Research on Traffic Dispatching Scheme of Multi-service Data Center for 5G Smart Grid

Aiming at the flow scheduling problem of 5G power network multi-service data center, this paper designs and proposes a multi-service-oriented multi-factor flow scheduling scheme. Based on the shortest path algorithm, this paper improves and perfects the traffic scheduling strategy, and introduces a set of scoring mechanism. The number of hops, the number of large flows and the remaining bandwidth of each path are used as scoring indicators, and the path with the highest comprehensive score is selected. As the forwarding path of traffic. Based on the Mininet simulation platform, use Ryu to control to generate probability model traffic to simulate data center network traffic. The experimental results show that the scheme improves the network throughput and bandwidth utilization, and improves the performance of the data center network.


The Concept of SDN for 5G Smart Grid
Software-defined networking (SDN) is a transformative network architecture, which brings hope for breaking through the limitations of current network infrastructure. First, it breaks the vertical integration of traditional devices by decoupling the control plane and data plane of the network. Secondly, with the decoupling of the control plane and the data plane, the network switch becomes a mere forwarding device, and the control plane is implemented in the controller, which simplifies the implementation of policies and the reconfiguration of the network. Table 1 summarizes the related architecture and design features of the SDN controller and control platform, in order to provide a reference for the design of the 5G grid controller.

The Architecture of SDN for 5G Smart Grid
The topological structure of the data center network has two major types, namely the switch type topology and the server type topology. Fattree, VL2, c-Through and Helios are the classic solutions of the former, and the network connection and routing functions are completed by a large number of switches; the typical topology of the latter is DCell, FiConn, BCube And uFix, etc., only one switch is connected to a large number of servers to form a layer 0 network. In order to build a multi-path traffic scheduling environment in the data center, this paper chooses the classic Fattree topology. The FatTree topology is proposed by Fares et al. of the Massachusetts Institute of Technology (MIT) on the basis of improving the performance of the traditional tree structure. The entire topological network is divided into three levels: from top to bottom are the access layer (edge), the aggregation layer (aggregate) and the core layer (core). The aggregation layer switch and the access layer switch form a region (Pod). The switching equipment is all commercial switching equipment.

The Multi-factor Flow Control (MFFC) Strategy
This paper proposes a Multi-factor flow control (MFFC) strategy for the data center-oriented traffic dispatch of SDN in the smart grid. The core of the strategy is to find the best path for the data flow. The main steps are as follows: First, use the KSP (k takes 5) algorithm based on the number of hops to calculate 5 candidate paths, and consider the number of hops, bandwidth, and the number of large flows in the 5-hop path for path selection. After that, the controller starts a statistics thread for sending port information requests and flow table information requests regularly, and processing the port information responses and flow table information responses returned by the OpenFlow switch. The port information contains the id and port number of the switch, the amount of data sent by the port, and the amount of data received by the port. The system takes the sum of the amount of data sent and received by the port as the port forwarding amount, and takes 1 second as the request interval. The port forwarding amount in the two adjacent response messages is calculated to obtain the bandwidth occupation of the corresponding port (in Mbit/s). After obtaining the bandwidth occupancy of each link and the number of large flows, the solution calculates the maximum bandwidth occupancy and the maximum number of large flows of all links on the path for each path in the alternative path, as the path Bandwidth occupancy value and the number of large flows. After obtaining the bandwidth occupancy value of each path and the number of large flows, each path is traversed. If the hop count of the path is less than or equal to the hop count of x other paths, the path score is increased by x; if the bandwidth occupancy value of the path is If the number of other large flows is less than or equal to the occupancy value of y other large flows, the score of the path is increased by y; if the number of large flows is less than or equal to the number of z other large flows, the score of the path is increased by z. Finally, the path with the highest score is selected. If multiple paths have the same score, the path with the lower subscript value in the result list of the KSP algorithm and the front position is selected.

Simulation Parameter Setting
In order to be consistent with other variables in the control experiment, each strategy uses the same experimental settings. The specific experimental settings are as follows:

Comparative Traffic Scheduling Strategy
In addition to the traffic scheduling strategy introduced in this paper, three other very well-known traffic scheduling strategies are also selected. The following briefly introduces these three traffic scheduling strategies: (1) ECMP: This experiment will implement ECMP by using flow entries and Select type group tables.
(2) Hedera: According to the existing experimental verification, its performance is better than ECMP.
(3) PureSDN: This scheme is the most traditional traffic scheduling strategy that uses SDN to implement the optimal path calculation scheme based on the number of path hops. It first uses KSP to select k paths with the least number of hops, and then selects the one with the lightest load as the best path.

Flow Model
This experiment simulated 8 traffic models of data center networks: random, stag_0.1_0.2, stag_0.2_0.3, stag_0.3_0.3, stag_0.4_0.3, stag_0.5_0.3, stag_0.6_0. 2. stag_0.7_0.2. "Random" means a traffic model formed by the host that initiates a large flow by selecting the opposite end in a completely random manner; other traffic models take stag_0.1_0.2 as an example, and stag_0.1_0.2 means the one below the access layer switch The traffic (that is, the traffic in the Pod) accounts for 10% of the total traffic, the traffic between different access layer switches in the Pod accounts for 20%, and the traffic between Pods accounts for 70%. Repeat the experiment 20 times for each flow model, and finally take the average of the experimental results as the final data. In the experiment, IperfTCP is used to simulate a large flow, and each host sends a large flow. It is equivalent to that each host is the client of the traffic, and then the server is selected according to the probability set by different traffic models.

Simulation Result Analysis
Throughput rate represents the actual amount of data that a network passes through a certain network in a unit time, which best reflects the performance of the network (Fig.1). When the proportion of traffic in the Pod is high, the average throughput of the four schemes is roughly the same as the standardized total throughput. But when the traffic in the Pod is relatively low, the performance of our proposed strategy is better than the other three strategies in terms of two indicators. The ratio of the number of links that have transmitted data to the total number of links in the network (Fig.2). The higher the utilization of the link, the more utilization of link resources, but when the utilization of the link is too high, it may cause excessive delay. In the simulation experiment of this article, because the network link is generally two-way, when calculating the link utilization and bandwidth utilization, a path is considered as two opposite directions, and the Pod and the access layer switch are not calculated separately. Direct links are included. In the case of low traffic in the Pod, the link utilization rate of ECMP is the lowest, and the MFFC scheme proposed in this paper is close to the link utilization rate of the Hedera and PureSDN strategies.  The bandwidth utilization rate of the network refers to the ratio of the bandwidth actually used by the link during data transmission to the total bandwidth. In general, low bandwidth utilization means that link resources are not fully utilized, and high utilization means that the link is very prone to congestion. As shown in Figure 3, in the case of high traffic in the Pod, the bandwidth utilization of the four traffic scheduling strategies is low. In the case of low traffic in the Pod, the links with lower utilization in the link bandwidth of ECMP account for a relatively high proportion, indicating that its link bandwidth resources are wasted a lot, while PureSDN and the MFFC strategy proposed in this paper are in the majority In this case, there are more links with link bandwidth utilization between 40% and 60%, which shows that the strategy proposed in this paper has better performance than ECMP and Hedera algorithms under the performance index of bandwidth utilization.

Conclusion
This paper proposes a multi-service-oriented multi-factor traffic scheduling scheme. It improves and perfects the traffic scheduling strategy, and introduces a set of scoring mechanism. Based on the Mininet simulation platform, use Ryu to control to generate probability model traffic to simulate data center network traffic. The experimental results show that the scheme improves the network throughput and bandwidth utilization, and improves the performance of the data center network.