the two cases when the signaling intervals are 260 ms, 250 ms, and 240 ms, respectively.
In Fig. 3.8a, we can find that the signaling overheads in our proposal are much lower than those in the case with OSPF. In Figs. 3.8b and 3.8c, we can clearly see the effects of signaling overheads on the performance of the two cases. In Fig. 3.8b, the throughput of our proposal remains nearly unchanged when the signaling interval is different. On the other hand, the throughput of OSPF, when the signaling interval is 240 ms, is much lower than that when the signaling interval is 260 ms or 250 ms. Thus, it may be inferred that the traffic congestion happens for the network using OSPF when the signaling interval is 240 ms. This is further demonstrated by the result in Fig. 3.8c which shows that the average delay per hop of OSPF, when the signaling interval is 240 ms, is nearly twice longer than that when the signaling interval is 260 ms or 250 ms. Moreover, we can find that when the signaling interval is 260 ms or 250 ms, the average delay per hop of OSPF is nearly the same as that of our proposal.
Through comparing the performance in the network using OSPF and our proposed routing strategy based on deep learning, we can find that our proposed deep learning based routing strategy has much lower signaling overhead, leading to better traffic control. The reason for the lower signaling overhead in our proposal is that only the edge routers instead of all routers require signaling packets since the edge routers can use the trained DBAs to build the whole paths and the inner routers do not need the signaling packets to compute the next nodes. However, in the network with OSPF, the edge routers cannot utilize current weights’ values of all links to build the practical whole paths as the paths computed through OSPF are only suitable for current network states. But during the packets’ transmission, the network traffic is changing and then the decided paths become unsuitable. On the other hand, for the routing strategy based on deep learning, the DBAs can find the complex relationship between the current traffic patterns and the real paths if we utilize the traffic patterns and real paths to train them. Therefore, the edge routers can utilize the trained DBAs to build the whole paths with only current network information.
based routing strategy outperforms the conventional OSPF in terms of the network packet transmission throughput and average delay per hop since our proposal has much lower signaling overhead. This demonstrated that the shift of routing computation from the traditional rule-based strategy to deep learning can improve the backbone network control substantially. In addition, the complexity of our proposed routing strategy was analyzed to evaluate that the GPU-accelerated SDR is much more efficient to run the proposed algorithms than the CPU-based SDR.
Online Learning Based Routing Strategy for Software Defined Communication Systems
4.1 Introduction
In last chapter, the supervised learning based routing strategy is proposed to tackle the increasing traffic overhead in the backbone networks. Since the labeled data impact on the performance of the considered deep learning architectures, it is a critical step to collect the training data. However, in many scenarios, it is very difficult to collect enough satisfying labeled data, for which the heterogeneous network is a good example.
As we know, for some heterogeneous networks, various kinds of communication tech-nologies, e.g. FiWi, D2D, and 5G, are utilized to meet users’ requirements in different scenarios [78, 79, 80]. Since these networks have various infrastructures and topology, to solve the difficulty in managing all these networks, researchers considered the SDN technology [33, 81]. As we mentioned in Chapter 1, the structures of routers and switches in the SDN scenarios get significantly simplified and unified due to the separation of com-plicated network logic. Similar to the cloud-based computing applications, the controllers conduct all the computation tasks for the switches [13]. Therefore, the utilized controller in SDN is usually composed by various computation platforms. To fit for the new network scenario as well as the improved computation capacity in SDN, the network algorithms should have been updated or redesigned. However, the packet forwarding algorithms in current Software Defined Communication Systems (SDCSs) [82] still follow the conven-tional manner [28]. Since the paths are computed according to fixed rules, when similar traffic patterns happen, the controller chooses the same paths even the decision has been previously proved wrong, which leads to unnecessary network performance deterioration.
S1
S2
S3
S4
S5
S6
S7
S9 Destined for
S8
(S8, S5) (S5, channel 6)
Routing Channel
assignment
SDN central controller Applications
S8
wireless links Legend chosen paths
Figure 4.1: The considered structure of SDCS.
This means that the current routing protocols lack the intelligence to learn from previous experiences. On the other hand, if we consider the deep learning based routing strategy similar to that in Chapter 3, it is still difficult to obtain the satisfied performance. This is because the data planes in the SDCSs consist of various communication technologies, different from the backbone networks. Therefore, the traffic patterns in the data plane are more complex and varying fast. Also, the bursty traffic is very common in the data plane. Therefore, even the considered architecture is trained with massive data, it can-not predict the paths accurately since the network surrounding may have changed. To solve this problem, in this chapter, we propose an online learning based routing strategy which periodically trains the considered architecture with real-time traffic patterns. The proposal consists of two steps: the initial phase and running phase. In the initial phase, the controller runs the conventional routing protocol while the switches record the traffic trace, which is utilized by the controller to initialize the utilized CNNs. Then, in the running phase, the CNNs are adopted in the controller to choose paths. Furthermore, to adapt the trained CNNs to the changing traffic patterns as well as reduce the training computation overhead, in the running phase, the switches keep recording the traffic trace for periodically retraining the CNNs in the controller.
The remainder of the chapter is organized as follows. Sec. 4.2 describes the problems in the routing strategies for current SDCSs, and then discusses our proposal to overcome the problems. The detailed procedures of our proposal are presented in Sec. 4.3. Sec. 4.4 analyzes the time and space complexity for the controller and switches in our proposal. We evaluate the network performance of our proposal in Sec. 4.5. Finally, Sec. 4.6 summarizes this chapter.
0 20 40 60 80 100
0 200 400 600 800 1000
Time (s)
Packet Generation Rate (Mb/s)
(a) The traffic patterns of switchesS1,S2, andS3.
0 0.1 0.2 0.3 0.4 0.5 0.6
0 200 400 600 800 1000
Packet Loss Rate
Time (s)
(b) The network packet loss rate.
0 50 100 150 200 250 300 350
0 200 400 600 800 1000
Delay (ms)
Time (s) (c) The network delay.
Figure 4.2: An illustrative example: when switches S1, S2, and S3 choose S5 as the next node to destinationS8,S5 will be the bottleneck, which means that traffic congestion will easily happen to S5.