Abstract
The paper presents a method for optimising the wireless optical network that carries elastic packet traffic. The particular focus is on modelling the effect of elastic traffic flows slowing down in response to the decrease of the optical transmission systems’ capacity at bad weather conditions. A mathematical programming model of the network design problem is presented that assumes that the packet rates of elastic traffic flows decrease fairly. While practically any subset of network links can be simultaneously affected by unfavourable transmission conditions, a particular challenge of solving the problem results from a huge number of network states considered in the model. Therefore, how the problem can be solved by generating the most unfavourable network states is presented. Moreover, it is proved that it is entirely sufficient to consider only the states that correspond to the decrease of capacity on a single link. Finally, as the general problem is nonlinear, it is shown that the problem can be transformed to a linear MIP problem and solved effectively when single-path routing of traffic flows is assumed.
1. Problem Definition
The paper considers the problem of designing a wireless optical network that carries elastic packet traffic. Each node of the considered network is a packet router, while each link connecting a pair of nodes is a packet link composed of a number of optical wireless transmission systems (cf. Figure 1).
Employing optical wireless transmission systems has many advantages, the major one being that little network infrastructure is required. Thus, the network can be installed or expanded quickly and the installation process is comparatively inexpensive. The network is also very flexible as far as reconfiguration is concerned—a node can be easily reconnected to a different set of nodes and the transmission equipment can be deinstalled and moved to another location quickly. However, there are also disadvantages. For example, the transmission systems are line-of-sight systems, which means that the optical transmitter must see the optical receiver. That limits the feasible network configurations and may heavily influence the network design as shown in [1].
The most important issue, however, is that the transmitted signal is vulnerable to atmospheric conditions (smoke, fog, smog, etc.) as the systems are wireless and no transmission medium is used (cf. [2]). The worsening of atmospheric conditions may heavily influence the quality of the received optical signal. With that in mind, the optical wireless transmission systems are designed to operate in a number of transmission modes. Based on a specific scheme of signal coding and modulation, each mode provides particular capacity of the transmission system and particular robustness of the optical signal to unfavourable transmission conditions—in general, the higher the system’s bit-rate the lower the signal’s robustness. Thus, using multiple schemes effectively solves the trade-off between signal robustness and system capacity.
At perfect weather conditions the transmission systems of the wireless optical network are supposed to operate using the highest-capacity transmission mode, which provides the maximum capacity of links. But when the signal propagation conditions deteriorate, the operating mode of each affected transmission system changes to a more robust one and the capacity of network links decreases. Whenever the resulting capacity of a network link is less than the total nominal bandwidth of the information flows assigned to the link, packet losses are inevitable.
With elastic packet traffic, packet losses will cause packet sources to adapt (lower) their packet rate to match the available network bandwidth. This behaviour is due to the per session end-to-end flow control mechanisms of the TCP data transport protocol. The packet rate of the TCP session is increased linearly every time the sender application receives an acknowledgement that a packet has reached the destination, and the rate is decreased geometrically every time a packet is lost (cf. Figure 2). Since the packet losses happen randomly, arguably, the decrease of the packet rates of the sessions on a given link must be fair, meaning that the sessions with higher packet rates decrease the rate first and all sessions decrease the rate to the same value.
Changing the rate at which packets are sent directly affects the perceived quality of service and should be controlled by careful network design. The network design problem considered in this paper consists in dimensioning the links of a given optical wireless network and routing a given set of elastic traffic flows so that the total cost of links (i.e., the total cost of the optical wireless transmission systems installed on the links) does not exceed a given budget , and in nonnominal transmission conditions the bandwidth that is assigned to any single traffic session is decreased by at most a given factor , . Alternatively, the objective of the design may be to minimise the total cost of links or to maximise the minimal bandwidth reduction. It is assumed that any subset of network links can be affected by unfavourable transmission conditions and that the decrease of packet rates on an affected link is fair.
The considered problem is similar to the classical problems of survivable network design (cf. [3]). However, there are a couple of major differences. First, in this paper a max-min fair distribution of traffic is considered instead of the general one; moreover, the fairness is considered at the level of individual sessions and not aggregated traffic flows. The application of the max-min fairness concept in network design and the approaches to solving max-min fairness problems are discussed in [4, 5]. Second, designing survivable networks is based on the notion of the set of network states, which usually consists of the nominal (failure-free) state and a number of nonnominal (failure) states. And the particular issue considered in this paper (apart from the fact that links do not fail but their capacity is decreased instead) is that in the case of the optical wireless network no meaningful set of nonnominal network states can be defined (what is the extent of fog?)—actually, any subset of links can be affected by unfavourable transmission conditions, which leads to a huge number of network states. Dealing with that issue is a major problem considered in this paper.
2. Problem Modelling
Let the optical wireless network be modelled with a directed graph , with being the set of nodes and being the set of links, and let be the set of paths in . For each , let denote the set of links of path , and, for each , let denote the set of paths that use link .
Let the links of the network be realised with one type of transmission system that can operate in two transmission modes—the primary (high-capacity) mode which is used in nominal transmission conditions and the secondary (low-capacity) mode which is used when conditions deteriorate. Without loss of generality it can be assumed that the capacity of the system is equal to in the high-capacity mode and is equal to , , in the low-capacity mode. Each link of the network can be equipped with a number of transmission systems. Let the link cost function define the cost of a single transmission system installed on the link.
Let denote the set of network states. Let denote the nominal network state of perfect transmission conditions (when all transmission systems operate in the high-capacity mode) and let set of nonnominal network states represent all assumed situations of unfavourable transmission conditions; thus, . In particular, could be defined by assuming that at most links of the network can be simultaneously affected by unfavourable signal transmission conditions. For each , let be the set of affected links in nonnominal state . It is assumed that whenever transmission conditions on a network link deteriorate, all the transmission systems installed on the link operate in the low-capacity transmission mode.
Let the network traffic be modelled with a set of demands . Let function define the number of sessions that correspond to the demand, function define the nominal packet rate of a single session, and function define the average packet length for the demand (it can be noticed that is the nominal bandwidth required by a single session and is the volume of demand ). Finally, for each , let denote the set of admissible paths of demand .
For each , let variable denote the number of transmission systems installed on link . For each and , let variable denote the number of sessions of demand assigned to path , and, for each , , and (it could be assumed that , let variable denote the reduction of the packet rate of the sessions of demand assigned to path in state (if equals , the packet rate is not reduced). The network design problem can now be formulated as a mathematical programme with the following constraints:Let , , and . Let denote the feasibility set defined by constraints (1a), (1b), (1c), (1d), (1e), (1f), and (1g) being the set of triples . Seemingly, to complete the formulation of the network design problem, additional constraints are still required to express the conditions that must be satisfied by variables ; the constraints should reflect the effect of fairly slowing down the elastic traffic sessions in response to the decrease of link capacities in nonnominal states . That effect can be modelled as follows.
Consider a given nonnominal state . Consider vector . Let denote the set of affected network links, that is, , and let be the vector of their capacities: (i.e., ). For each , , and , let be a binary variable that equals if the decrease of capacity of link causes the decrease of the packet rate of the sessions of demand that are routed along path and otherwise. For each , let be a binary variable that equals if link is saturated (i.e., the total flow on the link is equal to the link’s capacity) and otherwise; the decrease of capacity of a link can cause the decrease of a session’s packet rate only if the link becomes saturated. For each , let variable denote the common packet rate of the sessions that have their rate reduced due to the decrease of capacity of link ; is also the maximum session packet rate on link . The condition for packet rate reduction can be expressed as follows: Altogether, the following constraints must hold ( and are sufficiently large constants, e.g., and ) which express the relation between packet rate reduction and capacity decrease of links from (the constraints are based on the model of the max-min fair network flows proposed in [6]):Let denote the feasible set defined by constraints (3a), (3b), (3c), (3d), (3e), (3f), (3g), and (3h), and let denote the projection of onto the set of vectors . It can be shown that set is nonempty. Actually, for a vector to satisfy constraints (3a), (3b), (3c), (3d), (3e), (3f), (3g), and (3h), vector must be max-min fair. Thus finding vector can be formulated as a max-min fair optimisation problem having vector as a feasible solution.
Assuming the objective of minimising the total cost of links, the considered network design problem, denoted by , can now be defined as follows:
3. Problem Solving
Potentially, the number of nonnominal network states is huge as each subset of links can correspond to a nonnominal network state. Thus, in general, the number of constraints and variables of problem (4a), (4b), (4c), and (4d) is also huge. Still, problem (4a), (4b), (4c), and (4d) can be solved with the constraint and column generation approach, by generating nonnominal network states (and thus the constraints and variables that correspond to those states) on the “as-needed” basis.
Problem (4a), (4b), (4c), and (4d) defined with a restricted set of nonnominal states can be treated as the master problem. Let be an optimal solution to the master problem. For given and , the slave problem consists in finding such set of links affected by unfavourable signal transmission conditions that for at least one path of a demand the reduction of the packet rate is less than ; this set corresponds to a new nonnominal network state , which is then added to set of the master problem.
For each , let be a binary variable that equals if link belongs to the required set and otherwise. For each and , let variable denote the reduction of the packet rate of the sessions of demand assigned to path in the required state, and let . Let , and for each and , let be a binary variable that equals whenever . Then, a new nonnominal network state can be found by solving the following slave problem for and which minimises the value of over all sets (it could also be assumed that ):It can be noticed that constraints (3a), (3b), (3c), (3d), (3e), (3f), (3g), and (3h) defining set can be directly embedded into the formulation of .
If the optimal objective value of slave problem is greater or equal to , the optimal solution of the master problem is the optimal solution of the considered network design problem. Otherwise, the master problem must be modified by adding variables and constraints that correspond to a new nonnominal network state : the optimal solution vector of problem (5a), (5b), (5c), (5d), (5e), (5f), (5g), (5h), (5i), and (5j) defines the values of the characteristic function of set ; that is, .
Solving problem (4a), (4b), (4c), and (4d) with state generation approach does not guarantee that considering a large number of network states can be avoided. However, careful examination of the packet rate reduction condition (2) leads to a conclusion that, actually, it is sufficient to consider only single-link states in the design problem. This fact can be expressed formally with the following proposition.
Proposition 1. Consider an arbitrary set of nonnominal network states and set such that for each and . If is a feasible solution of problem , then there exists a feasible solution of problem .
Proof. Consider a set of vectors , , such that (recall that sets are not empty). It will be proved that is a feasible solution of problem .
Due to (4a), (4b), (4c), and (4d), it must be shown that . As is a feasible solution of problem , therefore , and it is sufficient to prove that vector satisfies constraints (1e); that is, for each , , .
Assume that there exist , , and , such that . Due to (2), there must exist link such that for ; obviously, that link must be saturated in ; that is, . According to the assumption, there exists state , such that . With being a feasible solution of problem , due to (4a), (4b), (4c), and (4d), and for each and , in particular, . Therefore, due to (2), for , and thus . Then, once again due to (2) (and the fact that ), for each , , and . Thus, , which is not possible.
The intuition behind the proof of Proposition 1 is that the packet rate on any given link decreases most if that link is the only overloaded link in the network; decreasing the capacity of other links can only decrease the load on that particular link and minimise the need for packet rate decrease. Consequently, the opposite of Proposition 1 is not true as stated by the following proposition.
Proposition 2. Consider an arbitrary set of nonnominal network states and set such that for each and . If is a feasible solution of problem , then there need not exist a feasible solution of problem .
Proof. Consider a network illustrated in Figure 3, which consists of nodes , and , links and , and demands from to and from to , such that , and ,. Let , such that . Let and . Consider a solution to problem (obviously, it is the optimal solution), such that and and and (index is skipped as there is only one path for each demand): it can be checked that and (again with index skipped). However, for state , such that , link is saturated since , and inevitably the packet rates of both flows must decrease as and . Thus, and , and then .
Proposition 1 leads directly to the following conclusion.
Corollary 3. Consider an arbitrary set of nonnominal network states and set such that for each and . The optimal objective function value of problem is equal to the optimal objective function value of problem .
Proof. Due to Proposition 1, the optimal objective function value of problem is an upper bound on the optimal objective function value of problem . But it is also a lower bound since problem is a relaxation of problem (due to the fact that ).
Proposition 1 and Corollary 3 not only enable considering a limited set of network states, but also allow for simplifying constraints (3a), (3b), (3c), (3d), (3e), (3f), (3g), and (3h) defining set , due to the fact that only one link for each state needs to be analysed. First, since , the number of variables and constraints is heavily reduced, and index is actually not required. Second, as it can be noticed that it is not necessary to examine directly whether the considered link is saturated (the link must be saturated if at least one is less than ), variables and constraints (3b) and (3c) are actually not required.
Still, problem (4a), (4b), (4c), and (4d) is nonlinear due to the nonlinearity of constraint (1d). However, if single-path routing of demands is assumed, the problem can be formulated as a MIP. For each and , let be a binary variable that equals if the sessions of demand use path and otherwise. Then variables can be replaced with variables using substitution . Finally, expression can be linearised requiring that and .
Thus, problem can be treated as a single-path routing MIP problem with additional constraints. And as analysed in [5, 7] the single-path routing problem can be solved quite efficiently. Hopefully, the additional constraints are not very demanding, as, in particular, the actual values of variables are not critical—it is only important if they are greater or equal to . However, the numerical experiments that are supposed to illustrate the computational complexity of the problem will be the subject of a separate paper.
Conflict of Interests
The author declares that there is no conflict of interests regarding the publication of this paper.
Acknowledgment
This work was supported by National Science Centre, Poland, under Grant 2011/01/B/ST7/02967.