1. Introduction

The relentless IP traffic increase and rapid adoption of cloud computing and the SDN (software defined network) environment require expansion of both optical fiber capacity and network functionality. Adoption of flex-grid and a new modulation format can increase fiber capacity to some extent, but increasing fiber degree per route can multiply the ‘network capacity’ [1–3]. The traffic volume may be further enhanced with the introduction of new wavelength services including Ultra-High Definition-TV (4k x 8k pixels, 72-144 Gbps per channel) services [4, 5], and future advanced wavelength service [6–9]. The explosion in traffic will force a rapid increase in the number of wavelength paths and hence fibers between adjacent nodes as indicated in [10, 11], and the importance of developing a large scale ROADM/OXC has been emphasized [1–3, 11, 12]. For example, when the capacity of the current 8 fiber degree of ROADMs/OXCs (in this paper, ROADMs and OXCs are used interchangeably) is completely filled, the traffic increase of 30% a year will require 80 fiber degree ROADMs/OXCs in nine years. Most present OXCs utilize WSSs. The highest port count of current commercially available WSSs is limited to 20 + , and further cost effective expansion will not be easy. If we cascade WSSs, the port count can be expanded, but the cost and loss increase; the two stage architecture of 1x9 WSSs (most commonly utilized size at present) yields 1x81 WSSs, but this requires ten 1x9 WSSs per incoming fiber and the loss is twice that of the single WSS. In addition, the increase in the port count prohibits the use of the present broadcast and select (OC + WSS) configuration since the OC (Optical Coupler) loss becomes excessive (with 81 fibers, the intrinsic OC loss is 19 dB). Hence, the route and select (WSS + WSS) architecture [13] would have to be utilized, which doubles the number of necessary WSSs. If we use 1x9(1x20) WSSs, a total of 1620(486) WSSs would be required to create an 81x81 OXC. The signal quality degradation would be particularly serious in metro networks where large number nodes must be traversed [12, 14, 15]. This clearly indicates that a new approach is necessary to create large port count OXCs cost-effectively.

To resolve this issue, we recently developed a new architecture that utilizes the interconnected multiple OXC-subsystem approach [16, 17]. The OXC-subsystems are created using small cost-effective WSSs; the subsystems are interconnected with a limited number of intra-node fibers. The use of subsystems allows the use of both small port count WSSs and the broadcast and select (OC + WSS) configuration, and as a result, the total number of WSSs needed in the network can be dramatically reduced, by more than 75% [17], and the maximum number of WSSs traversed by a wavelength path to less than 36% [17], compared to the equivalent large scale single OXC. These reductions not only substantially mitigate spectral narrowing and total node loss, but also greatly reduce node cost. All of the benefits are achieved with less than 1% offset in accepted traffic volume [18]. In these studies, the subsystems are interconnected in a ring-like manner to enhance subsystem connectivity, but the architecture does not allow hitless (no service disruption) OXC port count expansion. When expanding the OXC port count, subsystem interconnecting fibers must be cut off and the additional subsystems inserted.

Developing a scalable node architecture that allows hitless modular growth is crucial to cost-effectively introduce OXCs from day one that can be gracefully expanded adapting to the forthcoming traffic growth. In order to resolve the issue, we developed a linear interconnection scheme [19, 20] that allows addition of subsystems without any service disruption. The architecture, however, incurs enhanced intra-node blocking, since the linear interconnection is possible by removing some intra-node fibers from the original ring-type interconnection. In this paper, we investigate the performances of both ring- and linear- type interconnections in detail and clarify the differences and overall effectiveness.

We first explain the newly-developed scheme that efficiently accommodates inter-node (link) fibers with subsystem modular nodes, i.e. the use of intra-node fiber is reduced. We analyze the causes of blocking; blocking can occur at the link level (inter-node connection) and the node level which includes collisions stemming from insufficient capacity at intra-node subsystem connections and the bound for the number of intra-node fibers traversed, a parameter defined in the network control algorithm. An analysis shows that setting appropriate parameter values effectively improves the performances of the linear-type interconnection configuration without losing its merits. We then investigate the adaptability of the proposed OXC to traffic increases or the long-term performance, and the available cost savings. The overall performances of the proposed architecture that include the maximum/average number of traversed WSSs for optical paths and maximum/average total optical loss at nodes are also evaluated. Experiments prove that almost the same blocking performance (about 1% decrease in acceptable traffic volume at the blocking ratio of 0.1%) can be achieved as the conventional single large scale OXC. It is also demonstrated that the reduction in the number of necessary WSSs is enhanced as the traffic increases to exceed 80%. Parts of the preliminary studies on this were presented at international conferences [19, 20].

Another important issue to be resolved before creating a large scale OXC is the development of the add/drop part architecture [21–23]. We have already developed ones that can fully utilize the subsystem modular architecture [24, 25], and so this topic is not addressed here.

2. Subsystem modular node architecture

2.1 Subsystem modular architectures

Figure 1 shows typical OXC architectures that utilize WSSs. Since the present OXC degree is relatively small (< 8), most existing OXCs adopt the broadcast-&-select architecture (Fig. 1(a)), where optical couplers are used at the input side. As the fiber degree increases, the larger optical loss caused by the high port count optical couplers prevents the use of this architecture. Thus the route-&-select architecture shown in Fig. 1(b) has been thought necessary to realize large fiber degree OXCs. As mentioned before, the route-&-select architecture doubles the WSSs needed and worsens the filtering effects due to the increased number of WSSs traversed by an optical path at each node [12, 14, 15]. This can be a critical flaw, in particular when we consider the application of OXCs deep into the metro networks, where number of traversed nodes can be very large [12].

 

Fig. 1 WSS based OXC architecture.

Download Full Size | PDF

Our previously proposed subsystem modular node architecture that adopts ring-type interconnection is shown in Fig. 2 [16, 17]. It consists of small (m x m) OXCs that are bridged by a limited number of fibers. In this paper, we call these small OXCs “OXC-subsystems”. A node can accommodate more fibers by adding OXC-subsystems. When the subsystem scale is small (≦9), it can adopt the broadcast-&-select architecture (Fig. 1(a)). In Fig. 2, two pairs (input and output) of fibers are used for the subsystem interconnection, so seven fibers are for bridging (we assume here 9x9 subsystems). Inter-node fibers are established to connect the same layer subsystems of neighboring nodes (Fig. 3). If adjacent nodes have different numbers of subsystems, in addition to the connection, the top level subsystem of the node with fewer subsystems is connected to higher level subsystems of other adjacent nodes, as shown in Fig. 3. The number in the circle in Fig. 3 denotes the number of subsystems in each node.

 

Fig. 2 Subsystem modular architecture (ring-type interconnection).

Download Full Size | PDF

 

Fig. 3 Inter node fiber connection example between OXC-subsystems.

Download Full Size | PDF

2.2 Newly proposed hitless modular growth capable architecture

Modular growth of the OXC is possible with the architecture shown in Fig. 2, however, interconnecting fibers between a pair of adjacent subsystems must be cut when a new subsystem is to be added. In order to realize graceful expansion without service disruption, we introduce the modified interconnection shown in Fig. 4, where one or two pairs of intra-node fibers are not left unconnected and reserved for future subsystem addition; it forms a linear interconnection with one of two edge subsystems. Intra-node blocking can be increased since the routing ability is degraded compared to the ring-type interconnection. In general, intra-node blocking can be mitigated by developing an efficient RWA (Routing and Wavelength Assignment) algorithm that is aware of the intra-node blocking [26]. The developed algorithm is explained in the next subsection.

 

Fig. 4 Subsystem modular architecture (linear interconnection).

Download Full Size | PDF

2.3 Inter-node subsystem interconnection optimization

Network nodes are unlikely to have uniform numbers of OXC-subsystems; the number of fibers in each link differs. In previous studies [16, 17], inter-node fibers between neighboring node sub-systems are set to bridge them usually at the same levels (Fig. 3). Only if a pair of adjacent nodes has different numbers of subsystems is the top level subsystem of the node with fewer subsystems connected to the equal or higher level subsystems in the adjacent node (see Fig. 3). This intra-node connection scheme is not symmetric and hence if the intra-node routing ability is degraded, as expected with linear-type subsystem interconnection, the routing performance of the node may be deteriorated.

In this paper, we develop a new inter-node subsystem connection scheme where fibers are configured in a round robin manner (Fig. 5). Suppose that a pair of adjacent nodes, $a$ and $b$, have $u$ and $v$ subsystems, respectively, which are respectively labeled as $s_0^a$, $s_1^a$, …, $s_{u-1}^a$ and $s_0^b$, $s_1^b$,…, $s_{v-1}^b$. Suppose that w-1 fibers form the link. To add a new fiber to the link, the fiber should bridge from $s_{\left\{w\mathrm{mod}v\right\}}^a$ to$s_{\left\{w\mathrm{mod}u\right\}}^b$. If there is no vacant port for the subsystem, select another available subsystem randomly. This scheme can attain more symmetric subsystem interconnections between adjacent nodes, which is expected to avoid excessive optical path collisions at certain subsystems.

 

Fig. 5 Connection example between OXC-subsystems (round-robin) in the neighboring nodes.

Download Full Size | PDF

2.4 Routing and wavelength assignment that is aware of the subsystem modular node architecture

Since the intra-node fiber resources connecting OXC-subsystems are limited for subsystem modular nodes, blocking at the intra-node fibers may significantly limit the routing capability. Our solution is an intra-node blocking aware dynamic optical path control algorithm that considers the subsystem modular node architecture with ring-type interconnection [27]. We need to develop an algorithm suitable for the linear subsystem interconnection with subsystem connection between adjacent nodes as explained in 2.3. Another difference of the developed algorithm is that it adopts advanced intra-node subsystem interconnection as described in 2.3 and adaptively changes the F_intra value according to the length (number of hops) of paths as is discussed in 3.5.

For simplicity, we assume that no node offers wavelength conversion and that all inter-node link distances are the same. Each node offer full C/D/C (Colorless/Directionless/Contentionless) add/drop capability. We assume that wavelength paths are established/released dynamically in response to the connection requests. Dynamic path control operation is performed as follows: if a new set up request arrives, a wavelength path that connects the source and destination is searched for, and when found, the wavelength path is set up. Otherwise, the request will be blocked. When a path release request arrives, the path will be immediately torn down and terminated. Uniform link length is assumed throughout this paper to simplify the discussion, however the generalization to a non-uniform link length case or to incorporate a transmission performance limitation is straight-forward.

Here we upper bound the number of intra-node fibers traversed by a wavelength path: the bound is denoted by F_intra. Limiting F_intra can prevent the excessive use of intra-node fibers, which are limited resources, and avoid an excessive increase in the number of subsystems traversed or WSS traversed. For example, when F_intra = 2, each wavelength path can traverse up to two intra-node fibers or three OXC-subsystems in one node, or one intra-node fiber, or two OXC-subsystems, in each of the node pairs along the path as shown in Fig. 6. We also assume a hop slug, the maximum inter-node hop increment from the shortest hop path, to avoid excessive detour lengths.

 

Fig. 6 A path routing example for F_intra = 2.

Download Full Size | PDF

Since it is rather difficult to simultaneously manage the requirements of hop slug and the F_intra bound, we utilize an algorithm that finds suboptimal solutions. The suboptimality and computation viability are verified by experiments. For each setup request between a node pair, we create a list of available pairs of route and wavelength candidates by applying Dijkstra’s algorithm to find a shortest route connecting each pair of subsystems (source and destination nodes). Here the number of subsystem pairs is a multiple of each subsystem number in source or destination nodes. The number of candidates for each subsystem pair can be a parameter, but for most of the evaluations herein it is set to 1 since it has been confirmed to yield satisfactory results. Among pairs of route and wavelength candidates in the list, we find all the candidates that conform to the constraint that the number of traversed intra-node links must be equal to or less than F_intra. Then, among the candidates, we excluded the pairs whose inter-node hop count is greater than the sum of the shortest hop count and the given hop slug. Finally, we choose the minimum inter-node hop count route in priority in order to minimize the number of inter-node fibers traversed. If multiple candidates offer the minimum node hop count, we select the one that traverses the fewest intra-node links. Extensive experiments show that the method provides satisfactory solutions with short computation times.

3. Numerical experiment

3.1 Simulation setup

Physical network topologies used are the 5x5 poly-grid network, the COST266 Pan-European network [28], the US nationwide network [29], the French network. Major parameters of the tested networks are summarized in Fig. 7. This paper assumes that the link length is uniform to clarify the topology dependence of routing performance. Generalization to variable link length is straightforward. Each fiber accommodates up to 80 wavelengths. We assume that the traffic demand is uniform and randomly distributed and is represented as the average number of wavelength path requests between each node pair. The generation of path setup demands follows a Poisson process, the holding time of each connection follows a negative exponential distribution. We set the value of hop slug to 2. For the proposed node architecture, we assume that the OXC-subsystem size is 9x9 (namely, 1x9 WSSs are utilized) and each pair of adjacent OXC-subsystems is bridged by one pair of fibers (one input and one output fiber as seen in Fig. 2).

 

Fig. 7 Network topologies.

Download Full Size | PDF

The number of interconnected OXC-subsystems within each node is determined by the necessary number of input/output fibers to/from the node. Therefore, we first apply a static network design algorithm that assumes conventional single OXC nodes and evaluate the necessary node scale and number of fibers on each link for 20 different traffic distributions (the average number of wavelength path demand between each node pair was set at 7 or 14). Here a path-by-path sequential heuristic is applied; for each path accommodation, try to find a pair of wavelength and shortest route that minimizes the number of newly established fibers. For each link, we obtain the maximum number of fibers in 20 trials and then the number of interconnected OXC-subsystems for each node is that necessary and sufficient to connect the maximum fibers. We prepared the same number of inter-node fibers in both cases to compare the performances of the proposed node and conventional node. The simulation results were averaged over 10 different runs with different traffic distributions.

3.2 Classification of inter/intra-node blocking

F_intra is an important parameter for the proposed architecture. In order to analyze the performance of the proposed architecture and to clarify the effect of F_intra, we classified the causes of blocking using the procedure depicted in Fig. 8; an explanation is given below.

 

Fig. 8 Flowchart of blocking classification.

Download Full Size | PDF

<Type of blocking>

Please note that these three types are interrelated. For example, some types can be resolved by either adding a vacant intra-node fiber or increasing the increment of F_intra or adding inter-node fiber, however, we use the above classification hereafter for simplicity.

Figures 9(a) and (b) show the analyzed blocking types and their frequencies for the 5x5 poly-grid network when the average number of wavelength path demands between each node pair is set at 7 or 14, respectively. The tested subsystem modular OXC architecture is the ring interconnection. The inter-node subsystem connection scheme applied was the scheme originally used in the previous studies. The derived blocking ratios are normalized by these for the conventional node which is free from intra-node blocking, and hence the ratio of 1 is the lower bound. The absolute blocking ratio was around 10⁻³. When the traffic volume is small (i.e. in Fig. 9(a)), intra-node blocking hardly occurs. On the other hand, when traffic volume is doubled (i.e. in Fig. 9(b)), intra-node blocking and F_intra blocking are enhanced and they can be major causes of blocking. This is because when traffic volume increases, the necessary number of subsystems in each node increases, and then the necessary average number of intra-node fibers traversed to move from one subsystem to other subsystem increases.

 

Fig. 9 Blocking classification.

Download Full Size | PDF

Figure 9(c) depicts the result when the new inter-node subsystem connection method (Fig. 5) proposed in Sec. 2.3 is utilized, where the average number of wavelength path demands between each node pair is set at 14. The results show that intra-node blocking and F_intra blocking are substantially reduced and routing performance almost equivalent to that of the conventional OXC nodes with large port count WSSs is attained when F_intra is larger than 4. We also tested the performance of another inter-node subsystem connection method where a pair of subsystems is bridged by random selection. The blocking ratios are almost same as those shown in Fig. 9(c). These results prove that suppressing biased inter-node subsystem interconnections is effective in attaining better routing performance for subsystem modular OXC architectures. We also tested the linear architecture with the same conditions as Fig. 9(c), see Fig. 9(d). Compared to the ring-type subsystem interconnection, intra-node blocking slightly increases with this architecture when traffic volume is large (14), with no increase observed at small traffic volume (7, which is omitted here).

3.3 Performance comparison between ring-type and linear-type subsystem interconnection

Figures 10(a) and (b) indicate the traffic volume that can be accommodated by the ring-type and linear type interconnection when the target blocking ratio is 0.1%. The vertical axis plots the volume of accommodated traffic relative to that of the conventional network. The measured deterioration in routing performance was marginal for the different four network topologies, where the value of F_intra is set at 8. Even with linear interconnection, the fall in accepted traffic was less than 1%.

 

Fig. 10 Accepted traffic demand ratio.

Download Full Size | PDF

3.4 Performance evaluation under traffic increase scenario

The performance of the proposed architecture with subsystem linear interconnection was investigated in relation to the traffic expansion. In order to simulate the traffic increment, the initial demand (average number of wavelength paths per node pair) was set at 4 and periodically increased by 2. As mentioned before, for the proposed node architecture, we assume that the OXC-subsystem size is 9x9 (1x9 WSSs are utilized), and for the conventional node, we assumed that 1x9 or 1x20 WSSs are cascaded to create high port count WSSs.

As the initial setup, networks consisting of conventional single OXC nodes were designed for all traffic volume settings tested (from 4 to 18); necessary fiber numbers between adjacent nodes were determined to accommodate each traffic volume. Exactly the same fiber sets were tested for networks with the proposed nodes. The number of OXC-subsystems at each node for each traffic volume was derived. After fixing these conditions, blocking probabilities were evaluated by slightly changing the base traffic for all traffic volumes for networks with conventional single OXC nodes and those with the proposed node architecture, and the traffic volume that attained the fixed blocking rate was evaluated. The simulation results were averaged over 10 different random traffic patterns.

Performance comparisons between the conventional nodes and proposed nodes for the 5x5 poly-grid network and the COST266 Pan-European network are shown in Figs. 11 and 12, respectively. The traffic volume is periodically increased from the initial value and the number of fibers between nodes and number of subsystems in each node are accordingly incremented. The bar graphs depict the relative traffic volume that can be accommodated by the proposed node when the target blocking ratio is 0.1%, normalized by that for a single large scale OXC. It is shown that networks with the proposed nodes achieve almost the same network capacity as the conventional network for both network topologies. When the traffic demands increase, the normalized accommodated traffic volume slightly decreases, which stems from the increase in intra-node blocking. However, even when the average number of paths reached 18, which corresponds to six years after system introduction (assuming yearly traffic increase of 30%), the network capacity decrease is still only 1%. Please note that 1% of traffic corresponds to a 1/30 year (~2 weeks) time duration when yearly traffic increase of 30% is assumed, and thus our proposed node architecture requires only 2 weeks early system upgrade after 6 years.

 

Fig. 11 Network performance of proposed nodes compared to the conventional nodes (5x5 poly-grid network).

Download Full Size | PDF

 

Fig. 12 Network performance of proposed nodes compared to the conventional nodes (COST266 Pan-European network).

Download Full Size | PDF

The line graphs in Figs. 11 and 12 show the total hardware scale in terms of the total number of WSSs needed in each network. Here, for the conventional networks, we assume the route-&-select architecture (Fig. 1(b)) where large port count WSSs are created by cascading 1x9 WSSs or 1x20 WSSs, while for the proposed networks, which use linearly interconnected subsystem nodes, we can assume the broadcast-&-select architecture (Fig. 1(a)) realized with 1x9 WSSs. The number of needed WSSs increases with the traffic, and conventional networks exacerbate this trend. For the proposed subsystem architecture, the increase is linear, while the conventional one makes the increase almost proportional to the square of the number of fibers. For the topologies, the hardware reduction, number of 1x9 WSSs needed, reaches 79.8-81.4% when the average wavelength path demand between each node pair is 18.

3.5 Evaluation of the number of traversed WSSs and optical loss

We investigated the number of intra-node fibers traversed by wavelength paths in the network when F_intra was set at 8. We tested the 5x5 poly-grid network and COST266 Pan-European network where the traffic volume was set to yield the blocking ratio of 0.1%. Figure 13 shows the distribution of paths that traversed different number of intra-node fibers. It is shown that more than 80% of paths do not traverse intra-node fibers, which proves the effectiveness of our proposed intra-node subsystem connection arrangement.

 

Fig. 13 Distribution of paths that traversed different numbers of intra-node fibers.

Download Full Size | PDF

We also investigated the number of WSSs traversed by wavelength paths in the network. As in the above simulation, we tested 5x5 and Pan-European networks where the traffic volume was set to yield the blocking ratio of 0.1%. Here, the F_intra value was not set constant in the network but depended on the number of the shortest hops of paths; the larger the hop numbers between source and destination nodes, the smaller was the F_intra value. This arrangement can effectively limit the largest WSS traversed number, while minimizing the routing performance degradation as can be predicted from Fig. 13, where number of the paths that traverse many intra-node fibers is quite small. Figure 14 shows the average and maximum number of WSSs traversed in each network, they are normalized against that of the single OXC architecture. Compared to the single OXC architecture, which employs the route-&-select approach and cascaded 1x9 WSSs, the average value exhibits more than 70% reduction, and the maximum value a 60% reduction. This reduction is crucial in deploying OXCs in metro areas where the number of the traversed nodes can be much larger than that of core networks [12, 14, 15]

 

Fig. 14 Number of traversed WSSs.

Download Full Size | PDF

Finally, we investigated the end-to-end total optical node loss (fiber link loss is not included); we also evaluated the average and maximum value for all paths accommodated in each network. We assumed the optical loss of 1x9 WSS as 6.5 [dB] and 1: n optical-coupler, 10log n + 1 [dB]. Figure 15(a) depicts the average value; it shows that more than 10dB reduction can be achieved with the proposed node architecture. For the maximum value (Fig. 15(b)), the optical loss of the proposed architecture is almost the same as that of conventional single OXC architecture.

 

Fig. 15 End-to-end optical loss at nodes.

Download Full Size | PDF

4. Conclusion

We investigated the performance of our recently proposed novel subsystem modular OXC node that achieves graceful modular growth without any service disruption. Numerical experiments proved that the proposed OXC enables hitless expansion with a simple procedure while reducing the hardware significantly. Its network capacity penalty is very small; only 1% even after six years of traffic increases at a rate of 30% a year. The final WSS cost reduction can reach about 80%. Our OXC nodes can also reduce the average and the maximum　number of WSSs traversed and average end-to-end optical loss for optical paths, which shows the suitability of our proposed OXC architecture in particular to metro networks. We believe the proposed OXC architecture is a powerful solution for creating large scale OXCs that are cost effective and offer improved transmission performance.