference, 2006.

[21] M. Ruffini, D. Kilpler, D. O'Mahony, L. Doyle, “ Cost Study of Dynamically Transparent Networks” in Proceedings of OFC, 2008.

> Manuscript ID: 117376<

14

[22] D. B. Payne, R. P. Davey, “The future of fibre access systems”. BT Technology Journal, Vol. 20 , No. 4, pp.104-114, October 2002.

[23] A. Sánchez, S. Figuerola, G. Junyent, E. Kenny, V. Reijs, M. Ruffini, “A user provisioning tool for EoMPLS services based on UCLPv1.5”. In proceedings of the TERENA Networking Conference, 2007.

[24] E. Kohler, “The Click modular router”. PhD thesis, MIT, 2000.