Abstract
Network operators generally provide dedicated lightpaths for customers to meet the demand for high-quality transmission. Considering the variation of traffic load, customers usually rent peak bandwidth that exceeds the practical average traffic requirement. In this case, bandwidth provisioning is unmetered and customers have to pay according to peak bandwidth. Supposing that network operators could keep track of traffic load and allocate bandwidth dynamically, bandwidth can be provided as a metered service and customers would pay for the bandwidth that they actually use. To achieve cost-effective bandwidth provisioning, this paper proposes an autonomic bandwidth adjustment scheme based on data analysis of traffic load. The scheme is implemented in a software defined networking (SDN) controller and is demonstrated in the field trial of multi-vendor optical transport networks. The field trial shows that the proposed scheme can track traffic load and realize autonomic bandwidth adjustment. In addition, a simulation experiment is conducted to evaluate the performance of the proposed scheme. We also investigate the impact of different parameters on autonomic bandwidth adjustment. Simulation results show that the step size and adjustment period have significant influences on bandwidth savings and packet loss. A small value of step size and adjustment period can bring more benefits by tracking traffic variation with high accuracy. For network operators, the scheme can serve as technical support of realizing bandwidth as metered service in the future.
© 2017 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
In order to meet the demand of high-quality transmission, most of customers tend to rent a dedicated line from network operators. The ability of quickly providing end-to-end services is one of the most important requirements of network operators. Considering the variation of traffic load, customers usually rent peak bandwidth that exceeds the practical average traffic requirement. The peak period can last from tens of minutes to several hours or even longer. However, it is wasteful and prohibitively expensive to lease a private line to accommodate such services with the peak bandwidth. An optimal solution is that network operators provide customers with bandwidth as metered service [1]. An illustration of bandwidth as metered service is shown in Fig. 1. Compared to conventional bandwidth provisioning, bandwidth as metered service can effectively reduce the time-bandwidth product. It means that customers have access to potentially large bandwidth and only pay for the capacity they actually use. And the technical support of bandwidth as metered service is to realize dynamic bandwidth adjustment based on traffic load. In terms of providing cost-effective bandwidth service and improving resource utilization, it is highly desired for network operators to achieve dynamic bandwidth adjustment.
A lot of research has been conducted on bandwidth adjustment. An energy-efficient OFDM (Orthogonal Frequency Domain Multiplexing) transceiver system was proposed with power consumption following dynamically network traffic load [2]. In the solution, the bandwidth can be adjusted by adaptively tuning signal bandwidth, sampling rate, and number of parallel processing modules. However, the practical solution was only suitable for digital transceiver with OFDM modulation. The benefits of traffic engineering (TE) have been proved in terms of better utilization of network resources [3,4]. The objective of TE is to put the data traffic where the network bandwidth is available in an efficient way [5]. In TE, the allocation of network resources is based on current network status, which is a similarly static way. It means that the allocated resources will remain for some time. Considering the traffic variation, it is very likely that the original resource allocation would not be the optimal solution in the future time. Thus, the benefit of TE is limited in the scenarios of traffic variation.
As a centralized software control architecture, software defined networking (SDN) can realize the unified control over optical network resources [6–8]. The emergence of SDN technology provides an effective solution to implement the strategies of bandwidth adjustment. Based on the hierarchical control architecture in [9], the bandwidth on demand (BoD) service was deployed in the field trial of multi-vendor optical networks [10]. The strategies of bandwidth adjustment can be classified into two types, i.e., instant adjustment and timing adjustment. However, instant adjustment needs the manual operation to trigger the bandwidth adjustment while timing adjustment cannot match the dynamic change of traffic load. Meanwhile, the dynamic bandwidth adjustment requires the traffic prediction based on current traffic load. Big data analytics was applied for periodical IP traffic prediction to automate virtual network topology (VNT) reconfiguration [11–13]. The authors in [14] formally stated the incremental capacity planning (INCA) problem, and proposed an integer linear program (ILP) formulation and heuristic algorithm to support on-demand network planning in optical transport networks. Traffic prediction based on metro-flow model aggregation was proposed in [15] to reconfigure the VNT after metro-flow re-routing. However, to the best of our knowledge, there is no available solution to realize dynamic bandwidth adjustment based on real-time traffic load in the context of software defined optical networks.
To address the problem, this paper proposes a data analysis-based autonomic bandwidth adjustment (DA-ABA) scheme, which can dynamically track traffic load. In addition, the scheme was implemented as a function module in SDN controller based on OpenDaylight [16]. Field trial has been conducted in the software defined multi-vendor optical transport networks to demonstrate the feasibility and effectiveness of the scheme. Trial results show that the proposed scheme can automatically adjust bandwidth based on data analysis of real-time traffic load. Meanwhile, simulation experiments are conducted to evaluate its performance in terms of three metrics including bandwidth savings, the number of traffic overflow and average packet loss rate. We also investigate the impact of different parameters on the bandwidth adjustment in DA-ABA scheme.
2. Autonomic bandwidth adjustment scheme based on data analysis
The DA-ABA scheme runs periodically and tracks the variation of traffic load. The trend of traffic variation for the next period is predicted by analyzing the traffic monitoring data of current period. Based on the predicted traffic load, the DA-ABA scheme adjusts the bandwidth for the next period. For the sake of clarity, the following notations are introduced to describe the DA-ABA scheme (Table 1).
The flow chart of DA-ABA scheme in an adjustment period is illustrated in Fig. 2. In DA-ABA scheme, a series of parameters can be configured according to different traffic profiles. Once the DA-ABA scheme starts, traffic monitoring data is collected by sampling the traffic and the sampling interval is configurable. For each sampling point,, and are calculated based on Eqs. (1-3). If, it means that the traffic load of the sampling point has reached the upper threshold of current bandwidth. Similarly, denotes that the monitored traffic load has reached the lower threshold of current bandwidth.
After all the sampling points are collected in an adjustment period, we perform the data analysis of traffic load in current period and predict the traffic trend further for the next period. If at leastsampling points have reached the upper threshold of current bandwidth, the traffic capacity is more likely to increase in the next period. Thus, the DA-ABA scheme makes decisions to increase the bandwidth and the step number is calculated with formula (4) and (5). Formula (4) assures that after bandwidth adjustment, average traffic capacity in current period does not exceed the upper threshold of allocated bandwidth in next period. Formula (5) refers to the constraint of maximum bandwidth allocated for service connection. If does not exist, the bandwidth is set to the maximum bandwidth in the next period.
In the similar way, if at leastsampling points have reached the lower threshold of current bandwidth, the bandwidth is supposed to be decreased for the next period. The step number of decreasing bandwidth is calculated according to formula (6) and (7). Formula (6) guarantees that after bandwidth reduction, average traffic capacity in current period is higher than the low threshold of allocated bandwidth in next period. Formula (7) represents the constraint of minimum bandwidth provided for service connection. If is not found, the bandwidth is adjusted to the minimum bandwidth in the next period. Note that if and, the bandwidth will remain the same as current period.
We can analyze the time complexity of the proposed DA-ABA scheme. In single adjustment period, the DA-ABA scheme traverses all the sampling points and predicts the traffic trend based on current traffic load. Therefore, the time complexity of DA-ABA scheme in each period is, where is the length of traffic sampling in each adjustment period.
3. Field trial setup with multi-vendor equipment
Figure 3 shows the field trial setup of multi-domain optical transport networks, which is conducted in Fujian Province of China. The trial network is composed of three domains provided by three different vendors, HUAWEI, ZTE and FiberHome. As shown in Fig. 3, the trial networks include three planes: data plane, control plane and application plane.
In the data plane, 10Gbit/s inter-domain links are provided by 40Gbit/s inter-city optical transport networks. Intra-domain is equipped with 10Gbps links and supports GE and 10 GE client interfaces. Note that the capacity of inter-domain and intra-domain links does not affect the demonstration of bandwidth adjustment. With each domain providing three OTN nodes, there are nine nodes totally in the trial networks. The board card in OTN nodes can collect the monitored information, including service throughput, service delay and packet loss rate. Note that the traffic in the field trial was generated by JDSU Ethernet tester. We adopted JDSU MTS-6000 and JDSU ONT-503 for the GE and 10 GE interface, respectively.
In the control plane, hierarchical control architecture is adopted to realize the unified control of multi-domain optical transport networks. Each domain has a vendor-specified single-domain controller and its unique control protocol. Multi-domain controller is responsible for communicating with single-domain controllers of different vendors. The multi-domain controller is developed based on OpenDayLight controller, in which the DA-ABA scheme is implemented as a function module. OpenFlow protocol is extended to support Control Virtual Network Interface (CVNI) [9] between single-domain controllers and multi-domain controller.
In the application plane, the DA-ABA application is designed for network operators and it is connected to multi-domain controller via RESTful API [17]. After operators login the DA-ABA application software, they can perform the operation about DA-ABA scheme. On the one hand, the parameters of DA-ABA scheme can be configured according to customer requirement. Besides, after DA-ABA scheme starts, network operators have access to the record of bandwidth adjustment by selecting the service ID and specifying a certain time span.
Meanwhile, each network node has a corresponding IP address and the subnet mask is set as 255.255.0.0. All the single-domain controllers, the multi-domain controller and the APPs are connected to an Ethernet hub with China Telecom VPN. Note that bandwidth adjustment can be realized by configuring either client-side port (Ethernet) or line-side port (ODUflex) of OTN nodes. In the field trial, the client-side adjustment is adopted for gigabit Ethernet (GE) service.
4. Field trial results
The field trial results are shown in Fig. 4 and Fig. 5. In Fig. 4(a), we can see that a connection from node 10.130.24.3 to node 10.130.24.1 has been established successfully and its service name is admin-1. The parameters of DA-ABA scheme have been configured, including adjustment period, sampling length, maximum bandwidth. For instance, the maximum and minimum bandwidth assigned for this service are 800 Mbps and 300 Mbps, respectively. The IP addresses of DA-ABA application software and multi-domain controller are 10.130.5.1 and 10.130.5.11, respectively. Wireshark capture in Fig. 4(c) illustrates that configuration message is sent to multi-domain controller with PUT method of HTTP protocol [18] and its corresponding JavaScript Object Notation (JSON) [19] is illustrated in Fig. 4(b).
In Fig. 5(a), the adjustment information of service admin-1 is obtained and the time span is from 17:50:42 to 17:55:42. Since sampling interval is one minute, there are five sampling points in this time span. For instance, we can see that the real throughput of sampling point at 17:54:42 is 504Mbps. Meanwhile, as shown in Fig. 5(b), the JSON object of this sampling point includes sampling time, throughput and allocated bandwidth. Figure 5(c) describes the exchange messages for inquiring the record of bandwidth adjustment. The message is sent to multi-domain controller with GET method of HTTP protocol [20].
Network analyzer is used to monitor the bandwidth adjustment. As shown in Fig. 6, it is interesting to notice that the monitored bandwidth increases from 500Mbps to 600Mbps at 17:56:41. This can be explained as follows. Based on data analysis of traffic sampling information, there is high possibility that the traffic load will exceed current assigned bandwidth in the next period. Thus, the bandwidth of service admin-1 is adjusted from 500 to 600Mbps, indicating that the DA-ABA scheme is carried out successfully in the field trial.
5. Performance evaluation
Besides the field trial, simulation is conducted to evaluate the performance of the proposed DA-ABA scheme. We refer to the traffic profile, which is obtained by averaging the data collected by several European network operators [21]. Depicting the average daily traffic variation, this traffic profile serves as the input traffic of end-to-end lightpath connection in the simulation. In each experiment, we establish the single private line service. Note that the traffic profile is normalized to the peak value and the initially bandwidth allocated for the private line is the peak value (i.e. 1.0). In the simulation, sampling interval is set to 0.5 hour and the adjustment period is assumed to be an integer multiple of sampling interval. Thus the minimum adjustment period is 0.5 hour. We investigate the impact of different parameters on the bandwidth adjustment in DA-ABA scheme. In the simulation, we focus on four parameters including,,and .
The traditional private line is provided with peak bandwidth in the whole day. The proposed scheme dynamically allocates bandwidth based on the analysis of traffic load. However, due to the traffic burst, the allocated bandwidth cannot always meet the capacity demand, which leads to traffic overflow and packet loss in some cases. To quantify the benefits and weakness of the proposed scheme, the performance is evaluated in terms of three metrics, i.e. bandwidth savings, the number of traffic overflow and the average packet loss rate within 24 hours. We calculate the bandwidth saved by DA-ABA scheme in 24 hours according to Eq. (8). In Eq. (8), represents the allocated bandwidth at time and is a boolean variable. If the allocated bandwidth can meet the capacity demand at time, the value of is 1. Otherwise,is set to 0. It is important to note that the increment of is 0.5 since the minimum time granularity of bandwidth allocation is 0.5 hour.
In Fig. 7, the adjustment period of DA-ABA scheme is set to four different values (i.e. 0.5, 1.0, 1.5 and 2). The value of and is 0.8 and 0.6, respectively. For all the four cases, bothand are set to 1. The step size of bandwidth adjustment is 0.1. It is obvious that the DA-ABA scheme dynamically adjusts the bandwidth based on the variation of input traffic. However, the adjustment period has a great influence on the performance. The higher value of corresponds to less adjustment times in the same time span, which makes it difficult to track the traffic variation. With, the allocated bandwidth in the whole day fits well with dynamic traffic variation. From the performance comparison in Fig. 7, we can see that the case achieves the best performance in terms of bandwidth savings and packet loss. It is also the only case that does not suffer from traffic overflow and packet loss. With the growth of adjustment period from 1 hour to 2 hours, the number of traffic overflow increases from 1 to 4 and the average packet loss gets worse from 13.8% to 25.7%. Therefore, a smaller value of adjustment period is preferable to maximize the benefits of DA-ABA scheme.
We also consider the impact of step size of bandwidth adjustment in DA-ABA scheme. In Fig. 8, the step size of bandwidth adjustment is set to four different values (i.e. 0.05, 0.1, 0.15 and 0.2). In four cases, the adjustment period is equal to 0.5 hour, and the value of and is 0.8 and 0.6, respectively. If, the minimum adjustment to increase or decrease the bandwidth is 0.05 since the value is an integral multiple of. It is interesting to observe that the step size plays a significant role in DA-ABA scheme. With a smaller value of, the DA-ABA scheme is more sensitive to the traffic variation. In Fig. 8, we can see that the traffic between 5:30 and 8:00 is relatively flat. In this period, the bandwidth adjustment fits well with traffic trend for. However, for the case, the allocated bandwidth fluctuates greatly between 0.2 and 0.4, which does not follow the traffic variation. From the table, we can also find that the smallerbrings more bandwidth savings within 24 hours. In addition, the case even causes traffic overflow and the average packet loss rate is up to 20%. By contrast, there is no traffic overflow and packet loss for the other three cases.
Figure 9 shows how threshold number of sampling points affects the performance of DA-ABA scheme. The adjustment period is set to 1.5 hours and sampling length is 3. The value of and is 0.8 and 0.6, respectively. The step size of bandwidth adjustment is 0.1. We consider three different values of threshold numberin the simulation (i.e. N = 1, 2, 3). If, bandwidth adjustment will be triggered as long as one sampling point reaches theorof current bandwidth. However, for, the allocated bandwidth will be adjusted only if all the sampling points reach theor. Therefore, with higher value of, the requirement is stricter and higher to trigger the bandwidth adjustment. For example, the bandwidth from 8:30 to 10:00 is allocated based on the analysis of traffic load from 7:00 to 8:30. If or 2, the bandwidth increases from 0.2 to 0.3 to meet the demand of traffic growth. However, the bandwidth remains at 0.2 for since only two sampling points reach the of current bandwidth. The performance comparison of three cases is also shown in Fig. 9. We can find that the performance is almost the same for and. The results get worse for with more traffic overflow and higher packet loss rate.
The impact of and on the performance of DA-ABA scheme is shown in Fig. 10. The adjustment period is 0.5 hour and sampling length is set to 1. In addition, threshold number of sampling points is also equal to 1. The step size of bandwidth adjustment is 0.05. In DA-ABA scheme, the parametersand determine that how much bandwidth should be increased or reduced for the next period. Three different pairs are discussed in this simulation. Comparing the case with, we can note that the higher value of leads to the bigger drop of allocated bandwidth when traffic rate decreases from the 3rd hour to the 7th hour. Similarly, comparing the case with, we can observe that the lower value of corresponds to the higher increase of allocated bandwidth when traffic rate increases. Thus, the lower value of fits well with the sharp growth of traffic load while the higher value of is suitable for rapid decline of traffic load. The performance metrics of all the three cases are shown in the table. Even though the casecan bring the most bandwidth savings, the traffic exceeds the allocated bandwidth twice within 24 hours and the average packet loss rate is 6.7%. By contrast, there is no traffic overflow for the other two cases.
In Fig. 7-10, it is interesting to find that although the traffic rate decreases in the beginning hours, the bandwidth remains at the peak value. This is because the traffic rate does not reach the lower threshold and the bandwidth adjustment is not triggered. By analyzing the results in Fig. 7-10, we can conclude that all the four parameters will affect the performance of DA-ABA scheme. For the same input traffic, different parameter settings determine the value of bandwidth savings and packet loss. Although the DA-ABA scheme could lead to packet loss in some cases, we can reduce or even avoid packet loss by setting appropriate parameters. Especially, with a smaller value ofand T, the proposed DA-ABA scheme can achieve more bandwidth savings and reduce the impact of packet loss.
6. Conclusions
In conclusion, this paper introduces an autonomic bandwidth adjustment scheme based on data analysis of traffic load in terms of providing cost-effective bandwidth service. The DA-ABA scheme runs periodically and tracks the variation of traffic load. The trend of traffic variation for the next period is predicted by analyzing the traffic monitoring data of current period. Based on the predicted traffic load, the DA-ABA scheme adjusts the bandwidth for the next period. The scheme is implemented in OpenDaylight controller and verified in field trial network with commercial OTN equipment from three vendors. In addition, simulation experiments are carried out to evaluate the impact of different parameters in DA-ABA scheme. Simulation results show that more bandwidth savings and lower packet loss rate can be achieved with a smaller value of step size and adjustment period. The proposed scheme provides a promising solution for network operators to achieve bandwidth as metered service in the future.
Funding
National Science and Technology Major Project (2017ZX03001016); National Natural Science Foundation of China (NSFC) (61571058, 61601052).
Acknowledgments
Part of this work has appeared in the proceeding of Optical Fiber Communications (OFC) Conference, Los Angeles, USA, in March 2017 [22].
References and links
1. C. Yeo, S. Venugopal, X. Chua, and R. Buyya, “Autonomic metered pricing for a utility computing service,” Future Gener. Comput. Syst. 26(8), 1368–1380 (2010).
2. J. Hu, D. Qian, and T. Wang, “Energy Efficient OFDM Transceiver based on Traffic Tracking and Adaptive Bandwidth Adjustment”, in Proceedings of European Conference on Optical Communication (Optical Society of America, 2011), paper We.10.P1.53.
3. D. Awduche and Y. Rekhter, “Multiprotocol lambda switching: combining MPLS traffic engineering control with optical crossconnects,” IEEE Commun. Mag. 39(3), 111–116 (2001).
4. K. Zhu, H. Zhu, and B. Mukherjee, “Traffic engineering in multigranularity heterogeneous optical WDM mesh networks through dynamic traffic grooming,” IEEE Netw. 17(2), 8–15 (2003).
5. Y. Lee and B. Mukherjee, “Traffic engineering in next-generation optical networks,” IEEE Commun. Surveys Tuts. 6(3), 16–33 (2004).
6. Y. Yu, J. Zhang, Y. Zhao, Y. Lin, J. Han, H. Zheng, Y. Cui, M. Xiao, H. Li, Y. Peng, Y. Ji, and H. Yang, “Field demonstration of multi-domain software-defined transport networking with multi-controller collaboration for data center interconnection,” J. Opt. Commun. Netw. 7(2), 301–308 (2015).
7. H. Chen, J. Zhang, Y. Zhao, J. Deng, W. Wang, R. He, X. Yu, Y. Ji, H. Zheng, Y. Lin, and H. Yang, “Experimental demonstration of datacenter resources integrated provisioning over multi-domain software defined optical networks,” J. Lightwave Technol. 33(8), 1515–1521 (2015).
8. J. Zhang, Y. Zhao, H. Yang, Y. Ji, H. Li, Y. Lin, G. Li, J. Han, Y. Lee, and T. Ma, “First demonstration of enhanced software defined networking (eSDN) over elastic grid (eGrid) optical networks for data center service migration,” in Proceedings of Optical Fiber Communication Conference (Optical Society of America, 2013), paper PDP5B.1.
9. R. Jing, C. Zhang, Y. Ma, J. Li, X. Huo, Y. Zhao, J. Han, J. Wang, and S. Fu, “Experimental demonstration of hierarchical control over multi-domain OTN networks based on extended OpenFlow protocol,” in Proceedings of Optical Fiber Communication Conference (Optical Society of America, 2015), paper W4J.4.
10. C. Zhang, R. Jing, J. Li, Y. Ma, X. Huo, Y. Zhao, and J. Zhang, “Field trial of bandwidth on demand services based on hierarchical control over multi-domain OTN networks,” J. Opt. Commun. Netw. 7(11), 1057–1063 (2015).
11. F. Morales, M. Ruiz, and L. Velasco, “Virtual network topology reconfiguration based on big data analytics for traffic prediction”, in Proceedings of Optical Fiber Communication Conference (Optical Society of America, 2016), paper Th3I.5.
12. L. Gifre, L. M. Contreras, V. López, and L. Velasco, “Big data analytics in support of virtual network topology adaptability”, in Proceedings of Optical Fiber Communication Conference (Optical Society of America, 2016), paper W3F.6.
13. F. Morales, M. Ruiz, L. Gifre, L. M. Contreras, V. Lopez, and L. Velasco, “Virtual network topology adaptability based on data analytics for traffic prediction,” J. Opt. Commun. Netw. 9(1), 35–45 (2017).
14. L. Velasco, F. Morales, L. Gifre, A. Castro, O. Gonzalez de Dios, and M. Ruiz, “On-demand incremental capacity planning in optical transport networks,” J. Opt. Commun. Netw. 8(1), 11–22 (2016).
15. F. Morales, M. Ruiz, and L. Velasco, “Core VNT adaptation based on the aggregated metro-flow traffic model prediction,” in Proceedings of Optical Fiber Communication Conference (Optical Society of America, 2017), paper M2G.5.
16. OpenDaylight [Available Online]: https://www.opendaylight.org/
17. Wikipedia, RESTful API,https://en.wikipedia.org/wiki/Representational_state_transfer
18. Wikipedia, POST, https://en.wikipedia.org/wiki/POST_(HTTP).
19. JSON (JavaScript Object Notation), http://www.json.org/.
20. Wikipedia, GET, https://en.wikipedia.org/wiki/GET_(HTTP).
21. “Energy efficiency analysis of the reference systems, areas of improvements and target breakdown,” EARTH Deliverable D2.3, 2012.
22. Y. Li, Y. Zhao, X. Yu, H. Chen, R. Jing, C. Yu, and R. Cui, “Field trial of data analysis-based autonomic bandwidth adjustment in software defined multi-vendor OTN networks”, in Proceedings of Optical Fiber Communication Conference (Optical Society of America, 2017), paper W1D.4.