Application study of ant colony algorithm for network data transmission path scheduling optimization

: With the rapid development of the information age, the traditional data center network manage - ment can no longer meet the rapid expansion of network data tra ﬃ c needs. Therefore, the research uses the biological ant colony foraging behavior to ﬁ nd the optimal path of network tra ﬃ c scheduling, and intro - duces pheromone and heuristic functions to improve the convergence and stability of the algorithm. In order to ﬁ nd the light load path more accurately, the strategy rede ﬁ nes the heuristic function according to the number of large streams on the link and the real - time load. At the same time, in order to reduce the delay, the strategy de ﬁ nes the optimal path determination rule according to the path delay and real - time load. The experiments show that under the link load balancing strategy based on ant colony algorithm, the link utilization ratio is 4.6% higher than that of ECMP, while the tra ﬃ c delay is reduced, and the delay deviation ﬂ uctuates within ± 2 ms. The proposed network data transmission scheduling strategy can better solve the problems in tra ﬃ c scheduling, and e ﬀ ectively improve network throughput and tra ﬃ c transmis - sion quality.

determine the best path. The purpose of this study is to improve network throughput through traffic scheduling under different load conditions, reduce link congestion, and optimize user service quality. The contribution of this study is to break through the limitations of traditional network architecture in scalability and global visualization, and make the traffic scheduling algorithm compatible with SDN.

Literature review
The rapid development of network data applications has also brought pressure on global network security construction and network scheduling management. Ant colony algorithms (AG), by mimicking the behavior of biological groups, are often utilized by national and international scholars in the design of path optimization strategies in different fields. Sachdeva et al. developed an Android application using ant colony optimization and similar peer suggestions for delivering personalized and adaptive e-learning paths to learners. Experiments have shown that the application caters to many students and helps in reducing the time required to complete any subject or course [4]. Zhao et al. introduced an opposition-based learning mechanism in AG and proposed an improved AG ADNOLACO. Experiments have shown that the algorithm effectively improves the convergence of AG speed and improved convergence accuracy, while improving the ability to find a balance between local and global optimal solutions [5]. Cerda et al. proposed an algorithm based on the ant colony optimization metaheuristic to dynamically optimize the decision threshold provided by a pairwise trading investment strategy, and experiments showed that the ACO-PT algorithm converges quickly and can be effectively used in deep markets [6]. Mahfoud et al. proposed an ACO algorithm for direct torque control based on proportional-integral differential speed regulation theory study and experimentally demonstrated the effectiveness of the proposed ACO-DTC in the presence of system nonlinearity [7]. Umar IA et al. proposed an ACO algorithm for direct torque control based on proposed an optimal vehicle maneuver scheduling method based on simulated annealing ant colony optimization (saACO) image recognition algorithm and experiments showed that the in-vehicle lane detection system shows good performance of saACO based lane detection system and has better performance compared to standard ACO methods [8]. Wang et al. proposed a multi-intelligence and ant swarm optimization algorithm to solve the ship integrated power network reconfiguration problem, and experiments showed that the optimization method can reconfigure the integrated power system network accurately and efficiently [9]. Wang et al. integrated the concept of reinforcement learning into the movement of artificial ants and proposed a variable radius perception strategy to calculate the transfer per pixel probability for avoiding undetected and false detection of certain pixels in image edges. Experiments show that the algorithm can effectively extract the contour information of underwater targets, better maintain the image texture, and has ideal anti-interference performance [10]. Liu et al. proposed an improved wireless sensor network transmission method based on ant colony optimization and threshold agent re-encryption for securing data in IoT wireless sensor networks. Experiments show that the method can resist internal and external attacks and ensure secure and efficient data transmission [11].
Yi and Peng used an artificial bee colony algorithm to obtain the shortest path analysis for each cluster head node in IoT transmission and experimentally showed that the algorithm can effectively reduce the amount of data transmitted from sensor nodes to sink through cluster head node fusion, improve data collection efficiency, energy consumption balance and network reliability, and extend the network life cycle [12]. Mu et al. transformed the routing problems such as the limitations of protocols under rapid data growth conditions into a Markov decision process to solve the problem of high blocking probability due to increasing data volume by combining deep reinforcement learning. Experiments have shown that this method can significantly reduce the probability of data congestion and improve network throughput [13]. To take full advantage of multiple video transmissions and considering the connection between video coding and transmission, Li et al. proposed a joint optimization method for session HD video services. Experiments show that the method outperforms existing schemes in terms of data transmission and playback quality [14]. Pyeon and Yoon proposed an efficient multiplexed pipeline (EMP) transmission to support low-latency and energy-efficient large data transmission under various network conditions. Experiments show that EMP outperforms existing protocols in terms of transmission time and energy efficiency, and maintains improved EMP performance regardless of the network environment (e.g., link quality, hop count, and network density) [15]. Wang and Liu designed a distributed communication platform based on P2P network technology to implement internetwork on concurrent path transmission. Practical experiments on the simulation platform demonstrated the effectiveness of the approach [16].
To sum up, the application of AG in market research, vehicle scheduling, power, and other aspects has been very mature. At the same time, the link load balancing strategy of data center network in the existing literature often judges the link load condition by monitoring the link residual bandwidth in real time. However, in the actual network environment, if there are multiple large flows on a link, the instantaneous load fluctuation is large, so the instantaneous residual bandwidth information obtained by the controller at a certain time cannot fully reflect the load condition of the link under the current state. In addition, while improving the utilization of network resources, we cannot ignore the service quality requirements of the business, otherwise the user's experience of accessing the data center will drop sharply. This study uses AG to optimize link load scheduling, hoping to improve network performance and provide stable service guarantee. This study uses AG to optimize link load scheduling. Its main contribution is to actively combine the load balancing scheduling strategy of large and small traffic with AG. In the field of combinatorial optimization, which ACO is good at, it completely solves the routing calculation problem constrained by various conditions such as link load and delay.
3 Traffic scheduling method based on optimized AG

AG and data transmission scheduling optimization strategy research
Optimizing the network data transmission path scheduling first requires analyzing the topology and traffic characteristics of data network centers at the current stage. Traffic patterns in data center networks are characterized by dynamic and diverse changes due to the increasingly wide range of scenarios and domains in which data servers are used, leading to data centers hosting various applications and Web tasks from time to time [17]. Analyzing the data traffic characteristics in terms of overall web tasks, since most data center tasks are created by web activities, it can be concluded that most of the streams in data centers have short byte length and duration, while the traffic transfer tasks with large byte counts occupy more than 90% of the overall tasks, despite occupying only a small portion [18]. Therefore, the prerequisite for optimal scheduling of data traffic is to identify the size of the network traffic. The application of SDN architecture can capture information about the size characteristics of network flows for the study of optimization policies [19]. In addition, the data center architecture is classified into three types of core architectures: server core architecture, switch core architecture, and hybrid network architecture, and Figure 1 shows the topology of a network data center with a switch as the core architecture.
As shown in Figure 1, the fat-tree switch core network architecture contains several servers in the access layer, switches in the aggregation layer and switches in the core layer. The simple structure and diverse paths allow it to be effectively adapted to the current stage of big data networking tasks and cloud computing. In the strategy of this study, when judging the link load, consider combining the number of large streams on each link with the link real-time load, so as to find the light load path more accurately and reduce the probability of misjudgment. The more links a large stream passes through, the more resources it occupies. Moreover, there are many reachable paths between two end-hosts in the data center, and finding the optimal path from all the reachable paths consumes high resources and delay. At the same time, to ensure the quality of service of the stream, the first consideration should be the transmission delay of the stream. Therefore, the completion of these strategies needs to rely on intelligent colony optimization algorithm. In biology, the positive feedback mechanism presented by the colony behavior of ants and distributed computing can seek the shortest path optimization of foraging behavior. AG has the characteristics of distributed computing, positive feedback and robustness, and can also be integrated with other optimization algorithms, making it easier and more efficient to search for the optimal solution. Therefore, the study introduces the path selection behavior into the traffic path scheduling of data centers, and first describes the ant behavior strategy using equation (1).
where P ij a represents the probability that an individual ant a selects a path at location i to location j, τ ij represents the pheromone concentration between locationi and location j, η ij represents the heuristic function between the two locations, ω represents the pheromone weights, and θ represents the weights of the heuristic function. After the path from location i to location j is selected by the ant, the pheromone is left in the channel and the secreted pheromone varies with concentration time over time, and the reasons for the variation include natural volatilization of pheromone as well as pheromone increment, so the mathematical model formulation to calculate the variation in pheromone concentration is shown in equation (2).
where n represents the time unit, ( ) + τ n 1 ij represents the change in pheromone concentration in the path from position i to position j, φ represents the pheromone volatilization factor, m is the total number of ant individuals in the colony, and △τ ij k represents the pheromone increment in the pathway between two positions of the individual ants k. The formula for calculating the pheromone increment secreted by ants is shown in equation (3).

Core layer
Access layer

Convergence layer
Access layer

Access layer
Convergence layer Access layer Figure 1: Topology of fat-tree network data center with switch as core architecture.
Equation (3) includes the calculation of pheromone increment based on the total amount of pheromone secreted by ants and the length of the local path of ant movement. Q denotes the total amount of pheromone released by all ants in the colony, and d ij denotes the distance from position i to position j. For the sake of ensuring the stability of the algorithm and the complexity of computational convergence, the study integrates the overall pheromone and path length models to calculate the pheromone increment, which is described by equation (4).
where L k denotes the overall length of the path passed by the ant individuals, the longer the path, the less the pheromone increment, for the control of pheromone concentration and increment, avoiding the failure of the heuristic function during the calculation. Meanwhile, the network stream state information is collected with the help of SDN control module to discern the task data stream size.

AG algorithm construction for link load balancing policy
Link load imbalance is a situation where the traffic scheduling policy of the data network center does not match the actual demand of network task traffic and the state of network flow information, and link load imbalance can easily lead to network congestion and waste of broadband resources [20]. In the network data center, each large stream will preempt the larger transmission bandwidth in the link as much as possible, resulting in a large transient load fluctuation in the case of multiple large streams on a link. The instantaneous load information obtained at a certain time does not fully reflect the load situation of the link under the current state. At the same time, the current network services require high quality of network services, such as multimedia information, which not only has large data volume, but also requires real-time data transmission. Therefore the proposed AG based on the link load problem first needs to identify the realtime state of the links by judgment [21]. According to the study's classification of the characteristics of data traffic tasks in data network centers, although small streams of data occupy most of the tasks, their traffic size occupies a small part of the overall traffic scheduling, and the key factors influencing the link balance for the network are concentrated on the network large stream tasks, so the study solves the scheduling computation tasks of network large stream tasks by the AG, and small streams by the traditional ECMP scheduling [22]. According to the network flow information of the switch in the access layer of the SDN architecture to calculate the size of network task traffic, the SDN framework structure diagram is shown in Figure 2, which is mainly divided into SDN controller and OpenFlow switch, and the information interaction between the controller and the switch through the OpenFlow protocol, the controller and switch play their respective roles in the process of scheduling data flows by the system, and play The controller and switch play a crucial role in the process of scheduling data flows in the system [23].
In Figure 2, the data flow passes through the detection module, monitoring the network situation, and the traffic information is stored in the collection module; when the scheduling flow appears, the forwarding module calculates the forwarding path of the flow and sends it to the switch by the control module, and the simple structure makes it effectively adaptable to the current stage of big data network tasks and cloud computing. The network flow information transmission efficiency of the switch is calculated by the formula shown in equation (5).
where ψ represents the ratio of the transmission rate of the bandwidth to the maximum bandwidth of the link, by t1 is the number of bytes counted by the access layer switch at t 1 , by t2 is the number of bytes counted by the access layer switch at t 2 , B represents the maximum bandwidth of the link, and when the ratio of ψ is greater than 10%, the task data stream is determined to be a large stream. Aggregate the paths between hosts at both ends of the network task flow as and n indicates the number of links in the path aggregate. Set the bandwidth occupied by the linki in the path F to B i , then the overall load of the path is . And the available bandwidth on the path i is expressed as in equation (6).
where C ij is the total bandwidth capacity of the link, and B ij is the used bandwidth of the link j. And the average of all links of pathi can be calculated as shown in equation (7).
where q is the number of hops of the path. Meanwhile, setting the round-trip time difference from switch A to the controller as de a and from switch B to the controller as de b , the link transmission delay between the switches can be calculated, and the formula is expressed as shown in equation (8).
where de a b , denotes the transmission delay between switches A and B, de 1 denotes the packet transmission delay, and de 2 denotes the packet transmission reverse delay. Therefore, the delay of the path ensemble is derived and expressed as equation (9).
Therefore, the bandwidth load, transmission delay, and other parameters of the above links are used as indicators of link state evaluation, and the path with the highest evaluation is optimized according to the AG, and it is scheduled as the best path strength of the large flow. Since the number of links within the path set at both ends of the network data task is uncertain, directly calculating the optimal path using the AG in the used paths will lead to a large amount of computational work, thus consuming too many resources, so first use the information statistics module to filter the shortest path set between the sending switch address of the data stream and the destination switch address, and then use the AG to find the optimal path in the shortest path set, the optimal calculation model is shown in equation (10).   denotes the heuristic function 2 when the path i is at t, α denotes the weight factor of the pheromone, β denotes the weight factor of the heuristic function, and the definition of the heuristic function described in equation (11). we know that the more the number of large streams in the path or the heavier the load, the smaller the heuristic function, and the heuristic function is used as the numerator in the path selection probability formula, so the smaller the heuristic function indicates the smaller the probability of ants selecting the path. Meanwhile, according to the overall pheromone update method, the pheromone update calculation formula in the optimal path ensemble is shown in equation (12).
where ( ) + τ t n i represents the change in pheromone concentration on the path i, ( ) τ t i represents the pheromone concentration on the path i at t, and ρ represents the pheromone volatilization factor. ( ) τ t Δ i The increment of pheromone on the path i at t is calculated by dividing the total amount of pheromone constant released by the individual ant colony by the length of the optimal path, and its mathematical expression is shown in equation (13).
where Q denotes the total number of pheromones and L i denotes the path length. The study explores the strategy of scheduling network traffic to balance the network load of the link, and in order to present the path selection criteria clearly, the method design considers the delay and the real-time load of the link together, and uses the path state evaluation function to illustrate the load of the link, whose mathematical expression is shown in equation (14).
where e i represents the state function of the path i, θ represents the delay weight, de i represents the delay of the path i, and lo i represents the real-time network load of the path i. Finally, the minimum state evaluation value is the optimal path, and the optimal path is expressed as shown in equation (15).
According to the link load balancing method proposed in the study, the switch of SDN architecture is used to judge the traffic size, the AG algorithm path optimization is performed for large flows, and the traditional ECMP algorithm scheduling is selected for small flows, and the overall optimization process of network link load balancing is shown in Figure 3.
As shown in Figure 3, since the uneven distribution of large flows is the main factor that causes the imbalance of network load, the strategy proposed in this chapter develops ACO dynamic traffic scheduling algorithm for large flows. However, in view of the small flow's small impact on network congestion and high delay requirements, if dynamic traffic scheduling algorithm is used to route it, it will increase the complexity, so the small flow still uses static routing algorithm ECMP for scheduling. When the switch receives the data flow sent by the host, it determines whether it is a direct connection (that is, whether the destination host is a host under the switch of the same access layer) according to its destination address. If it is, it forwards it directly downward, otherwise, it needs to determine the flow. When the switch determines that the flow belongs to small flow, it uses ECMP to schedule directly. When the switch determines that the flow belongs to a large flow, it sends a message to the controller. According to the topology, load, delay, and other network information obtained from the switch, the controller uses the load balancing algorithm proposed in this chapter to calculate the large flow forwarding path, and sends the flow table information. Then, the switch forwards the large flow according to the distributed flow table.

Performance analysis of network data transmission path scheduling optimization strategy
In order to verify the performance of the link load balancing AG optimization algorithm proposed in the study, this experiment uses the Mininet simulation platform to construct a fat-tree network topology with the number of core switches and pods of 4 within it, a large flow duration of 60 s, and an occupied broadband size interval of 100 Mbps to 1 Gbps. The remaining parameters of the simulation experiment are shown in Table 1.
For the purpose of determining the delay weights, the study compares the link utilization and delay, average throughput, and delay deviation of the AG link load balancing algorithm by selecting the weight values of 0.5, 0.6, 0.7, and 0.8 from the weight range of 0.5-0.8, as shown in Table 2.
From Table 2, we can know that the AG algorithm network link load scheduling strategy has the highest link utilization when the delay weight is 0.6, and the link utilization ratio is 0.849. The AG link load balancing strategy has the lowest link delay when the weight value is 0.7, and the value is 1.512 ms. At the same time, the algorithm model has the highest average throughput when the delay weight is 0.5, which is 4.831 Gbps. In the delay deviation comparison, the deviation is 0.617 ms when the delay weight is taken as 0.7, which is lower than the deviation when the other weights are taken. In the overall comparison, the AG link load scheduling performance is optimal when the delay weight of the algorithm model is taken as   In order to test the performance of the proposed AG balancing strategy to improve the network load balance under low load, mainly from the aspect of link utilization, the experimental comparison with the classical traffic scheduling strategies ECMP and Hedera is carried out to simulate the load situation of nine data networks (0.2-1.0) and the impact on the strategy performance in nine traffic models. Figure 4 shows the link utilization results of the three strategies. Figure 4(a) shows the change in link utilization of the three strategies with the increase in network load. In Figure 4(a), the link utilization of ECMP strategy is the lowest because ECMP strategy schedules large flows in equal paths, which limits the transmission paths to a certain range and cannot balance all network resources well, Hedera and AG algorithm strategies both schedule flows according to real-time network state, which achieves load balancing to a certain extent compared with ECMP. The Hedera and AG algorithm strategies both schedule traffic according to real-time network status, which can achieve load balancing to some extent, and improve link utilization by 3.3 and 4.6%, respectively, compared with ECMP.
The results of chain utilization in different traffic models are shown in Figure 4(b). Under random model, the difference of three strategies utilization is not significant, in staggered model, when the percentage of traffic between transmission zones increases, the upper links are fully utilized and the link utilization of three strategies is generally higher. When the percentage of traffic within transmission zones increases, only the bottom links are utilized and the link utilization of three strategies link utilization decreases. Before s.4 model, the link utilization of Hedera and AG algorithm strategy improves 1.8 and 2.6%, respectively, compared to ECMP, and after s.4 model, the utilization of ECMP improves 0.65% compared to Hedera, and the link utilization of AG algorithm strategy improves 1.2% compared to ECMP, which indicates that Hedera and AG algorithm strategies are suitable for transmission inter-zone traffic and ECMP is suitable for the case of high intra-zone traffic. In addition, the average network throughput comparison of the three network traffic scheduling strategies is shown in Figure 5.
In Figure 5, the average network throughput represents the amount of data elaborated by different link load scheduling strategies per unit time, and the larger value indicates better data transmission performance of the link. From Figure 5, it can be seen that the average link throughput of all three network traffic   scheduling strategies increases with the number of large streams sent by the hosts. At the same time, the gap between the algorithms gradually widens as the number of large streams sent by the hosts increases. The performance of the three algorithms is superior and the gap between algorithms is small when the number of large flows is small. The traditional ECMP algorithm performs the worst among the three algorithms, with an average throughput of only 7.21 Gbps at six large streams, while the AG algorithm achieves the highest average throughput five times during the test with the number of large streams from 1 to 6. The experiments show that the AG algorithm is more stable in processing large flows on the link, and the AG link scheduling strategy takes into account the actual situation of the link and the analysis of network flows, which can effectively reduce the waste of link bandwidth resources. To test the quality of service of the proposed AG strategy under high load, the packet loss rate under different traffic models is examined and compared with ECMP and Hedera, and the results are shown in Figure 6. In Figures 6, s1-s6 denote the traffic patterns stag-0.8-0.1, stag-0.6-0.3, stag-0.5-0.2, stag-0.4-0.3, stag-0.2-0.2, and stag-0.1-0.1, respectively, and R denotes the random pattern. From Figure 6, we can see that the three scheduling policies have the lowest packet loss rate in stag-0.4-0.3 traffic pattern, which are 0.01604% for AG link load balancing policy, 0.01618% for Hedera scheduling policy, and 0.016% for ECMP policy. The three scheduling policies have the highest packet loss rate in the stag-0.1-0.1 traffic pattern, which are 0.01673% for AG link load balancing policy, 0.01664% for Hedera scheduling policy, and 0.01731% for ECMP policy. And in the random traffic pattern, 0.01649% for AG link load balancing policy, 0.01653% for Hedera scheduling policy, and 0.01673% for ECMP policy. The overall curve trends of packet loss rate and images in random mode show that as the percentage of traffic within pods increases, large flow collisions under the traditional scheduling strategy of ECMP lead to an increase in packet loss rate, while the AG algorithm strategy reduces the possibility of network blockage occurrence through large flow scheduling. Meanwhile, the experiments show the quality of service of the optimized strategy from the round-trip delay test of the traffic, and the average round-trip delay results in different traffic models are shown in Figure 7.
From Figure 7, we can see that the three strategies have uncertainty under different traffic models. In general, with the increase in traffic between transmission zones, the round-trip delay increases, because the switch needs to discriminate traffic for path forwarding, the transmission efficiency decreases and the delay increases. With the data flow increases in the link transmission zones, the round-trip delay of all three strategies shows a decreasing trend, because the traffic delivery under the same switch can. When Ant Colony Economic fragment calculates the forwarding path, it gives priority to the path with less fluctuation in the remaining bandwidth of the link, which makes the remaining bandwidth distribution on the network link more balanced and therefore significantly reduces the round-trip delay. To further test the transmission efficiency of the strategies, the delay deviation curves of the three strategies with network load changes are obtained as shown in Figure 8.
From Figure 8, it can be seen that ECMP has a delay deviation range of about −5 to 10 ms and a stability deviation range of −5 to 5 ms when loading load, Hedera has a stability delay deviation range of about ±3 ms, and link load balancing-recombinant fragment (LLB-RFrag) has a stability delay deviation range of about ±2 ms, indicating that the LLB-RFrag scheduling optimization strategy proposed in the study has good service quality. Finally, in the experiment, the AG link scheduling method proposed in this study is compared with the scheduling strategy of traditional time sensitive network and traditional graph neural network, as shown in Table 3. Stag in the network link data flow mode taken in this study as _ 0.1_ 0.2 is used as an example to generate traffic. The number of large streams sent increases from 2 to 6. The performance test usually takes 10 s. It can be seen from the table that the AG link scheduling optimization algorithm proposed in the study has the shortest running time. Only 249 s is used for 6 maximum flows, which reduces the scheduling running time by 60 s compared with the traditional graph neural scheduling method and time sensitive scheduling.

Discussion
According to the research, the traffic in the data center network can be divided into two types: large flow and small flow. In the network, the traffic duration is long, the traffic volume is large, and the data flow exceeds 10% of the link bandwidth, which is called large flow. In the data center network, the elephant stream accounts for about 80% of the total traffic bandwidth. Usually, applications include virtual machine migration, data backup, etc. The data stream is small in size and short in duration. The common applications of small streams include online business, viewing web pages, conducting business, etc. In the specific data network, due to the high proportion of large flow and the high bandwidth, if the data flow is not better protected, it will easily lead to corresponding network congestion, and then make the corresponding network face paralysis. Due to the different transmission performance requirements of these two data streams, the large stream requires high throughput, while the small stream requires high latency. Therefore, how to schedule the large stream in the data center network reasonably and efficiently is the driving force for the development of various network services at present. Based on this, this study proposes the idea of classifying large flow and small flow in dynamic scheduling. In the traditional data center network, most of them adopt multi-element topology, which contains multiple equivalent paths. The equivalent multi-path algorithm (ECMP) is the most widely used in this structure. Its idea is that there are multiple equivalent available paths to the same destination node. Therefore, the data flow of multiple identical target addresses can be evenly distributed to multiple equivalent paths in a certain way to achieve link load balancing at one   time. However, in the real network environment, many traffic has dynamic properties, and the relevant parameters on many paths in the network meeting also have differences, such as available bandwidth, packet loss rate, etc. In these cases, it is difficult to improve the utilization of network links if the network overhead above each path is considered to be consistent. Hedera scheduling method is actually a scheduling based on SDN technology, and shows different performance from static scheduling, which significantly improves the scheduling quality of large flows. However, its disadvantage is that the interaction of massive information between the controller and the switch increases the resource consumption when detecting the flow information. The strategy and method of SDN combined with ACO algorithm for largeflow scheduling and small-flow ECMP scheduling proposed in this study uses AG to solve the optimal value of link state to evaluate and define the best path. In order to verify the performance of the proposed strategy, this section compares AG with the classic static traffic scheduling strategy ECMP and the dynamic traffic scheduling strategy Hedera. It mainly verifies the performance of the strategy proposed in this chapter in improving the network load balance and ensuring the network quality of service. The load balance is mainly reflected in the link utilization and network throughput, while the quality of service is mainly analyzed from the aspects of the flow round-trip delay and round-trip delay deviation. This study describes the proposed traffic scheduling strategy in detail, including the comparison of different scheduling strategies, the performance difference under different traffic modes, and the comparison of scheduling optimization time consumption.

Conclusion
Aiming at the problems that a large amount of data are easy to jam the network, high data transmission delay, and reliability, the traffic scheduling strategy based on AG is studied to find the optimal transmission path for large traffic data. In order to test the performance of the proposed strategy, simulation experiments are conducted. In the link utilization experiment, the AG load balancing strategy can maximize the use of link resources and accurately find the light load forwarding path, which is 4.6% higher than the link utilization of ECMP. The AG algorithm achieves the highest average throughput for five times during the test of 1-6 large streams. In addition, the traffic round-trip delay of the AG strategy is low. At the same time, the proposed scheduling optimization strategy has good service quality, and its stable delay deviation range is about ±2 ms, which is lower than other scheduling algorithms. Experiments show that the proposed strategy can better solve the problems caused by different network loads, improve network throughput, improve the stability of traffic transmission, and ensure the quality of service. Although this research has made some achievements, there are still many shortcomings, and the application scope of the proposed strategy needs to be further expanded.
Funding information: The research is supported by Research and implementation of remote automatic installation operating system based on PXE (No. 171280).
Author contributions: Peng Xiao collected the samples, analysed the data, analysed the results and wrote the manuscript.