BaProbSDN: A Probabilistic-based QoS Routing Mechanism for Software Defined Networks Ahmed Al-Jawad, Ramona Trestian, Purav Shah and Orhan Gemikonakli Design Engineering and Mathematics Department, School of Science and Technology, Middlesex University, London, UK E-mails:
[email protected], {r.trestian, p.shah, o.gemikonakli}@mdx.ac.uk Abstract—Over the past decade there has been an exponential increase in the Internet traffic especially with the proliferation of cloud computing and other distributed data services. This explosion of data traffic with its dynamically changing traffic patterns and flows might result in degradation of the network performance. In this context, there is a need for an intelligent and efficient network management system that delivers guaranteed services. To this extent, this paper proposes BaProbSDN, a probabilistic Quality of Service (QoS) routing mechanism for Software Defined Networks (SDN). The QoS routing algorithm employs the bandwidth availability metric as a QoS routing constraint for unicast data delivery. BaProbSDN makes use of Bayes’ theorem and Bayesian network model to determine the link probability in order to select the route that satisfies the given bandwidth constraint. The performance of the proposed probabilistic QoS routing algorithm was tested in a simulationbased environment and compared against the widest-shortest path routing (WSR) algorithm. The results demonstrate that BaProbSDN can achieve up to 8.02% decrease in the bandwidth blocking rate when compared to WSR in the presence of link update inaccuracies of threshold and time delay. Keywords—Software Defined Networks, OpenFlow, Quality of Service Routing, Bayesian Network
I.
INTRODUCTION
In the recent years, the networking environment witnessed a rapid evolution in applications like cloud computing. As service diversity widened in general, the access to centralised resources and services contributed to an increase in data traffic volume. This lead the technology to impose the necessity for intelligent network state and data flows monitoring solutions in order to meet both the customer and the service needs. However, the traditional approaches are not enough to accomplish this, and new ways of managing the network components and balancing the data traffic need to be investigated. Recently, both academia and industry have paid great attention at adopting the Software Defined Networks (SDN) concept [1], which is seen as a new paradigm that enables virtual, dynamic, and flexible networking. The concept of SDN is to separate the control plane from the data plane giving the network administrators a more fine grained control over the traffic flows. In this context, a central SDN controller is used to administrate all the switches within a network. The switches handle the data-path functionality while the controller, a standard server, will handle the high-level networking decisions (e.g., routing, load balancing, failure recovery, etc.). c 978-1-4799-7899-1/15/$31.00 2015 IEEE
Additionally, Quality of Service (QoS) data delivery has become an active field of research especially for the applications that require data delivery under certain QoS constraints (e.g., multimedia and voice data). To this end, the QoS routing can make use of the probability distribution information to find the most probable feasible path that has the best chance to satisfy the QoS constraints. To date, the traditional QoS routing approaches mainly rely on the instantaneous metric observations without the consideration of imprecision in the state information. However, in general, the global knowledge of the network state information contains uncertainties caused by various perturbation factors, such as: the link state update policy, link propagation delay, synchronization of switches’ state update, hierarchical aggregation, metric computation methods, etc. Thus, the employment of probabilistic schemes into QoS routing might be favourable to the traditional costoptimization solutions. To this extent, this paper proposes BaProbSDN, a probabilistic-based QoS routing mechanism for Software Defined Networks that makes use of Bayes’ theorem and Bayesian network to determine the link probability. The paper considers the amount of overhead while evaluating the QoS performance in terms of bandwidth blocking rate. The frequent updates exchanged between the switches and the controller in SDN introduces overhead into the network and it can be resource consuming especially in large networks. Therefore, the introduction of link state update policies reduces the amount of state information packets exchanged between the control and data planes. II.
RELATED WORKS
The rapid adoption of Software Defined Networking brought significant changes in today′ s enterprise datacenter networks architectures and various revenue models appeared. SDN enables now the network operators to control and efficiently move the data flows through the network in order to achieve load balancing and efficient use of the network resources. Thus, new ways of managing the network components and the traffic are enabled through SDN. Within the research community, more and more SDN solutions and deployments are expanding, especially for QoS routing. The authors in [2], [3] present an end-to-end QoS routing solution for multimedia delivery over SDN. The proposed routing algorithm considers the packet loss as the QoS metric for feasible path computation. However, the impact of uncertainty in the QoS
parameters was not investigated on routing accuracy. A similar solution was presented in [4] where the authors discuss the adaptation of classical QoS techniques for SDN. However they did not quantify the performance of their approach. In the case of traditional networking, numerous researches have investigated the design and implementation of various QoS routing approaches. The work in [5], [6] provides a comprehensive review of different QoS routing algorithms. The authors identify the existing problems in QoS source-based routing, such as: the management of the global network state information update which might degrade the performance of QoS routing. However some of these problems need to be considered with respect to the SDN. The relationship between the probability methods and QoS routing was also investigated in other works. Ghosh et al. [7] discussed the probability approaches using the hierarchical QoS routing. On the other hand, [6], [8] introduced the theoretical derivation of the probability approach for the endto-end delay with bandwidth constraint. The authors concluded that the end-to-end delay probability is more complicated in solving than the bandwidth problem alone. This is because the delay metric is additive, while the bandwidth metric must satisfy the minimum bound along the QoS path. In this paper, we explore the use of probabilistic-based QoS routing in the context of SDN and we propose BaProbSDN, a probabilistic-based QoS routing mechanism. Previous researches introduced the solution of finding the most probable path that satisfies the QoS requirements [6], [8], however the probability was obtained by different schemes [7], [9]. In this work, we propose a method that makes use of Bayes’ theorem and Bayesian network chain to obtain the probability of the link bandwidth availability. Generally, the traditional cost minimization-based methods can outperform the other methods if the network maintains an accurate information state [9]. However, in practice this is not the case, as networks are in general large and by their nature there will always be uncertainty in the state information due to many factors, such as: the large propagation delay along the link or aggregation [8].
III.
PROPOSED SYSTEM ARCHITECTURE
Figure 1 illustrates the proposed functional units built on top of the classical SDN architecture. The approach exploits the benefits of SDN in terms of centralised management and configuration. The communication between the SDN switches and the SDN controller is done using the OpenFlow protocol. In the forwarding layer, the current OpenFlow-based switch does not provide a passive push-based method based on the measurement data. Therefore, a Link State Update Policy is integrated so that the link state information from the SDN switch is pushed to the SDN controller every time a certain criterion is met. As each link has its own bandwidth capacity, the SDN switch computes the bandwidth availability metric through the Metric Computation unit. The application layer provides two units, namely the proposed Probabilistic QoS Routing Algorithm, BaProbSDN and the Global Information State Management which maintains a consistent global network state information matrix.
Application Layer
BaProbSDN (Probabilistic QoS Routing)
Global Information State Management
Northbound Interface (API)
Network Services
SDN Controller
Control Layer
Network Operating System
Southbound Interface (OpenFlow API)
SDN Switch (1)
Forwarding Layer
Fig. 1.
Link State Update Policy
SDN Switch (1) Metric Computation
Link State Update Policy
Metric Computation
SDN architecture with proposed functional units
A. Link State Update Policy In practice, the state of a network changes over time according to several factors, such as: node or link failure, massive traffic overload on a certain link or group of links, etc. In the case of SDN architecture, the SDN controller maintains the global network state information which needs to be updated every time something changes in the network. However the update of each state change of all the SDN switches can lead to overhead and waste of resources. Therefore the link state update policy is used in the SDN switch to prevent such a network overhead. By integrating the update policy, the switch advertises its state information only when certain conditions are fulfilled. The Link State Update Policy consists of a threshold-based triggering policy combined with a holddown timer (HDT) [10]. Thus, the link state update is triggered only if a certain threshold thr was exceeded and the HDT expired. In this paper the used QoS metric is the available bandwidth so that the threshold-based policy will first check blast −bk if blast > thr. where blast is the last updated value of the available bandwidth metric and bk is the current state of the metric. B. BaProbSDN This paper proposes BaProbSDN, a probabilistic-based QoS routing algorithm that makes use of a directed acyclic graph Bayesian network and Bayes’ theorem to obtain the link probability of bandwidth availability. The probability is determined along a window of observation, referred to as Window Size (WS) (i.e., a window size gives the observation period that contains a number of measurement samples which are updated by the SDN switches). The following states are defined to model the bandwidth availability of a link as described by the Bayesian network listed in Fig. 2: •
L ≡ bk > breq : Does the link has enough bandwidth
•
∆B ≡ breq − blast : The amplitude of last advertised bandwidth value blast relative to the requested bandwidth breq
•
T ≡ t (blast ) = i : The temporal order of last advertised bandwidth t (blast ) where i is the location of bandwidth measurement in the observation window
Hence given the above mentioned states, T and ∆B are independent but they become dependent when L is given,
therefore the prior joint probability is defined as: P r [L, ∆B, T ] = P r [L|∆B, T ] P r [∆B] P r [T ]
T
(1)
Here in order to find the probability of a link, we are interested to compute the posterior probability P r [L|∆B, T ] which informs whether the link supports a bandwidth larger than the requested one when the last advertised measurement and the requested bandwidth are given. It is rather difficult to calculate directly the probability of hypothesis L given the evidence of ∆B and T , therefore Bayes rule simplifies the problem by considering the observables quantities as formulated in 2: P r [L|∆B, T ] = P r [∆B, T |L] P r [L] P r [L, ∆B, T ] = P r [∆B] P r [T ] P r [∆B] P r [T ]
B
(2)
The likelihood P r [∆B, T |L] is estimated from the system observations of a large training dataset in advance. Here the likelihood is calculated approximately by the means of a counting process. The proposed approach is based on Bayes’ theorem and the causal relation in Bayesian network. Given the location of last advertised bandwidth availability within the observation window and the value of current requested bandwidth, the model estimates the link probability based on this information. Basically, a link is more likely to support the request as the location of the last update approaches the request arrival time and the difference ∆B becomes larger. However, the model reaches moderate probability as the difference ∆B becomes smaller, this indicates that there is no evidence whether the link supports lower or higher bandwidth at the time of the arrival request. The probability-based routing algorithm is formulated as a path finder of p∗ that is most likely to satisfy the bandwidth constraint B, and the following condition is met for every other path p [5]: πB (p∗ ) ≥ πB (p) (3) Q where πB (p) = (i,j)∈p P r [b (i, j)] and b (i, j) is the residual bandwidth of the link (i, j). The solution of eq. 3 is carried out by assigning to each link a weight of −log (P r [b (i, j)]), so that the problem is transformed to additive operation in path finding. To find the path πB (p∗ ) the standard Dijkstra shortest path algorithm is executed on the associated weights with a computational complexity of O(L logN). This algorithm is known as the Most-Probable Bandwidth Constrained Path (MP-BCP) problem [5]. According to the observations, the model produces different probabilities depending on the order of the observation sequence. When a request arrives, the probability is determined by equation 2 for each link in network. Therefore the group of links that shape a QoS path between the source and destination hosts can be calculated by solving the MP-BCP problem. In this paper, the problem solution was modified so that the group of paths that most likely satisfy the QoS requirements are selected first. Then the path which contains the links with higher probabilities is chosen. Additionally for lonely threshold-based policy, the BaProbSDN algorithm can determine if the bandwidth availability that is not advertised by the switch supports enough bandwidth for the QoS request. This can be achieved by checking if the requested bandwidth lies in the range of breq ∈ / [bk−1 (1 + thr) , bk−1 (1 − thr)].
L
Fig. 2.
Bayesian network model in BaProbSDN for link probability
IV.
PERFORMANCE EVALUATION
The performance of the proposed BaProbSDN algorithm is evaluated through simulations using Matlab and it is compared against the widest-shortest path routing (WSR) [11]. The WSR algorithm selects first the shortest path group that matches the bandwidth availability constraint and then it picks the widest path among the match group (i.e., the path which offers the larger bandwidth availability). For both algorithms, we introduce a pruning method by ignoring the links if their last observed bandwidth state is smaller than the requested amount. A. Network Simulation The Internet Service Provider (ISP) topology as illustrated in Figure 3 was chosen for performance evaluation. A number of several hosts could be connected to the SDN switches generating data traffic. For the considered topology, the number of switches (N) is 18 and the number of links inter-connecting the switches (L) is 30 while the link capacity is set to 100Mbps. The network simulation employs a traffic generation model that loads the network using two kinds of request arrivals: one that arrives at each host to generate the best-effort traffic while the other arrives at the source host for QoS routing execution. The traffic arrival follows a Poisson distribution with rate λ while the active period of each connection is distributed exponentially with a mean of 1/µ seconds [12]. The destination node is chosen at random other than the source node. For the traffic load, the request arrival rate λ is chosen to determine the overall load. In the simulation, two traffic patterns are used to load the network, the first behaves for long duration with a mean of µ = 70 seconds and small bandwidth requirements distributed uniformly between 1Mb and 10Mb, while the second indicates a short duration with a mean of µ = 10 seconds and larger bandwidth requirements of uniform distribution between 20Mb and 30Mb. The two patterns were shared equally in the simulation. The links are assumed to be bi-directional with the same bandwidth capacity C. For the QoS routing traffic, the source and destination pairs are fixed during the simulation period and their requested bandwidth is uniformly distributed between 10Mb and 40Mb. Within the SDN network, every switch samples the link metric at every 1 second interval and it sends the link state information to the controller based on the link state update policy decision as previously described. The used performance metric is the bandwidth blocking rate which reflects the success of the algorithm to set an endto-end QoS path according to the bandwidth metric. A rejected request indicates that the end-to-end QoS path contains at least one link that does not satisfy the bandwidth constraint. Therefore by measuring the rate, it shows if the algorithm generates an end-to-end QoS path P with the following rule min(i,j)∈P ⌊b (i, j)⌋ > breq . The bandwidth blocking rate is
Application Layer
BaProbSDN (Probabilistic QoS Routing)
Global Information State Management
100
Update Rate of Information State (%)
Northbound Interface (API)
Network Services
Control Layer
SDN Controller Network Operating System Southbound Interface (OpenFlow API)
Forwarding Layer
Source Host
Destination Host
HDT = 0s HDT = 10s HDT = 20s HDT = 40s
90 80 70 60 50 40 30 20 10
Fig. 3.
Considered ISP topology
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Threshold Value
defined as in eq. 4 [13].
(4)
B. Results and Discussion This section presents the performance comparison of two algorithms under various settings of the link state update policy. The results were averaged over 50 simulation runs and they were obtained with a 95% confidence interval. As SDN separates the control plane from the data plane, the control and information messages are exchanged frequently between the switches and the controller for reasons like checking of resource availability or capacity planning. In particular, for routing purposes, a significant amount of overhead is often introduced to keep the network state at the controller, as accurate as possible especially when the network size becomes larger. Therefore it is essential to quantify the algorithm performance and the overhead introduced due to frequent update messages. The aim is to minimize the overhead associated with the routing in the SDN network while ensuring that the QoS guarantees are satisfied. To this extent, the thresholdbased triggering policy together with the HDT are introduced. Figure 4 illustrates the state information update rate under various threshold and HDT values. It can be noticed that the introduction of threshold-based link update policy reduces the number of advertised states in the network. For example, for the threshold value of 0.5 and HDT=0s there is 84.05% decrease in the state information update rate. It can be seen that the variation of update percentage decreases with the increasing threshold and HDT values. For a threshold value of 0.3 and HDT=20s the update rate reaches 3.78%. As previously mentioned, the BaProbSDN method computes the link probability on a set of observations taken over a WS. Therefore, Fig. 5 shows how the WS impacts the performance of BaProbSDN method. The result shows that the blocking rate becomes steady as WS increases. For example, for HDT = 10s and threshold value of 0.3 BaProbSDN achieves 26.90% in the bandwidth blocking rate when WS = 100. Thus for the rest of simulation, the WS is set to 100 seconds. Figure 6 illustrates a comparison of the algorithms under varying threshold values. It can be seen that BaProbSDN reduces the bandwidth blocking rate when compared to the WSR algorithm. For example, for Thr=0.6, BaProbSDN can achieve
100 HDT = 10s HDT = 20s HDT = 30s HDT = 40s
90 80 Bandwidth Blocking Rate (%)
bandwidth blocking rate = P Bandwidth of rejected requests P Bandwidth of arrived requests
Fig. 4. State information update rate for different HDT and threshold-based link update policy
70 60 50 40 30 20 10 0
20
40
60 80 100 Window Size (sec)
120
140
Fig. 5. Bandwidth blocking rate of BaProbSDN method under different WS (Threshold=0.3)
up to 5.50% decrease in the bandwidth blocking rate when compared to WSR. Moreover, WSR has a close performance to the BaProbSDN method because the group of links that support enough bandwidth depend on the requested bandwidth lying in the range of breq ∈ / [bk−1 (1 + thr) , bk−1 (1 − thr)]. WSR will present an improved performance in the presence of imprecision only if those links that contribute to the QoS path do not lie in this range. Figure 7, presents the results in terms of bandwidth blocking rate, obtained when HDT value was varied between from 5 to 40 seconds. The results show that as HDT becomes larger the performance of the two algorithms decreases noticeably. It can be seen that BaProbSDN outperforms the WSR algorithms. For example, for HDT=15s and Thr=0.5, BaProbSDN can achieve up to 9.43% decrease in the bandwidth blocking rate when compared to WSR. In order to study the impact of the traffic load, the traffic arrival rate λ was varied while other simulation parameters were kept fixed. Figure 8 shows the performance of the two algorithms as a function of the network traffic load. Due to the aggregated cross traffic the link will be more exploited as the traffic load increases. It can be seen that under highly loaded network, the blocking rate of algorithms increases. However, when compared to the WSR, BaProbSDN achieves
100 BaProbSDN WSR
90
Bandwidth Blocking Rate (%)
80 70 60 50 40 30 20 10 0
Fig. 6.
0
0.1
0.2
0.3
0.4 0.5 Threshold Value
0.6
0.7
0.8
0.9
Bandwidth blocking rate for different threshold values and HDT=0s
100 Thr = 0.5 Thr = 0.8
90
Bandwidth Blocking Rate (%)
WSR
60 50 40 30
BaProbSDN
20 10 0
Fig. 7.
5
10
15
20 25 Hold Down Time (sec)
30
35
40
Bandwidth blocking rate for different HDT values
100 BaProbSDN WSR
90
Bandwidth Blocking Rate (%)
80 70 60 50 40 30 20 10 0 80
90
100
110 120 Traffic Load (Mbits/s)
130
140
Fig. 8. Bandwidth blocking rate for different traffic loads (Threshold=0.3, HDT=10s)
up to 7.96% decrease in the bandwidth blocking rate for a traffic load of 140Mbps. V.
As future work, we plan to extend the model by considering other QoS metrics (e.g., packet loss, delay) and by integrating an adaptive monitoring scheme for policy violation detection. Furthermore, we plan to evaluate the performance of the proposed solution through real experimental scenarios. R EFERENCES
80 70
order to decrease the overhead on the control plan between the SDN switches and the controller, BaProbSDN makes use of a threshold-based triggering link update policy combined with a hold-down timer. The results show that the overhead can be greatly reduced with less significant impact on the performance in terms of bandwidth blocking rate. Moreover the proposed BaProbSDN algorithm was compared against the widest-shortest routing (WSR) algorithm in the presence of link state update policy. The results show that BaProbSDN can achieve up to 7.41% decrease in the bandwidth blocking rate when compared to WSR, for a threshold value of 0.5, HDT=10s, and a reduction of 95.38% in the control messages overhead.
CONCLUSIONS
This paper proposes, BaProbSDN a probabilistic-based QoS routing algorithm for Software Defined Networks. BaProbSDN makes use of a Bayes’ thoerem to determine the link probability in terms of bandwidth availability. In
[1] Software-Defined Networking: The New Norm for Networks. Accessed on September, 2014. [Online]. Available: https://www.opennetworking.org [2] H. Egilmez, S. Dane, K. Bagci, and A. Tekalp, “OpenQoS: An OpenFlow controller design for multimedia delivery with end-to-end Quality of Service over Software-Defined Networks,” Signal Information Processing Association Annual Summit and Conference (APSIPA ASC), 2012 Asia-Pacific, pp. 1–8, Dec 2012. [3] H. Egilmez, B. Gorkemli, A. Tekalp, and S. Civanlar, “Scalable video streaming over OpenFlow networks: An optimization framework for QoS routing,” Image Processing (ICIP), 2011 18th IEEE International Conference on, pp. 2241–2244, Sept 2011. [4] R. Wallner and R. Cannistra, “An SDN Approach: Quality of Service using Big Switchs Floodlight Open-source Controller,” Proceedings of the Asia-Pacific Advanced Network 2013, vol. 35, pp. 14–19, 2013. [5] P. Van Mieghem, F. A. Kuipers, T. Korkmaz, M. Krunz, M. Curado, E. Monteiro, X. Masip-Bruin, J. Sol-Pareta, and S. Snchez-Lpez, “Quality of Service Routing,” Quality of Future Internet Services, Lecture Notes in Computer Science Springer, vol. 2856, pp. 80–117, 2003. [6] T. Korkmaz and M. Krunz, “Bandwidth-delay constrained path selection under inaccurate state information,” IEEE/ACM Trans. Netw., vol. 11, pp. 384–398, 2003. [7] D. Ghosh and R. Acharya, “A probabilistic scheme for hierarchical QoS routing,” Networks, 2001. Proceedings. Ninth IEEE International Conference on, pp. 416–421, Oct 2001. [8] D. Lorenz and A. Orda, “QoS routing in networks with uncertain parameters,” Networking, IEEE/ACM Transactions on, vol. 6, no. 6, pp. 768–778, Dec 1998. [9] X. Yuan and G. Yang, “Empirical probability based QoS routing,” Communications, 2003. ICC ’03. IEEE International Conference on, vol. 3, pp. 1713–1717, May 2003. [10] G. Apostolopoulos, R. Gu´erin, S. Kamat, and S. K. Tripathi, “Quality of service based routing: A performance perspective,” ACM SIGCOMM Computer Communication Review, vol. 28, no. 4, pp. 17–28, 1998. [11] D. Medhi and K. Ramasamy, Network routing: algorithms, protocols, and architectures. Morgan Kaufmann, 2010. [12] A. Shaikh, J. Rexford, and K. G. Shin, “Evaluating the Impact of Stale Link State on Quality-of-Service Routing,” IEEE/ACM Transactions on Networking, vol. 9, pp. 162–176, 2001. [13] K. Kowalik and M. Collier, “ALCFRA - A robust routing algorithm which can tolerate imprecise network state information,” in Proceedings of 15th ITC Specialist Seminar, Wurzburg, Germany, July 2002.
