1. Introduction

Recent advances in core transport technologies are aiming to maximize the “bandwidth-times-length product” since the steadily increasing traffic load will soon require core network links with tens of Tb/s capacity per fibre. At the same time new switching technologies are emerging for building the next-generation switching nodes able to support the emerging requirements for aggregate capacity approaching Pb/s and dynamic traffic patterns exhibiting high spatial and temporal asymmetry. Thus, the objective is to provide the necessary bandwidth under stringent QoS requirements while allowing for a significant curb in the overall cost and power consumption.

Under the above framework, the advances of optical transmission technologies are profound and evident [1,2]. In striking contrast, the identification of an efficient switching mode is still an open issue raising considerable controversy. This controversy is part of a much broader skepticism: is it reasonable to believe that existing networking paradigms could guarantee the end-to-end QoS performance across a core network under acceptable CAPEX (capital expenditure) figures? If not, we need to have a major rethinking with respect to the existing premises.

Today, all developments are aiming at the proliferation of the Internet. A necessary step to safeguard the sustainability of Future Internet is the identification of networking modes performing equally well across an extended set of parameters that include traditional QoS like (i) capacity and throughput, (ii) end-to-end packet loss, (iii) end-to-end packet delay and jitter, plus a set of additional parameters like (iv) degree of resource utilization, (v) availability, (vi) physical layer performance (Q-factor, BER), (vii) power consumption. Obviously, interest is on solutions offering a balanced performance across all indexes. Additionally, each solution should be investigated in conjunction with the associated switching mode for it is exactly this interplay that makes feasible to attain the requested performance.

Apparently, a stumbling block in achieving these objectives is that the most established optical networking modes/paradigms fail to provide either mechanisms for efficient and coherent traffic grooming facilitating the transportation and switching of different granularities ranging from an IP packet to a WDM waveband, or mechanisms achieving statistical multiplexing gains (the main feature of packet switching) equivalent to the robust two-way resource reservation (the main feature of circuit switching). Historically telecommunication networks have evolved following two distinct approaches, namely the connection-oriented and connectionless communication modes. Today’s existing offspring solutions (pure IP, MPLS-TP, PBT-TE, etc.), have the characteristics of one of these modes i.e. they are in essence either circuits or packets in their origin and they are trying to embrace the features of the complementary mode (i.e. statistical multiplexing and guaranteed performance, respectively). Unfortunately, they fail to do so because of the inherent granularity mismatch between the transmission granularity of the electronic layer and that of the WDM channel: the maximum identifiable bandwidth granularity stemming from the “electronic layer” is too small (compared to the round trip time in a Metro/Core network) to allow for an efficient two-way reservation. Moreover, this granularity mismatch has a secondary effect: IP-over-WDM networks are bound to frequent transitions between the optical and electronic layer (o/e-e/o conversions) followed by OTN aggregation in order to minimize the total amount of resources used i.e. the number of WDM channels and the number of optical switching ports. This incurs the penalty of excessive information processing leading to complex, high-CAPEX, solutions requiring considerable power consumption.

In this work, we benchmark the performance of the main networking modes in core networks against the extended set of QoS parameters. The networking modes under consideration include Optical Circuit Switching (OCS), Optical Burst Switching (OBS), Optical Fast Circuit Switching (OFCS) which can also be understood as dynamic OCS, and a new core optical networking architecture we have proposed named Clustered Architecture of Nodes in Optical Networks (CANON). The mode of operation of the aforementioned networking paradigms are described and analyzed in the following section (Section 2). The rest of the paper is organized as follows: in Section 3 we discuss the network dimensioning and operational parameters that affect performance in each case, define the performance metrics of interest and elaborate our benchmarking methodology aiming to have a consistent, coherent and fair way to simulate the essential aspects of these mechanisms. In Section 4 the results of the performance benchmarking over several figures of merit are presented. Through extensive simulation results we quantify performance gains and we show that CANON demonstrates the highest efficiency achieving both targets for statistical multiplexing gains and QoS guarantees. Finally Section 5 concludes our work.

2. Overview of dynamic and static resource reservation schemes and related work

OCS provides the equivalent of circuits in legacy telecommunication networks. “Optical circuits” are established through lightpaths [3] and are classified to transparent, translucent, opaque or hybrid. For future reference in this work, transparent is the configuration where a lightpath is not wavelength converted when cross-connected in intermediate nodes, translucent is the configuration with intermediate wavelength conversion [4] but no traffic grooming, and opaque is as before including OTN traffic grooming. Finally, a hybrid mode is the case where there could be traffic grooming in a selected number of nodes (to be decided during the planning phase) whilst in other network paths optical bypassing is deployed (off-loading). Optical bypassing [5] is proposed as a means to provide WDM channel routing while obviating expensive electronic switching. In this case, lightpaths are pre-provisioned resulting in semi-static network connections making use of, mature, slowly reconfigurable WDM technology i.e. Reconfigurable Add-Drop Multiplexers (ROADMs) and/or Optical Cross-Connects (OXCs). Their set-up/release procedure can be either centralized or distributed since the interconnection patterns do change but on a very large time scale.

Regarding OCS, several studies have shown that over-provisioning of lightpaths is essential to guarantee QoS mandating a considerable number of transceivers and large port-count switches [6–10]. Thus, the requested performance is attained at the expense of a high overall cost and poor resource utilization.

In order to overcome the scalability limitations of OCS, schemes allowing for a) statistical multiplexing gains enabling sharing of resources and b) traffic grooming at sub-wavelength granularity, are sought. So far, various dynamic resource allocation schemes like Optical Burst/Packet Switching (OBS/OPS) were studied as an alternative to OCS. In OBS presented in [11] and [12], the incoming packets are queued at edge nodes. Each node aggregates traffic towards a particular destination node and casts it into a burst after transmitting a reservation message informing the intermediate nodes for the upcoming burst transmission. Burst generation (also called “burstification process”) is performed based on local queue status information. Burst switching nodes rely on a more mature technology than OPS [13] and for this reason, in this work we focus on OBS rather than OPS. However, there are serious concerns regarding OBS and specifically that the one-way reservation scheme may lead to high burst loss probabilities at high loads [14–20].

The comparison of OCS and OBS networks has been conducted under several conditions [7–19]. The work in [7–10] and [17,18] present quantitative results regarding the comparison of OBS against OCS under specific conditions based on simulation. In [14,15] the authors approximate boundary conditions analytically under the necessary simplifying conditions in order to obtain a tractable problem formulation. Through this analysis they quantify potential efficiency improvements due to statistical multiplexing, using an OBS version employing centralized scheduling called Wavelength Routed-OBS (WR-OBS) originally proposed in [16], instead of OCS. In [7] the first results demonstrating performance gains of OBS vs. OCS under highly dynamic traffic patterns where demonstrated for a given network topology with a static dimensioning and some limiting specific implementation details including alternate routing in OCS and deflection routing in OBS. In [8–10], the authors provided an integrated model for packet and flow based traffic modeling for a more fair comparison. In [17] modeling a large core network capable of grooming millions of ingress packet flows and optimally assigning wavelengths as well as wavebands in the case of OCS the authors claim that in order for OBS to achieve an acceptably low blocking probability, a large increase in the network capacity is required. Only recently the conditions under which the above conclusions may hold were related to the traffic grooming capabilities of the network in [10], [18] and [19]. Due to the limited scalability of solutions based on centralized scheduling like WR-OBS or on real-time characterization of flows of optical bursts in order to identify in-profile and out-of profile bursts as proposed in [10] in large size core nodes we focus in our study only on typical implementations of OBS employing one-way reservations. In [18] the authors correctly identify the impact of the network topology (size and average nodal degree) and available capacity on the performance in each case. To incorporate this in their study they employ a planning methodology in order to optimally assign wavelengths according to traffic demands (i.e. assuming that traffic can be terminated in intermediate nodes in order to achieve traffic grooming through electronic processing) also modeling the relative cost in each case. Finally, in [19] the comparison also includes aspects of the overall network cost (CAPEX) following specific cost estimation models comparing OBS to an alternative circuit-switching oriented model called Optical Flow Switching (OFS). OFS employs a hierarchical network organization similar to CANON, where distributed MANs are interconnected through OXCs via established light-paths to carry inter-MAN traffic, employing different scheduling algorithms for inter- and intra-MAN communication. Since [18,19] focus more on the approximation of CAPEX vs. traffic level performance (level of burst/packet losses) under the simplified assumption of a same static average traffic demand they fail to provide valuable insight related to application level performance at the time scales of a packet switching vs. a circuit-based, pre-provisioned per-flow, network. This has been better addressed in terms of fair comparison in [8–10]. Evidently, direct quantitative comparison of data losses and overall performance in an OBS vs. an OCS network is quite challenging mainly due to lack of realistic traffic models, a large number of variables and parameters to consider, and a long simulation time involved in simulating at a granularity as fine as packets. In [19] finally, as the authors observe the underlying assumptions for OFS to be efficient is first, that large transactions are needed in order to be able to overcome the lightpath setup and teardown times, and second, a suitable (centralized) control mechanism is needed for OFS to be able to fully use that lightpath. In order to overcome these limitations in this work we evaluate OFCS as a flow-based switching paradigm under the same simulation framework with the other paradigms in order to quantify the trade-offs between flow set-up delay vs. packet losses vs. buffer and link dimensioning and resulting utilization.

In [21] a mechanism termed Optical Fast Circuit Switching (OFCS) employs a time-limited wavelength reservation, a process that is disassociated from the time-frame needed for forming-up the transportation unit (packet/burst). Thus, OFCS is implementing a two-way reservation protocol (e.g. RSVP) where each reservation lasts for an unknown period and, hence, it may serve a number of consecutive number of packets/bursts (Fig. 1 ).

 

Fig. 1 Wavelength reservation and data transmission (a) OFCS (b) OBS.

Download Full Size | PDF

In OFCS, similarly to OBS, when packets arrive at an ingress node, they are buffered (if no established channel is found available at the time of arrival) until their transmission over the optical link is triggered by the node controller. The system estimates periodically the required capacity, i.e. number of virtual (the term is used to include the broader case of translucent core network architectures) wavepaths (VWPs), with period T_S and then sends the required reservation messages. If successful, each reservation remains active until the system decides to release the resources. The estimation is based on real-time measurement of the data queue occupancy status and the ratio R_S(t) of the average number of packets (which are considered of fixed-size L_F in bytes) arriving during the time interval T_S at the node, over a number of intervals N. The reservation process of OFCS is schematically illustrated in Fig. 1(a), while for comparison purposes the scheme for OBS is shown in Fig. 1(b).

In [22–24], we have proposed CANON as a solution for networks exhibiting considerable spatial and temporal traffic asymmetry. As it is shown in this work, CANON demonstrates (a) efficient grooming at the optical layer via a distributed multiplexing based on ring topology networks (obviating any electronic grooming), (b) a balanced deployment of dynamic resource allocation with two-way reservation mechanisms and (c) statistical multiplexing gains per network segment (as opposed to the achievable gains per single node).

With reference to Fig. 2 , we consider a partially-mesh connectivity core network (upper right). Following CANON, a Core/Metro network is decomposed into a number of geographically limited areas (clusters). The partitioning is based on various criteria like administrative domains, topological characteristics, traffic patterns, legacy infrastructure, and so on. An important consideration is that each of these clusters comprises a group of nodes in geographical proximity. We have developed a formal methodology for clustering original mesh-topology networks but the details of this algorithm are outside the scope of this paper.

 

Fig. 2 Clustered Architecture for Nodes in an Optical Network (CANON).

Download Full Size | PDF

A cluster consists of a number of nodes and the resultant structure is further decomposed to a number of ring topology networks; the ring topology becomes a building block for network engineering purposes. Thus, a single –hierarchically flat- core network is now seen as an assembly of clusters generating a single inter-cluster and a number of intra-cluster sub-networks. Since the objective is the ring nodes to share the same wavelength(s) in TDMA fashion, there is no shortest path routing in the intra-cluster segment. In each cluster there is a single node with a dual purpose: it acts as a gateway between the particular cluster and other clusters and it coordinates the transmission of all other nodes of its own cluster. Consequently, two distinctive node classes are identified: a number of Metro-Core Edge Nodes (MENs; called Regular Nodes in [23]) with a ROADM in the optical section and a single Core Transit Node (CTN; called Master Node in [23]) that is an OXC and is the gateway node. Moreover, CANON postulates two new sub-wavelength granularities in the optical layer, namely, the optical slot and the optical frame (Fig. 3 ). Having defined these, the end-to-end operation is as follows: in a given MEN, the incoming traffic is classified per class-of-service (CoS) and it is groomed (under the CoS priorities) into fixed size containers called Optical Slots (slots, hereafter) based on MEN destination criteria; contiguous slots are also envisaged. The details of this process are discussed in length in [23]. Under the supervision of a Medium Access Control (MAC) protocol executed at the CTN, a given MEN would launch slots in the ring. The slots from different MENs in a ring destined towards nodes residing in a specific remote cluster are transported towards the CTN over the same wavelength channel(s). Hence, the role of the CTN is twofold: (i) to arbitrate, by means of a central scheduler, the collision-free introduction of the slots from all MENs in the ring like in [25] acting as a distributed multiplexer, and (ii) to handle the transportation in the inter-cluster network segment of the generated frames towards their destination cluster. A long cascade of slots effectively forms a larger container called Optical Frame (frame, hereafter). To accommodate for traffic variations the process is repeated in a time-frame equal to the duration of the frame. Based on fairness SLA constraints etc. one MEN may add more slots than another leading to statistical multiplexing gains over the ring under spatial and temporal traffic asymmetries.

 

Fig. 3 CANON layering and transmission granularities.

Download Full Size | PDF

To schematically illustrate the operation, in Fig. 3 the size L of the slots is in the order of several microseconds and could be standardized e.g. to accommodate one or more OTN G.709 [26] frames. A frame (of size F in the order of several milliseconds) is a container that is formed up when slots are casted together towards specific distant clusters. For example, in Fig. 3, two wavelength channels are supposed to transport traffic from Cluster 1 to Cluster 2 and one channel from Cluster 1 to Cluster 3. The upper part of Fig. 3 is showing the grooming hierarchy in CANON in both electronic and optical domains. Efficient grooming is achieved, per node, by means of standard electronic L2/L3 systems whilst the CANON intra-cluster operation, similar to “distributed multiplexer”, as said, is leading to statistical multiplexing gains across the ring. It is pointed out that this is a collective effect, having its origin in the coordinated operation of many network nodes, and not the outcome of an individual node; thus, allowing for gains that are not feasible using schemes like OBS/OPS etc.

In the inter-cluster segment (between CTNs), a wavelength channel may transport either a single frame or a cascade of frames to a distant cluster. This wavelength channel may be statically [23] or dynamically reserved as a response to traffic fluctuations [24] so the communication mode between CTNs may exploit OCS, OFCS or OBS. To avoid confusion, for the rest of this work when the communication mode is OCS, OFCS or OBS between edge nodes this will be designated as “mesh” hereafter, and when it refers to the inter-cluster segment transportation it will be designated as “CANON”.

Overall, in the intra-cluster segment where the optical slots are short in duration compared to the round trip time the lossless transportation of slots is based on a MAC protocol which has proved to be an efficient arbitration mechanism in access networks. Due to the distributed slot multiplexing and the statistical multiplexing gains achieved we will show that the traffic profile is smoothed-out which means that optical slots are well-groomed in optical frames.

From the above discussion, we can conclude that CANON offers a complete switching framework addressing all aspects and providing solutions based on widely explored technologies to achieve both geographical network partitioning as well as traffic grooming through an efficient real-time control plane handling resource allocation of appropriate transport layer hierarchies and efficient concentration/switching node architectures. In this work it is the first time that a thorough comparison of CANON versus static and dynamic wavelength reservation and switching architectures under a comprehensive set of conditions and performance metrics is presented. Since, it is also interesting to asses cost-related aspects, apart from the estimation of the required resources in order to achieve specific QoS levels, a cost comparison is attempted following the approach of [18]. More details related to the impact of CANON on node and network dimensioning resulting in CAPEX and OPEX savings can be found in [27,28]. To allow for a direct comparison we model the same switch architecture [27] capable of fast switching (msec frames/bursts scale) employing o-e-o conversion and single frame buffering in core nodes/CTNs only for synchronization at the slot/frame level (i.e. without any queuing at intermediate nodes for contention resolution) and wavelength conversion. It is worth noting that slow switching (OCS/OFCS), without wavelength conversion capabilities would allow for simplified i.e. less costly operation by means of all optical switching architectures like MEMS avoiding the o-e-o completely. However, since a detailed cost comparison is outside the scope of this work we only focus on the functional characteristics of the switching nodes, which are described in more detail in the following section. Additionally, hybrid architectures based on electronic traffic grooming as studied in [17–19] (allowing for what is also called as optical bypassing) are not included in this study, since a fair comparison in such cases would only be meaningful under an accurate techno-economic study. In contrast we only focus on the fair and direct comparison of the different networking and switching paradigms under similar conditions in order to expose the merits and drawbacks of each solution. Additionally, we evaluate the performance of sub-wavelength switching modes (OBS, OFCS, CANON) under different operational parameters regarding burst/frame sizes and time thresholds and show the impact on control plane load, which should not be ignored. The network dimensioning for the expected traffic matrix in each case is based on the theoretical lower capacity bound as proposed in [15], while new utilization metrics are introduced in order to provide better insight into the operation of each networking mode and demonstrate the reasons of higher resource utilization. Below we discuss the details of our modeling and then present the simulation results for a number of scenarios detailing our main observations and findings.

3. Benchmarking methodology

The main objective of the following sections is to present the methodology for and the results of the benchmarking of CANON against the other networking paradigms; namely: mesh OCS, OBS and OFCS. The profound difficulty in this process is to grasp and simulate the essential aspects of these mechanisms in a coherent and fair way. This is not a straight-forward task given the -inherently orthogonal in some cases- operation modes on subjects that are ranging from the resource reservation process to the subsequent switching granularities, to node functionality (e.g. buffering, synchronization) etc. A further concern is to select the traffic/network level parameters that are meaningful to carry out the modeling work as well as the selection of appropriate network topologies, routing schemes and the available network resources (i.e. link capacities) that will facilitate comparison on the same grounds.

Regarding the switching granularity, it is worth pointing out that OCS and OFCS employ static and dynamic, respectively, two-way channel reservations before data transmission between any two nodes. OBS and CANON on the other hand employ burst data transmission on a burst or frame level, respectively. While OBS has been initially conceived to operate based on variable length bursts [6], it has also been studied in [12,20] in its slotted variant. Slotted OBS (S-OBS) is an OBS solution where all bursts are constrained to be of a fixed size and all control plane messages are transmitted under a specific time relation with respect to data units creating a slotted system. Apparently, S-OBS offers many advantages not only due to a smaller burst collision probability [20] but also due to the reduced switching node control and scheduling complexity, when variable slot allocations need to be scheduled in real-time (only at the cost of synchronization at each port, which is not considered to have a critical impact). Following the above remarks and in order to allow for direct comparison with CANON, our benchmarking uses this OBS variant. Hence, the S-OBS burstification and the CANON frame generation processes share some common features described below.

For each destination node/cluster (OBS/CANON respectively) upon receipt of the first packet (at the local queue of an OBS edge node) or request (at the CTN in CANON) to reach this destination, a timer T_k is set. Whenever the timer reaches a maximum acceptable delay value T_MAX (variable queues and thresholds could be implemented depending on the required QoS) burst/frame generation is triggered irrespective of the total amount of accumulated packets/requests. In case, during the time interval between setting T_k and its expiration after T_MAX, the sum of the packets/requests expressed by R_tot exceeds the burst/frame capacity F i.e. R_tot > F, a burst/frame (of fixed size F) generation procedure is initiated. Finally in case R_tot exceeds a minimum acceptable burst/frame fill level U_MIN, after a reasonable waiting time T_MIN again a burst/frame generation is triggered, in order to expedite data forwarding at the cost of an acceptable level of burst/frame underutilization (in which case U_loss = F- U_MIN).

The next step in our analysis is to select reference network topologies and a traffic load profile per core network node. In this work, three topologies are considered: the Pan-European core network of [29] that consists of 16 nodes interconnected in a partially mesh topology b) an ideal network topology of 16 equidistant nodes as an abstract case and a network with a larger number of nodes using the national backbone network of Telecom Italia [30] (called hereafter TI network) consisting of 40 nodes with an average nodal degree 3.15 facilitating the extraction of global conclusions.

Figure 4(a) depicts the Pan-European network topology for both partially-meshed node connectivity and CANON. As discussed in [31], the latter introduces a node connectivity modification shown in the inset of Fig. 4(a). The (ideal) equidistant node topology is illustrated in Fig. 4(b). In this case, to simulate OCS, OFCS and S-OBS a Torus network connectivity is assumed, while to simulate CANON the selected node connectivity follows the argumentation of [31]. Finally the original mesh topology TI network is shown in Fig. 4(c) (lower part). In order to modify the node connectivity following CANON, we performed node clustering resulting in the (asymmetric) clustered network shown in Fig. 4(c) (upper part). For the inter-cluster segment, the partial-mesh connectivity of CTNs exclusively uses existing links. The implicit assumption has been made that each CTN is connected with at least 2 CTNs while the inter-cluster network retains the original average nodal degree.

 

Fig. 4 Pan-European network in a mesh and clustered (CANON) configuration (a), Ideal equidistant 16 node network following a 4x4 Torus (mesh) and clustered (CANON) topology (b), Telecom Italia mesh and clustered network (c).

Download Full Size | PDF

Regarding the traffic load profile, the following statistics for the incoming traffic were assumed: fixed-sized data units (packets) arrive at edge nodes in the case of S-OBS and aggregated data slots formed at CANON MENs, with size L equal to 0.125 msec. Under the same assumptions of [13,20] inter-arrival times are exponentially distributed and destinations are stochastically varying following a uniform probability distribution. For each node, the inbound traffic is transported over four/ten incoming wavelength channels (M = 4/10) for the 16-node/40-node topologies respectively. The burst/frame size F is considered equal to 5 msec and the minimum frame fill level for early frame generation U_MIN is selected equal to 0.5⋅F. Since in CANON frame generation depends on the queue status of all the MENs in a given cluster (4 up to 6 in the topologies under study), in the case of S-OBS U_MIN is set to be equal to ¼ of the value used in CANON in order to make the comparison meaningful. However, in order to investigate the impact of U_MIN as well as to evaluate S-OBS even under more favorable conditions (reducing loss at the cost of higher delay) in addition to U_MIN = 0.125⋅F, we also considered scenarios selecting U_MIN = 0.5⋅F. It is pointed out that this threshold might not be the optimum value neither for S-OBS nor for CANON. Following the same rationale, queues (per destination cluster) of 400 slots (50 msec) in CANON MEN nodes are considered while in mesh S-OBS, queues (per destination node) of 100 slots (12.5msec) are considered per edge node. Also, it has been assumed that core nodes (CTNs only in CANON) implement the λ-S-λ architecture of [27] allowing for full wavelength conversion. For synchronization purposes in slotted modes, the actual propagation delay of each link is rounded to an integer multiple of the frame/burst size F. In reality, the synchronization is achieved in the λ-module of the λ-S-λ node architecture by placing the necessary slot/frame buffer between the fixed-receiver, tunable transmitter pair connected back-to-back. It is worth noting here that the o-e-o conversion per input port is employed only to provide signal 3R regeneration and frame synchronization and not buffering and queuing to assist contention resolution, which would drive the complexity, cost and power consumption to much higher levels. Finally the intra-cluster round trip delay in CANON is set equal to F i.e. 5 msec.

Performance evaluation was conducted by means of an event-driven network simulator. All simulations were run for sufficient time to obtain steady-state results. In general, thirty million time units were simulated per point in each curve. Model validation followed the variance reduction technique with 40 replicated simulations using different seeds obtaining results with 95% confidence.

4. Performance metrics and simulation results

In this work, the aforementioned networking modes are benchmarked against the extended set of QoS parameters which are used as “figures of merit”. Although the definition of parameters like loss and delay is straightforward, parameters like network utilization and efficiency, that correlate traffic performance with CAPEX, overall network resources and their utilization level, need further elaboration. The merit functions we used are the following:

- Packet or Slot Loss Probability (SLP) measures the percentage of the lost L3/L2 packets that arrive at ingress nodes which are then transmitted either as they are (i.e. with no further L2 processing like in OCS, OFCS) or they are aggregated into burst/frames (OBS, CANON). It is pointed out that in our work SLP reflects the fraction of the incoming packets to the edge node that are lost and not the percentage of burst (OBS case) or optical frame (CANON) losses. The origin of these losses is either due to collisions (output port blocking in an intermediate network node) or buffer overflows at edge nodes.

- Delay measures the average end-to-end time needed for the packets (at the input of an ingress node) to arrive to the final recipient node. The delay is a collective event including the queuing delay at the ingress node and the propagation delay. The retransmission delay, which is important in OBS due to collisions, is not taken into account.

At this point we should stretch our attention to the impact of the overall network dimensioning (i.e. the capacity assigned to each link) on network performance. The link capacity can drastically alter the performance of certain figures or merit e.g. it would lead to lower SLP figures (referring to output port contention in OBS or blocking in OFCS; assuming nodes providing full wavelength conversion) while it has a dominant share in the overall system cost both in terms of hardware interfaces (CAPEX) as well as in terms of power consumption (OPEX). Therefore, SLP and delay – albeit their key role in assessing network performance – cannot be exclusively used to determine the overall effectiveness of a given solution. These figures should be assessed in comparison to the capacity assigned to different links. The policy adopted by most of the network and service providers is to attain certain SLP and delay figures while minimizing cost (CAPEX, OPEX) and maximizing the utilization of network resources. For given networking conditions i.e. technology, traffic characteristics and performance targets, one needs to establish the grounds over which network efficiency is measured.

To do so, we define the Total Network Capacity (C) as C = Σc_i, with c_i being the capacity i.e. the number of WDM channels, of link i. For mesh connectivity networks this is equal to the number of transceivers (Tx/Rx pairs) making the implicit assumption of one Tx/Rx pair per channel, per node port. Obviously, the value of C is dictated by: a) the network topology (number of physical unidirectional links – where a “physical link” refers to the transmission medium that can operate independently of and in parallel with other physical links to transport traffic) and b) the input load (I) and the target throughput (T). Apparently, T ≤ C. So, at this point we introduce the following network utilization factors:

- Maximum Capacity Utilization (U) = Throughput/ Total Network Capacity (T/C); relates the attainable throughput T to the corresponding network CAPEX (in turn related to the installed capacity C); in other words, it shows the trade-off between cost and SLP. T refers to the entire network i.e. it expresses the sum of the successfully served traffic from all nodes.

- Reserved capacity/Total Network Capacity (R/C) provides a useful insight into the way the resource reservation mechanisms operate and demonstrates the long term average capacity they require as a ratio of the total installed capacity C. Obviously static solutions like OCS reserve available resources at the highest level to service the peak rate demand.

- Used capacity/Reserved capacity (S/R) provides a useful insight into the effectiveness of the corresponding networking architecture since it shows the long-term average of the transmitted data units for the reserved resources. Obviously framing mechanisms that allow transmission of partially filled frames/bursts or predictive reservation schemes (like OCS and OFCS) allowing transmission of empty/idle frames in case of absence of data (assuming synchronous operation) result in lower values of this factor.

As mentioned earlier, a critical issue that strongly affects network performance is the link dimensioning. The point of departure of our benchmarking is the number of WDM channels required to serve the input traffic load in static OCS designated, hereafter, as C_OCS. Given this figure we demonstrate the performance trade-offs that sub-wavelength resource allocation can achieve and benchmark CANON against S-OBS, OFCS and static OCS. In OCS, each node is interconnected with every other node by means of at least one VWP. In CANON, the inter-cluster network consists of fully interconnected CTNs by means of a mesh logical topology.

For OCS each source-destination pair is served by means of one VWP (i.e. one wavelength) and wavelength assignment is based on shortest path routing. For fully dynamic networking solutions i.e. S-OBS, OFCS, CANON, there is no static VWP reservation between any pair of nodes. In these cases, the links are dimensioned to what we call “minimum required capacity” (C_MIN) which is defined as follows: it is the capacity (more accurately, the number of transceivers) needed to serve the expected long term average load of each link i according to the resulting traffic matrix (always under the shortest path routing assumption) at 100% input load or else the lower bound on the wavelength requirements as defined in [15]. It is worth noting that for the resulting traffic matrix C_MIN turns out approximately 30% of C_OCS in average for the above scenarios, which is also the condition evaluated in [10]. In order to evaluate the performance trade-offs we have examined the performance of the dynamic networking schemes in two scenarios with link capacity more than C_MIN but still lower than that of the C_OCS bound, as is also performed in [10]. So, for S-OBS and OFCS we studied two additional cases for capacities 45% and 75% of C_OCS. To benchmark under equivalent conditions i.e. the same aggregate C (more accurately, the number of transceivers), CANON was scaled to 60% C_OCS which is the value stemming from the CANON clustering and operation in order to result to the same number of transceivers as above. The overall Total Network Capacity C for the 16-node networks (Figs. 4(a), 4(b)) is shown in Fig. 5 .

 

Fig. 5 Total capacity C (number of Tx/Rx) for the 16-node networks.

Download Full Size | PDF

Figure 5 is very indicative with respect to the necessary resources each network requires to achieve the corresponding QoS performance. Evidently, all dynamic solutions based on sub-wavelength reservations allow for significant savings in terms of the number of transceivers and, hence, CAPEX. In contrast, a mesh connectivity network aiming to achieve the requested QoS performance exploiting OCS incurs the highest CAPEX. However, since the exact impact of link dimensioning on cost is determined by the relation of the switching complexity to the total capacity and switching granularity, a detailed analysis, should take into account the specific implementation details of each solution. Such an analysis is out of the scope of this paper and we plan to include it in future extensions of this work. This mode consumes an excessive amount of resources, compared to the actual demand, but it retains a high level of networking performance with no losses and minimal delay – irrespective of traffic load – as shown in the average results summarized in Table 1 .

Table 1. Traffic performance (SLP, Delay) and utilization factors for OCS (independent of traffic load)

View Table | View all tables in this article

Conclusively, OCS outperforms the other networking modes but it does this at an unacceptably high cost that severely limits the scalability of this solution. It should be pointed out here that even though OCS appears to have excellent performance, in the extreme cases where the traffic rate increases close to the service rate, packet losses would inevitably occur. Actually, the system for dedicated capacity under the Poisson arrival pattern can be modeled as an M/D/1 queuing system (if we assume an ideal system with infinite buffering capability). It is well known from queuing theory [32] that M/D/1 performs reversely proportional to (1-ρ) where ρ = λ/μ is the system utilization factor, λ is the average arrival rate of the Poisson process and μ is the service rate. Thus, congestion (queue size and delay) is expected to worsen (actually tending to infinity or – practically – high SLP) when λ→μ. However, in our case a dedicated wavepath served by one WDM channel is more than enough to accommodate the traffic matrix we have assumed leading to guaranteed robust performance at the cost of underutilization: the data of Table 1 indicate U = 0.1/0.15/0.13 for the Pan-European, Torus and TI cases, respectively, whilst R/C = 1 meaning that all available wavelengths are reserved but they are poorly utilized since S/R = 0.27/0.27/0.26. Last but not least, the lower number of transceivers for the ideal Torus network is due to the higher connectivity pattern (larger average nodal degree) of this topology mandating fewer wavepaths for node interconnection.

Given that OCS trades network performance with high CAPEX under time-varying traffic patterns, dynamic reservation schemes like S-OBS and OFCS aim to exploit statistical multiplexing, per single node, to set a lower figure on the required resources as well as to limit resource under-utilization. Certain conditions under which OBS can achieve improved utilization figures are reported in [7–16].

In Fig. 6 , the performance of S-OBS over a number of QoS related parameters for the Pan-European network is shown. In Fig. 6(a) the SLP is shown for various values of the capacity C, burst size F and U_MIN. The figures for C = 0.45∙ C_OCS, were omitted for clarity but the same qualitative conclusions hold. One can observe that when C <C_OCS, (even if C is significantly higher than C_MIN), S-OBS SLP performance deteriorates rapidly. Evidently, providing a higher number of wavelength channels allows reducing the collision probability (in absence of buffers at intermediate nodes, a prospect that would increase complexity and cost). So, in S-OBS a higher C value allows improved SLP figures at the expense of poorer utilization figures as shown in Figs. 6(c)-6(d) and in Table 2 . Since the achievable throughput is directly proportional to the available capacity (up to the point of maximum possible throughput before congestion) we do not plot the entire curve of T/C vs. load but only present in Table 2 the maximum achievable ratio T/C at 100% load. For the figures R/C and S/R though, it is of interest to present the utilization of available resources under all loads in order to draw interesting conclusions on the limitations of the reservation mechanism and the way the available resources are exploited. Moreover, the results of Fig. 6 demonstrate that SLP in S-OBS can improve either by increasing the burst utilization factor (U_MIN) or by reducing the maximum burst size (F). Increasing U_MIN enforces denser packing of data into bursts (Fig. 6(d)), which leads to improved utilization in Table 2 but again at the cost of significantly higher delay (almost doubles the delay penalty at lower traffic loads and sacrifices the advantage of low latency of OBS) as shown in Fig. 6(b). Reducing F has the same effect as above since it also leads to more frequent transmission of bursts, which are better utilized improving all performance figures of S-OBS. However, the burst size F cannot be reduced arbitrarily since this severely affects the scalability of an OBS network, as is shown next.

 

Fig. 6 S-OBS performance in terms of SLP (a), Average end-to-end delay (b) and resource utilization expressed as R/C (c) and S/R (d).

Download Full Size | PDF

Table 2. Throughput/ Total Network Capacity (T/C) for S-OBS

View Table | View all tables in this article

The burst size F directly affects the transmission efficiency (due to physical layer overheads) but more important it affects the control channel load since the control plane protocol overhead in OBS is directly proportional to the number of bursts transported over the network. This overhead is shown for different values of the average message load (across all links) per node in the network (expressed as a percentage of the channel capacity) in Fig. 7 , where one can observe the inversely proportional relation between F and control channel load.

 

Fig. 7 S-OBS and OFCS total average message load per node.

Download Full Size | PDF

This could become critical in networks with high numbers of nodes and many WDM channels per link, leading to severe congestion and, eventually, SLP degradation [33]. Thus, we can conclude that S-OBS in a meshed connectivity network trades one QoS performance parameter for another; overall, it has a very poor performance unless the network capacity is significantly over-provisioned, while the utilization of the available resources is modest at best (Fig. 6 / Table 2).

The QoS performance of the Pan-European mesh connectivity network topology employing the OFCS solution is shown in Fig. 8 . Similarly to S-OBS, the performance of a network adopting the OFCS mode is affected by the sensitivity of the sampling period T_S, during which the decisions for channel reservation or release are taken, and a threshold T_h for rounding to the closest number of VWPs [21]. For a direct comparison with S-OBS, the QoS indexes are parameterized against the same set of values. However, their actual impact on OFCS performance is lower than that of the corresponding parameters in S-OBS. What is important to point out is that OFCS, unlike S-OBS, actually improves the SLP performance (Fig. 8(a)). Specifically, when C = 0.75∙ C_OCS, the Pan-European network under the OFCS mode exhibits lossless performance. However, this performance is attained at the expense of significantly higher end-to-end delay as it is indicated in Fig. 8(b) that may exceed 100 msec (average value) in extreme cases (i.e. at higher loads and limited capacity). This is an unavoidable outcome due to the delay introduced by to the capacity estimation process. OFCS utilizes as many resources as possible to prevent losses due to buffer overflow at edge nodes and, as a result, it demonstrates a much better utilization performance compared to S-OBS as observed in Figs. 8(c)-8(d) and in Table 3 .

 

Fig. 8 OFCS performance in terms of: SLP (a), Average end-to-end delay (b) and resource utilization expressed as R/C (c) and S/R (d).

Download Full Size | PDF

Table 3. Throughput/ Total Network Capacity (T/C) for OFCS

View Table | View all tables in this article

Conclusively, the simulation results for S-OBS and OFCS show that both networking paradigms trade the performance of one QoS parameter for the other. This is inevitable since statistical multiplexing and resource reservation are carried out at a single node level.

CANON on the other hand demonstrates an almost lossless operation in all cases as shown in Fig. 9(a) . Losses are observed only at very high loads and only for a dynamic inter-cluster network (based on OFCS or S-OBS). CANON has a higher end-to-end delay (Fig. 9(b)) compared to OCS and S-OBS but lower than that of OFCS; the delay remains bounded to acceptable limits in all cases. The number of transceivers in CANON is slightly higher than OFCS and S-OBS but still remains much lower than that of OCS. On the other hand, it is very important to point out that although in “mesh” networking solutions the number of transceivers coincides to the number of the number of WDM channels in the network (in reality, the product of number of fibers times the WDM channels per fiber), in CANON the number of WDM channels is significantly reduced due to the statistical multiplexing on network level and not per node, as will be shown below.

 

Fig. 9 CANON performance in terms of SLP (a) Average end-to-end delay (b) and resource utilization expressed as R/C (c) and S/R (d).

Download Full Size | PDF

Last, but not least, the optimal performance is when employing the minimum network resources (i.e. when the capacity of each cluster is selected equal to the aggregate capacity of the core inter-cluster links emanating from the cluster). Actually, for the traffic load distributions examined in this study, CANON achieves optimal performance employing static resource reservation in the inter-cluster part (OCS). It is worth noting that CANON leads to much higher utilization of reserved wavelengths (higher even than the lower bound set by U_MIN) as a result of statistical multiplexing at the optical layer (Fig. 9 and Table 4 ). Dynamic reservations over inter-cluster links would only be preferable in case of traffic profiles that present higher asymmetries. In such extreme cases over provisioning (e.g. 60% of C_OCS as shown in Fig. 9) would diminish SLP at the cost of lower utilization.

Table 4. Throughput/ Total Network Capacity (T/C) for CANON

View Table | View all tables in this article

Having, reached the above conclusions we evaluate for the scenario (set of parameters) giving the best results for S-OBS and OFCS presenting in Figs. 10(a) –10(c) comparatively the benchmarking results in terms of SLP, delay and utilization U = T/C for S-OBS, OFCS and CANON when the ideal network of Fig. 4(b) is used (Torus mesh and clustered respectively).

 

Fig. 10 S-OBS, OFCS and CANON performance for the Torus network and OFCS and CANON performance for the TI network in terms of SLP (a, d) average end-to-end delay (b, e) and resource utilization expressed as U = T/C (c, f).

Download Full Size | PDF

Finally, under the same conditions we evaluate the case of the larger TI network of Fig. 4(c). Since S-OBS performance has been shown in many studies to degrade with the size of the network (e.g. [17,20,34]) we only focus on the performance of OFCS and CANON presenting the results in Figs. 10(d)-10(f). As we observe from Fig. 10(d) the impact of the size of the network on SLP is minimal under the same dimensioning, yielding only a lower T/C ratio due to the large number of links, which in turn result in longer multi-hop VWPs. Thus, we can confirm that our observations above hold for any network size and topology.

The main advantage of CANON exposed above is that it makes feasible two-way reservations within geographically limited domains dynamically adapting resource reservation. However, it is far beyond the geographical partitioning that improves performance, since CANON integrates this feature within a complete framework addressing also traffic grooming through an efficient control plane, appropriate transport layer hierarchies and efficient concentration/switching node architectures. The macroscopic effect of this grooming in the intra-cluster (ring) is a traffic profile shaping in the inter-cluster part. This effect can be observed in Fig. 11 , which shows the distribution of the probability of a WDM channel being occupied by frame transmissions over time serving traffic between clusters. The resulting output traffic profile between three source-destination cluster pairs is shown (clusters C3 to C5, C1 to C3 and C5 to C1 of Fig. 4(c)) at 80% input load. The impact of cluster size asymmetry is also evident. We observe that in all cases there is a very high probability (>90%) of a specific number of WDM channels being constantly occupied while the occupancy probability for a number of channels above this number is low (and the probability for a lower number of channels is nearly zero). The different pairs of clusters yield different distributions since the amount of traffic destined to a cluster depends on the traffic matrix (which -following the assumption for uniform traffic distribution- depends only on the number of nodes of the source and the destination clusters). In all three cases, there is a lower bound C_Si,j (actually equal to C_min discussed above) on the number of channels needed to serve any pair (i, j) of clusters leading to a more or less static traffic profile. In order to further elaborate on the impact of the grooming functionality we summarize in Table 5 the results related to the reduction of node complexity following the approach of [18]. Although a direct comparison would require a thorough technoeconomic evaluation, due to size limitations we only present here in Table 5 the total number of components required in each case assuming that the OXC architecture follows the λ-S-λ (wavelength selective switch) architecture discussed in detail in [27]. Although OCS could be implemented via simpler architectures (e.g. MEMS providing transparency under physical layer performance trade-offs) to allow for direct comparison allowing for wavelength conversion and 3R regeneration at each node (translucent network as discussed above) we base our comparison on this common OXC architecture. Following the methodology of [18] (where a cost comparison of OCS vs. OBS is attempted through the actual dimensioning of the switching nodes) we list in Table 5 the total number of the most critical components [27] i.e. fixed transceivers (Tx/Rx), tunable transmitters (TTx), wavelength routers (AWG), couplers, semiconductor optical amplifiers (SOA) and amplifiers (EDFA). Table 5 demonstrates the cost reduction that traffic grooming at the optical layer can achieve and the optimal balancing of traffic level QoS performance, network resource utilization and cost vs. the other static or dynamic networking paradigms.

 

Fig. 11 CANON inter cluster traffic profiles: 3 source-destination cluster pairs at 80% load.

Download Full Size | PDF

Table 5. Total number of components (CAPEX)

View Table | View all tables in this article

5. Concluding remarks

An extensive benchmarking of several networking modes has been presented for three networking topologies and over a range of operation parameters and conditions. The conclusion of these studies is that OBS and OFCS require a considerable number of WDM channels just to avoid collisions/blocking. On the other hand, CANON not only keeps loss quite low in all cases, even when employing an OBS-like solution over the inter-cluster network, but also it efficiently grooms slots into frames allowing for statistical multiplexing of data over a larger number of contributing nodes, increasing resource utilization as shown in Fig. 9 and Table 4. The same conclusions hold in all topologies and, as also shown in [30] for the clustered topology of Fig. 4(a), the improved performance cannot be attributed to the topology transformation alone. On the contrary, it is the interplay of the hierarchical aggregation, reservation and switching mechanisms that reduce collision domains (a cause of high loss probability) while they are leading to the efficient utilization of available resources and a limited end-to-end delay. CANON outperforms other solutions even when OCS is employed over the inter-cluster network, under the traffic conditions examined in this study, and provides optimal loss and delay performance with a minor increase of the required capacity compared to S-OBS, yielding though high utilization and reduced overall network cost. However, for better scalability both in larger node count network topologies and higher traffic volumes, CANON with OFCS in the inter-cluster network may prove a more cost-effective solution.