Multilayer Protection with Availability Guarantees in Optical WDM Networks

Abstract

Survivability is a key concern in modern network design. This paper investigates the problem of survivable dynamic connection provisioning in general telecom backbone networks, that are mesh structured. We assume differentiated services where connections may have different availability requirements, so they may be provisioned differently with protection (if needed) based on their availability requirements and current network state. The problem of effectively provisioning differentiated-service requests, that has been widely investigated for connections routed at the physical layer, assumes peculiar features if we consider sub-wavelength requests at the logical layer that have to be protected (or more generically, whose availability target has to be guaranteed), but also have to be groomed for an efficient use of network resources. An integrated multilayer approach is necessary that considers requirements and grooming of connections at the logical layer as well as their routing and availability at the physical layer. Joint availability-guaranteed routing and traffic grooming may lead to a negative interaction, since the objective of the first problem (guaranteeing a given level of availability to the connections) clashes with the objective of the other problem (minimizing resource consumption). For a multilayer WDM mesh network, we propose new multilayer routing strategies that perform effective availability-guaranteed grooming of sub-wavelength connections. These strategies jointly considers connection availability satisfaction and resource optimization and are developed under two different practical hypotheses: guaranteed target, i.e., a connection is routed only if its availability target is satisfied, and best-effort target, a connection is always routed and, when the availability target cannot be guaranteed, the path with the best possible availability is provisioned. Numerical results are reported and discussed for the two approaches mentioned above. In both cases, the results show high effectiveness of our provisioning strategy.

Appendix: On the Availability Evaluation for Shared Protected Connections

There is a relevant body of literature developed in these last years, regarding analytical models for availability calculation for protected connections, based on the classical reliability theory [23]. In the case of dedicated protection, the evaluation of the availability of the protected connection can be easily obtained with the following expression

$$ A= A_w + A_b(1-A_w) $$

where A _w and A _b represent the availability of the working and the backup path, respectively.

On the contrary, various methodologies have been proposed to solve the problem of availability evaluation with different degree of complexity and precision. In fact, the availability calculation in the case of shared protection has been proven to be NP-complete, and we have to resort to approximations to find practical models for the availability calculation.

The key problem to be solved is how to account for the connections that are “conflicting” with the backup path of the connection under analysis (say b), i.e. those connections that in case of failure (single or multiple) compete with b for the same shared backup capacity. Intuitively, the count of the conflicting connections becomes more and more complex for an increasing number of concurrent failures, and the corresponding number of possible conflict scenarios to be considered increases exponentially. In conclusion, it is not computationally feasible to account for all the possible scenarios.

A first approximated formulation appeared in [24], where the authors consider as “conflicting” connections all the connections that are sharing with b some capacity along their backup path (the set of these connections is referred to as Protection Group). In this case, the formula for the availability A _i of the generic protected connection i is given by

$$ A_i \approx A_{w_i}+\left(1-A_{w_i}\right)A_{b_i} \prod_{\forall h \in PG_i (h \neq i)}A_{w_{h}} $$

(2)

where A _{w_h} is the availability of the working path of a generic connection h belonging to the protection group PG _i of connection i.

An extension of this model can be found in [25], where a more accurate count of the conflicting connections is proposed. The study in [25] recognizes that only some of the links along the working paths w _h in the PG _i do actually induce a conflict in case of failures, and that only those links have to be accounted for. The formula 2 is evolved in

$$ A_i = A_{w_i}+ A{b_i}A{c_i} \cdot A_{w_i}A{b_i}A{c_i} $$

where A c _i is the product of the availability of all the links conflicting according to the new definition. The authors propose a data structure, called conflict vector, to evaluate the set of conflicting links c _i given a connection i. Please refer to [25] for further details.

Also in [11] and [28], the authors propose similar approaches. They use the concept of grabbing probability, i.e. the probability that the connection “grabs” the shared capacity in case of contention. We omit further details for these two works since they are based on a similar formulations to the one in [25], and are leading to a similar performance.

A completely different approach is proposed in [26]. The authors investigate a two-step approach: a first step to be executed once, during the network planning phase and a second step to be applied for each connection whose availability is currently evaluated. The calculation is based on a Markov chain model. An extension of this work is reported in [27], where an investigation of the upper and lower bounds of the previous analytical approach is conducted.

Finally, an extremely precise methodology is proposed in [17]. This work is the first one to apply the concept of conflicting links to each link of the working path. In other words, for each link of the working path the authors evaluate which other links would compete (or “conflict”) with it in case of double failure, while previously, as in [25] and [26] conflicting links where considered as conflicting with any link of the working path. Now a set conflicting links is associated to each link of the working path and not generically to the working path. We will see that this novelty avoids a significant overestimation of connection unavailability.

For this method, each of the competing connections has a probability 0.5 to acquire the shared capacity. Formula 3 describes the calculation in [17]. For each link i on the working path, a “modified” unavailability is calculated. Consider that Cond _CT indicates the set of links conflicting with i, i.e., those links that in case of another concurrent failure together with i would cause the loss, for the connection in exam, of a traffic quantity equal to NT, over a total amount of traffic d _s,t , then:

$$ U_i^* = U_i \cdot \left[ \sum_{j \notin W \cap i \in B} U_i + 0.5 \cdot \sum_{j \in Cond_{CT}} \left( U_j \cdot \frac{NT_{i,CT}^{i,j}}{d_{s,t}} \right) \right] $$

(3)

returns U ^* _i as the modified link availability that captures the likelyhood of a conflict (note that U _i is the actual link availability). Once the U ^* _i has been calculated, under the approximation of rare-events, the entire path unavailability can be expressed as:

$$ U_{path} = \sum_{i=1}^{N} U_i^* $$

where N is the number of links in the working path. Please refer to [17] for further details.

It is worth mentioning that for all these works, the main approximation is that no more than two concurrent failures are explicitly taken into account (as in [25]).

In Fig. 12, we compare some of the approaches described above over a simple network topology: a single protected connection consists of a working path with two links and and a protection path with three links. The three backup links can be shared by no other connections up to ten other connections (x-axis in the Fig. 12). All the links and sharing connections have an availability equal to 0.9. Monte Carlo simulation [24] provides a benchmark for all the approaches. The approach in [17] returns the estimation that is closest to the actual value of availability. So, in our simulation in Sect. 5, we decide to use the approach in [17].

Fig. 12

Comparison of the availability estimation provided by the various availability models provided in this section (U _link = 0.1)