Distributed Consensus-based Auctions for Wireless Virtual Network Embedding Flavio Esposito Francesco Chiti fesposito@exegy.com Advanced Technology Group Exegy, Inc., St. Louis, MO francesco.chiti@unifi.it Department of Information Engineering University of Florence, Italy Abstract—Software-Defined Networks (SDNs) based approaches represent an opportunity for easing the deployment and the management of wide-area wireless network services. In this paper, we focus on a particular SDN management mechanism that wireless infrastructure providers need to adopt to support such services: the wireless virtual network (VN) embedding problem. We formulate the problem leveraging on optimization theory, analyzing its complexity, and proposing a general distributed auction mechanism. This mechanism leverages on the max-consensus literature to guarantee bounds on the embedding time and on the performance with respect to a Pareto optimal solution. Using extensive simulations, we confirm superior resource utilization when compared with existing distributed VN embedding solutions, proving to be an attractive and flexible resource allocation approach for wireless SDNs. Many VN embedding solutions have been proposed for wireless [8], [9] and wired distributed service architectures, although most of them with embedding goals tailored to specific applications or providers’ goals [10]. Only recently, distributed policy-based embedding solutions were designed as an effective way to customize the embedding, based on the Infrastructure Providers (InPs) wishes (see e.g., [11]–[14].) In a wireless settings however, to our knowledge, a distributed VN embedding that enables InPs to choose their own policies does not yet exist. Moreover, to our knowledge, this paper is the first to propose a wireless virtual network embedding solution that provides guarantees on both the convergence time to a VN embedding, and on the performance of their heuristics. I. INTRODUCTION To this aim, we present a Consensus-based Auction for Wireless (CAW) virtual network embedding. Auctions are an effective resource allocation mechanism when the value of the resources is unknown. In distributed settings, the bandwidth available at neighbor nodes can be at best estimated. Our CAW mechanism allows wireless physical nodes bid on the virtual resources to be embedded, providing guarantees on a Paretooptimal VN embedding solution, and on the time to reach such solution — a critical metric especially in delay-sensitive SDN applications. Moreover, wireless nodes running CAW can instantiate their own (private) policies, and adapt the VN embedding based on their goals. For example, physical nodes can (collectively or individually) choose a bidding function that would balance their load, or a adapt a “bin packing” bidding strategy to leave idle a higher number of nodes to save energy. We already presented a non-wireless version of CAW for the virtual network embedding problem [11], [12]. In this paper, we extend the problem formulation to the wireless scenario, and we show how not only the approximation algorithm, but even the resulting complexity becomes harder (from NP-hard to strongly NP-hard), but it can be reduced to polynomial in a particular case with only two wireless interfaces. Managing a distributed control plane in a wireless network is challenging, even for a WLAN with a small number of access points. Allocating radio resources on a limited spectrum, implementing efficient handover mechanisms, or balance the load between cells are only a few of these challenges. The network management complexity is then acerbated by the scalability requirements of such networks, usually centralized but deployed by several communicating processes; for example, a typical IEEE 802.11 [1] enterprise WLAN can have thousands of access points, while cellular networks serve hundreds of thousands of customers. Software Define Networks (SDNs) based approaches represent an opportunity for easing the deployment and management of such wireless network. Recent work have demonstrated how traditionally hard-to-implement features are indeed becoming a reality with the SDN-based networks. Few applications include, for example, seamless mobility through efficient handovers [2], [3], load balancing [2], [4], on demand virtual access points creation [2], [5], downlink scheduling (e.g., an OpenFlow switch can do a rate shaping or time division) [5], dynamic spectrum usage [5], or inter-cell interference coordination [3], [5]. Not only in the business community, but even in the research community wireless SDN are used to manage wide-area virtual network testbeds [6], [7], where the physical channel is sliced to allow simultaneous access to multiple researchers. Before running any of these virtual network (VN or slice) services, wireless infrastructure providers need to create such slices, solving a VN embedding problem, i.e., the (NP-hard) problem of finding, for each virtual node, a physical hosting node, and for each virtual link, at least one loop-free physical path. The rest of the paper is organized as follows: in Section II we define the wireless VN embedding problem, and we study its complexity, showing how a centralized solution can be found in O(1.344n ), where n is the number of wireless physical nodes, and only using polynomial space. We then present CAW, our policy-based distributed auction mechanism in Section III, and give examples of such policies in Section IV. We further compare the performance of representative policy instantiations with two existing distributed embedding solutions in Section VI and conclude our paper in Section VII. 2 II. T HE W IRELESS V IRTUAL N ETWORK E MBEDDING In this section we formulate the wireless virtual network embedding problem as a centralized optimization problem, and we show that it can be reduced from the edge coloring problem, therefore justifying the need for an efficient heuristic or approximation algorithm. By leveraging a recent result by Kowalik [15], we also show how, a centralized solution to the problem can be found in O(1.344n ), where n is the size of the physical wireless network, using only polynomial space. maximize Np Nv X X Uijk (xij , fk ) i=1 j=1 subject to Nv X Cj xij ≤ Ci ∀i ∈ VG j=1 Np X xij ≤ i=1 Np Nv X X 1 ∀j ∈ VH xij = Nv A. Centralized Wireless Virtual Network Embedding Given a wireless virtual network H = (VH , EH ) and a physical network G = (VG , EG ), the wireless virtual network embedding problem is the problem of finding, for each virtual node in VH , one wireless hosting physical node in VG , and for each virtual link in EH , at least one loop-free physical path, where the physical links composing the path are chosen from the set EG . A wireless (e.g., Wi-Fi) virtual network embedding has a crucial additional constraint: every physical node in the multigraph G representing the physical network can only have a limited (three in the case of Wi-Fi) outgoing edges, as there are only limited (three) wireless network interfaces in the wireless (IEEE 802.11) standard that allow interference-free transmissions. 1 We assume that each wireless physical node i ∈ VG , where |VG | = Np belonging to some InP, has a positive utility Uijk when hosting a virtual node j ∈ VH , where |VH | = Nv , or when relaying traffic within the flow fk , where k ∈ K is the set of the loop-free physical paths in G. We also assume that InPs cooperate to provide a wide-area service, seeking a Pareto optimality, i.e., they aim to maximize the summation of their utilities, without behaving selfishly. We then model the wireless virtual network embedding problem with the centralized optimization Problem (1a), where xij = 1 if virtual node j is assigned to physical node i and 0 otherwise. The vector xi ∈ {0, 1}VH , whose elements are the xij variables, represents the assignments for physical node i. The first three constraints express, respectively, the requirements that the number of virtual node capacity assigned to a wireless physical node i should not exceed its physical rate capacity Ci , each virtual node should be assigned to at most one wireless physical node, and all virtual nodes should be assigned to at least a physical node. The fourth constraint ensures that no two adjacent interfaces use the same frequency z ∈ Φ, to avoid interference: Iz (u) is an indicator variable equal to one if physical node u transmits using frequency z, and zero otherwise. The remaining equations are the existential, and the flow conservation constraints. Such constraints ensure that the net flow on each physical link is zero, except for the source sk and the destination tk , that have virtual link demand dk . 1 In the rest of the paper we use wireless and Wi-Fi interchangeably. i=1 j=1 Φ X (1a) Iiz ≤ 1 ∀ i ∈ VG z=1 K X fk (u, v) ≤ C(u, v) ∀ (u, v) ∈ VG k=1 X fk (u, w) = 0 ∀k ∀ u 6= si , ti w∈W fk (u, v) = −fk (v, u) ∀k ∀ (u, v) ∈ VG X X fk (sk , w) = fl (w, tk ) = dk ∀ sk , tk w∈W w∈W xij = {0, 1}, fk ≥ 0 ∀k, ∀i ∈ VG , ∀j ∈ VH B. Wireless Virtual Network Embedding Complexity From the optimization problem defined in the previous section we note how the problem is at least NP-hard. This is because the first set of constraints (the first inequality) is equivalent to the capacity constraint of the set packing problem, known to be an NP-hard problem. Moreover, the flow constraints represent the classical multi-commodity flow constraints, which is known to be solvable in polynomial time, if each virtual link (flow) is allowed to be split among two or more loop-free physical paths, and otherwise known to be strongly NP-hard. We assume that our physical network runs the one of the IEEE 802.11 b/g/n protocols whose frequency spectrum is 100 MHz wide and made up of 11 channels centered 5 MHz apart. Each 2.4 GHz channel is 20 - 22 MHz wide, making the spectrum crowded. In the case of Wi-Fi, channel overlapping is undesirable. This is why, to be able to transmit data without interference, once a virtual network is formed, each physical nodes can only typically use at most three channel interfaces: the interface to channel 1,6, and 11. Given a physical network where each physical node has these three interfaces, each virtual node needs to be mapped to an interface such that the physical outgoing edges do not use the same channel. Therefore the following result holds: Theorem II.1. The 3-channels wireless VN embedding problem is NP-complete. Moreover, it is possible to test whether a 3-channels embedding exists in time O(1.344n ), where n is the size of the physical network, only using polynomial space. 3 Initialization Bundle Construction (Sub-modular Utility) Communication Winners & bundles Conflict Resolution (Assignment Rules Table) N Consensus reached? Y End Fig. 1. CAW workflow: a virtual link embedding phase (k-shortest path) follows the virtual node bidding and agreement phases (virtual node embedding). Proof. (sketch) Let us first show that the problem is NPComplete. Given a multigraph G = (VG , EG ), with a set of vertices VG , and a set of edges EG , the graph coloring problem is the NP-complete problem of assigning a color to each edge so that no two adjacent edges share the same color. Let us consider G to be the physical hosting network and let three colors representing channel 1, 6 and 11 of the IEEE 802.11 b/g/n spectrum. In polynomial time, an algorithm A takes each virtual link of the virtual network and assigns one of the three colors to it at random. Given a virtual network embedding solution, we can check in polynomial time if the edge coloring problem is satisfied on the hosting physical network G, hence our problem is NP-complete. The remaining part of the theorem can be proved by construction, and we omit it for lack of space, since it is a direct corollary of the recent Theorem 2 of Kovalik [15]. For delay-sensitive wireless applications, maximizing the convergence time —the time it takes to find an embedding may be more important than having the physical network fully utilized. If we are willing to avoid transmissions on one of the three channels, then the wireless virtual network embedding problem can be solved in polynomial time. It is in fact trivial to verify if a graph can be edge-colored with (one or) two colors only — we only have to verify the capacity constraints. III. C ONSENSUS - BASED AUCTIONS FOR W IRELESS N ETWORK E MBEDDING In this section we present CAW network embedding mechanism. Auctions are a well known approach to allocate goods when the value of such goods is unknown to the bidders, or when the bidders’ budgets are unknown to the auctioneer. In distributed settings, the congestion level of the network is unknown a priori, or at best, estimated. A distributed auction on the flows (virtual links) to be embedded (allocated) on a possibly congested software-defined wireless network is hence an effective congestion avoidance mechanism. The mechanism. CAW consists of two phases: a bidding, and an agreement phase. The two phases iterate synchronously, that is, a second bidding phase for a virtual link (flow) does not start until the first agreement phase on a previous flow terminates. Physical nodes act upon messages received at different times during each consensus phase; therefore, the consensus phase is asynchronous. This is also different from the results presented in [11] and [12], where the consensus phase was assumed to be synchronous. In the rest of the paper, we denote such rounds or iterations with the letter t and we omit t when it is clear from the context. Bidding phase (Algorithm 1.) Each wireless physical node i ∈ I uses a private utility function (see Equation ??) to bid on a {flow, frequency} tuple j ∈ F, and stores all bids in |F | a vector bi . Each element bij ∈ bi ∈ R+ is a positive real number representing the highest bid known so far on the {flow, frequency} tuple j. The mechanism is general enough to work with any private utility function, but in this paper we assume that the wireless physical nodes bid higher values if their residual congestion on a given physical flow is higher, and if they have higher residual capacity. In particular, we consider the notion of stress as the product of the physical node capacity, times the residual available capacity on all the adjacent physical outgoing links. The lower is the stress of a wireless physical node, the higher will be its bid. The wireless physical nodes store the IDs of the flowfrequency tuple on which they are bidding in a list (bundle vector) mi ∈ F Ti , where Ti is the target number of flows that i can allocate, F the set of available frequencies, and |mi | is the number of flows that a node are allowed to bid simultaneously. Wireless physical nodes need to be also aware |F | of a mapping vector fi ∈ R+ (with size equal to the number of flows to be embedded) between the frequencies (currently) assigned by the winner physical nodes and their ID. Agreement or Consensus phase (Algorithm 2.) After the private bidding phase, each physical node exchanges the bids |F | with its neighbors, updating the assignment vector ai ∈ VG and the flow vector f with the latest information on the current assignment of all flows, for a distributed auction winner determination. We assume that nodes are aware of its firsthop neighbors, and the routing table. A. Conflicts resolution When it receives a bid update, a node i has three options: (i) ignore the received bid leaving its bid vector and its allocation vector as they are, (ii) update according to the information received, i.e. wij = wkj and bij = bkj , or (iii) reset, i.e. wij = ∅ and bij = 0. When |mi | > 1, the bids are not enough to determine the auction winner as virtual nodes can be released, and a node i does not know if the bid received has been released or is outdated. Bids should be ignored or reset if they are outdated, and subsequent bids to an outbid virtual node should be released, to maximize the summation of the utilities. To avoid storing bids computed with out-of-date utility values, physical nodes simply reset their own bundle at the beginning of every bidding phase (Algorithm 1, line 4). Remark (reducing the communication and runtime complexity.) In order to reduce the communication cost, we let the nodes send their values updates only when they first learn 4 Algorithm 1: bidding for wireless node i at iteration t 1: Input: ai (t − 1), bi (t − 1) 2: Output: ai (t), bi (t), mi (t), fi (t), 3: ai (t) = ai (t − 1), bi (t) = bi (t − 1) 4: mi (t) = ∅, fi (t) = ∅ 5: if biddingIsNeeded(ai (t), Ti ) then 6: if ∃ j : hij = I(Uij (t) > bij (t)) 6= 0 then 7: η = argmaxj∈F {hij · Uij } 8: mi (t) = mi (t) ⊕ η // append η to bundle 9: biη (t) = Uiη (t) 10: update(η, ai (t)) 11: update(η, fi (t)) 12: Send / Receive bi to / from k ∀k ∈ Ni 13: Send / Receive ai to / from k ∀k ∈ Ni 14: Send / Receive fi to / from k ∀k ∈ Ni 15: end if 16: end if Algorithm 2: agreementPhase for node i at iteration t 1: Input: ai (t), bi (t), mi (t), fi (t) 2: Output: ai (t), bi (t), mi (t), fi (t) 3: for all k ∈ Ni do 4: for all j ∈ F do 5: if IsUpdated(bkj ) then 6: update(bi (t), ai (t), mi (t), fi (t)) 7: end if 8: end for 9: end for about them, not at every round. This strategy is, however, not sufficient to decrease the order of magnitude of the worst case message overhead [16]. In a system implementation, the message overhead could be further reduced by compressing the attributes to be sent. IV. T HE P OLICIES One of the design goals of CAW is its flexibility, i.e., the ability to create customizable allocation algorithms to satisfy desired policies, rules, and conditions in any wireless settings. We describe here few representative examples of such policies, and later in this section we show how CAW can be instantiated to satisfy providers’ embedding goals. A straightforward example of policy is the (normalized) utility function U that InPs use to bid on virtual resources (nodes.) In our evaluation (Section VI) we use: X Ti − Sij Ti = Ci + Cik , Uij = (2) Ti k∈Ni where Ti is the target virtual (node and links) capacity that is allocatable on i, and Sij the stress on physical node i, namely, the sum of the virtual node capacity already allocated on i, including virtual node j on which i is bidding, plus the capacity of the virtual links allocated on the adjacent physical links in the previous round. Note that, due to the max consensus definition [16], the bid bij at physical node i on virtual node j is the maximum utility value seen so far. The normalization factor T1i ensures that such bids are comparable across physical wireless nodes. We have seen from related work, e.g. [17], [18], how embedding protocols may require SPs to split the VN. CAW is able to express this requirement by enforcing a limit on the length of the bid vector bi , so that physical nodes bid only on the released VN partition. Each InP can also enforce a load target Ti on its resources, by limiting the bundle size. Another policy is the assignment vector ai , that is, a vector that keeps track of the current assignment of virtual nodes. ai may assume two forms: least and most informative. In its least informative form, ai ≡ xi is a binary vector where xij is equal to one if physical node i hosts virtual node j and 0 otherwise. In its most informative form, ai ≡ wi is a vector of winning physical nodes so far on virtual nodes; wij represents the identifier of the physical node that made the highest bid so far to host virtual node j. Note that when ai ≡ wi the assignment vector reveals information on which physical nodes are so far the winners of the auction, where if ai ≡ xi physical node i only knows if it is winning each resource or not. As a direct consequence of the max consensus, this implies that when the assignment (allocation) vector is in its least informative form, each physical node only knows the value of the maximum bid so far without knowing the identity of the bidder. We also leave as a policy whether the assignment vector is exchanged with the neighbors or not. In case all physical nodes know about the assignment vector of the virtual nodes, such information may be used to allocate virtual links in a distributed fashion. Instead, if ai ≡ xi , to avoid physical nodes flooding their assignment information, i asks the SP about the identity of the physical node hosting the other end of the virtual link and attempts to allocate at least one loop-free physical path. V. C ONVERGENCE AND P ERFORMANCE G UARANTEES All wireless nodes need to be aware of the mapping, by exchanging their bids with only their first-hop neighbors, therefore a change of information needs to traverse all the physical network, which we assume has diameter D. This result states that D hops are also enough, that is, a necessary and sufficient condition to reach max consensus on a single virtual node allocation is that each node is visited once. An interesting observation that follows from the result is that the number of steps for CAW to converge on the embedding of a slice of |VH | virtual nodes is always D · |VH | in the worst case, regardless of the size of the bundle vector. This means that the same worse-case convergence bound is achieved if CAW runs on a single or on multiple virtual nodes simultaneously. Formally, we have the following result: Proposition V.1. (Convergence of CAW.) Given a virtual network H with |VH | virtual nodes to be embedded on a physical wireless network with diameter D, the utility function of each physical node is sub-modular, and the communications occur over reliable channels, then the CAW mechanism converges in a number of iterations bounded above by D · |VH |. 5 We omit the proof of as it is equivalent to the proof obtained for wired VN embedding [11]. We showed in Section III that the problem of embedding a wireless virtual network becomes intractable as the size of the physical or virtual network grows. In general, this means that the sum of the utility of the providers after an embedding could be arbitrarily low. However, because we assumed a decreasing marginal gain in adding new virtual nodes to the previously allocated, this does not happen as we inherit the following bound: Proposition V.2. (CAW Approximation.) The CAW embedding algorithm yields an (1 − 1e )-approximation w.r.t. the optimal node assignment solution, and w.r.t. the VN embedding, if we consider physical networks whose diameter D is at most one. We omit also this proof of as it is equivalent to the proof obtained for wired VN embedding [12]. Note how, although virtual nodes and links are tightly connected, the wireless node embedding is the only computationally intractable subproblem if we allow path splitting. CAW splits the embedding in two phases (see Figure 1), so once there is agreement on which wireless physical nodes will host the source and the destination virtual node, for each virtual link, the loop-free physical path assignment phase (run as a k-shortest path) can be found in polynomial time, as a result of a fractional multi-commodity flow problem. A virtual link can be in fact mapped onto multiple physical paths [19]. VI. E XPERIMENTAL R ESULTS To test the proposed distributed consensus-based auction algorithms, we developed our own trace-driven simulator. Our results confirm the theoretical bounds and show how a representative CAW policy instantiations outperform similar (policy-based) distributed virtual network embedding solutions. Although they were not designed for wireless networks, we have chosen to compare CAW against the first VN embedding solution [18], and against PolyVine [14], to our knowledge, the first distributed policy-based VN embedding solution. A. Simulation Environment Physical Network Model: We use the BRITE topology generator [20], to build a flat topology using either the Waxman model [21], or the Barabasi-Albert model [22] with incremental growth and preferential connectivity, respectively. We tested our algorithms with various physical network sizes, varying the number of physical nodes and physical links (as in [19].) Our simulations do not consider delay constraints, while link capacity constraints are discussed later in this section. The results are similar regardless of the topology generation model and the physical network size, and we only show the results obtained for physical networks of size n = 50 and a BarabasiAlbert physical topology. Virtual Network Model: we use a real dataset of 8 years of Emulab [23] requests. For each simulation run, we sample a request from a dataset of 61968 requests; the average size of a request is 14 with standard deviation of 36 virtual nodes; 99% of the requests have less than 100 virtual nodes, and 85% have at most 20 virtual nodes. Excluding the 10% long-lived requests that cause the standard deviation of VN lifetime to exceed 4-millions seconds, the duration of the requests is on average 561 with 414 seconds of standard deviation. As the dataset does not contain neither the number of virtual links nor the virtual network topology, we connect each pair of virtual nodes at random with different average node degree (Figure 2b,c.) All our simulation results show 95% confidence intervals; the randomness comes from both the virtual network topology to be embedded, and the virtual constraints, that is, virtual node and link capacity requirements. Similarly to previous work [19], we let capacity requirements be real numbers uniformly distributed between 1 to 100 units. We impose both virtual-physical node and link average capacity ratio R = {50, 100, 500}. The results are similar and we only show plots for R = 100. Moreover, we add a node capacity physical constraint to make sure each physical node supports at least the capacity of its adjacent physical links. Comparison Method: we compare our CAW mechanism, instantiated instantiated with different policy configurations, with another policy based distributed virtual network embedding algorithm, PolyViNE [14], and with the first published distributed virtual network embedding algorithm [18], that we call Hub and Spoke due to the adopted heuristic, that partition the VN to be embedded in hub and spoke partitions, and tries to embed sequentially all the VN partitions. A VN is rejected as soon as one partition cannot be embedded B. Simulation results Our main simulation results show how CAW policies that balance the load on the wireless physical network (e.g., SAD, as in Single-item Auction for Distributed embedding) are more effective than bin packing policies (e.g., MAD, as in Multiitem Auction for Distributed embedding.) This is because the only three available interfaces (channels 1,6 and 11 of IEEE 802.11) constraint the physical network multigraph (a graph with multiple edged between the nodes) so that the outgoing link capacity is quickly exhausted, if too many virtual nodes are mapped to the same wireless physical node (Figure 2b.) This result is confirmed by the physical node utilization (Figure 2d), lowest for the SAD policy. We also show results for a path bidding policy (PAD), in which physical nodes bid on virtual paths of size p, to be embedded on physical loop-free paths of size less or equal than p, to avoid physical node relays (a drawback of the SAD and MAD policies). The efficiency of a wireless embedding is also evaluated with the VN allocation ratio — ratio between VN successfully embedded and requested, and with the resource utilization — physical node and link capacity utilized to embed the VN requests, as well as with the endurance of the algorithm, i.e. the number of successfully allocated requests before the first VN request is rejected (Figure 2c). Finally, we evaluate the convergence time, — number of algorithm steps until a valid 6 60 40 20 0 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 size of VN request 0.8 SAD MAD PAD Hub & Spoke PolyVINE 0.6 0.4 0.2 0 3 4 5 6 Average Virtual Node Degree (a) 7 (b) SAD MAD Hub & Spoke PolyVINE 100 80 60 40 20 0 5 10 15 Virtual Network Size [# of nodes] (c) 20 Physical Node Utilization 80 Continuously Allocated VNET 1 Slice Allocation Ratio Convergence Time Steps 100 1 0.8 0.6 0.4 0.2 0 S MH P S MH P S MH P S MH P Distributed Slice Embedding Algorithm (d) Fig. 2. (a) The number of steps required for CAW to convergence is linear in the size of the VN as shown theoretically. (b) CAW’s instantiated with the SAD policy allocates more slices since it balances the load on the physical network. A single shortest path is available. (c) CAW’s instantiated with the SAD policy is able to allocate more VN before the first one is rejected. (f) CAW’s instantiated with the SAD policy better balances the load on physical nodes. S, M, H and P indicate SAD, MAD (the two CAW’s policies), Hub and Spoke and PolyViNE, respectively. embedding is found (Figure 2d), confirming our theoretical results (Proposition V.1). VII. C ONCLUSIONS We presented the problem of embedding a virtual network on a wireless physical network, via a centralized optimization problem, and we studied its complexity. We then focus on a specific wireless protocol, the IEEE 802.11, that has only three available transmissions channels. This assumption led to a simplification, with respect to the general virtual network embedding problem. In particular, the multigraph representing the wireless physical network can have at most three edges. This restriction allowed us to show, leveraging on a known solution to the edge-coloring problem, that the time complexity of finding a centralized solution is O(1.344n ), where n is the size of the physical network. After motivating the need for a polynomial time algorithm to solve the problem, we proposed CAW, a distributed approximation algorithm, that leverages our previous results on VN embedding to converge to a conflict-free assignment of virtual nodes and links on the wireless physical network. We then developed our own event-driven simulation and showed that CAW outperforms existing distributed VN embedding solutions, in terms of VN allocation ratio and in terms of physical network utilization. R EFERENCES [1] IEEE. 802.11. [Online]. Available: http://standards.ieee.org/about/get/ 802/802.11.html [2] L. Suresh, J. Schulz-Zander, R. Merz, A. Feldmann, and T. Vazao, “Towards Programmable Enterprise WLANS with Odin,” in Proc. of the First Workshop on Hot Topics in SDN, ser. HotSDN ’12. ACM, 2012, pp. 115–120. [Online]. Available: http://doi.acm.org/10.1145/ 2342441.2342465 [3] L. E. Li, Z. M. Mao, and J. Rexford, “Toward software-defined cellular networks,” in Proc. of European Workshop on SDN, ser. EWSDN ’12, 2012, pp. 7–12. [4] A. Gudipati, D. Perry, L. E. Li, and S. Katti, “Softran: Software defined radio access network,” in Proc. of the ACM SIGCOMM Workshop on Hot Topics in SDN, ser. HotSDN ’13. ACM, 2013, pp. 25–30. [5] J. Vestin, P. Dely, A. Kassler, N. Bayer, H. J. Einsiedler, and C. Peylo, “CloudMAC: towards software defined WLANs,” Mobile Computing and Communications Review, vol. 16, no. 4, pp. 42–45, 2012. [6] The GENI testbed http://www.geni.net. [7] Raychaudhuri, D. et al., “Overview of the ORBIT radio grid testbed for evaluation of next-generation wireless network protocols,” in Wireless Comm. and Networking Conf., 2005 IEEE, vol. 3, March 2005, pp. 1664–1669 Vol. 3. [8] Leivadeas, A. et al, “An architecture for virtual network embedding in wireless systems,” in Network Cloud Computing and Applications (NCCA), 2011 First International Symposium on, Nov 2011, pp. 62–68. [9] D. Yun and Y. Yi, “Virtual network embedding in wireless multihop networks,” in Proceedings of the 6th International Conference on Future Internet Technologies, ser. CFI ’11. New York, NY, USA: ACM, 2011, pp. 30–33. [Online]. Available: http://doi.acm.org/10.1145/ 2002396.2002404 [10] F. Esposito, I. Matta, and V. Ishakian, “Slice Embedding Solutions for Distributed Service Architectures,” ACM Computing Surveys, Volume 46, Issue 1, March 2014). [11] Esposito, Flavio and Di Paola, Donato and Matta, Ibrahim, “A General Distributed Approach to Slice Embedding with Guarantees,” in Proc. of the IFIP Networking, 2013. [12] F. Esposito, D. Di Paola, and I. Matta, “On distributed virtual network embedding with guarantees,” ACM/IEEE Trans. on Networking (to appear), Nov 2014. [13] F. Esposito and I. Matta, “A decomposition-based architecture for distributed virtual network embedding,” in Proc. of the 2014 ACM SIGCOMM Workshop on Distrib. Cloud Computing, ser. DCC ’14. New York, NY, USA: ACM, 2014, pp. 53–58. [14] Chowdhury, M. et.al, “PolyViNE: Policy-Based Virtual Network Embedding Across Multiple Domains,” ser. SIGCOMM VISA, 2010. [15] L. Kowalik, “Improved edge-coloring with three colors,” Theoretical Computer Science, vol. 410, no. 38-40, pp. 3733–3742, 2009. [16] N. A. Lynch, Distributed Algorithms, 1st ed. Morgan Kauf., Mar. 1996. [17] Y. Zhu, R. Zhang-Shen, S. Rangarajan, and J. Rexford, “Cabernet: Connectivity Architecture for Better Network Services,” in Proceedings of the 2008 ACM CoNEXT Conference, ser. CoNEXT, 2008. [18] I. Houidi, W. Louati, and D. Zeghlache, “A Distributed Virtual Network Mapping Algorithm,” in ICC ’08. IEEE International Conference on Communications, May 2008, pp. 5634 –5640. [19] M. Yu, Y. Yi, J. Rexford, and M. Chiang, “Rethinking Virtual Network Embedding: Substrate Support for Path Splitting and Migration,” SIGCOMM Comput. Commun. Rev., vol. 38, no. 2, pp. 17–29, 2008. [20] Medina, A. et al, “BRITE: An Approach to Universal Topology Generation,” in Proc. of the Inter. Symp. in Modeling, Analysis and Sim. of Computer and Telecom. Systems, ser. MASCOTS ’01. IEEE Computer Society, 2001, pp. 346–. [21] B. Waxman, “Routing of Multipoint Connections,” Selected Areas in Comm., IEEE Journal on, vol. 6, no. 9, pp. 1617 –1622, dec. 1988. [22] A. L. Barabasi and R. Albert, “Emergence of Scaling in Random Networks,” Science, vol. 286, 1999. [23] B. White, J. Lepreau, L. Stoller, R. Ricci, S. Guruprasad, M. Newbold, M. Hibler, C. Barb, and A. Joglekar, “An Integrated Experimental Environment for Distributed Systems and Networks,” SIGOPS Operating System Review ’02.