Data center flow scheduling method based on differential evolution fusion particle swarm optimization

Abstract

The invention belongs to the technical field of networks, and particularly relates to a data center flow scheduling method based on differential evolution fusion particle swarm optimization, which comprises the following steps that: a switch receives a data flow and judges whether a destination address of the data flow belongs to a local direct connection host, if so, the data flow is forwarded; otherwise, judging the size of the data stream, and if the data stream is a small data stream, scheduling the data stream through an ECMP algorithm, directly distributing a path and issuing a flow table; if the data stream is a big data stream, calculating an optimal path by adopting a PSODE algorithm, optimizing a link load and time delay according to the optimal path, and issuing path information to a switch to execute scheduling; according to the method, the advantages of the SDN and the heuristic algorithm are combined, link load balancing optimization and time delay control are realized in a data center network traffic high-load scene, and the overall performance of the network is further improved.

Description

Technical Field

[0001] The present invention belongs to the field of network technology, and particularly relates to a data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm. Background Art

[0002] With the development of cloud computing and big data technologies, internal traffic scheduling in data centers has become a key issue for improving network performance. Existing scheduling methods mainly rely on static policies or traditional optimization algorithms, which are difficult to adapt to the dynamically changing network environment, resulting in low resource utilization and poor load balancing.

[0003] The OpenFlow technology based on SDN (Software Defined Network) has become a key technology for solving network management problems. SDN separates the control plane and the data plane of the network, and realizes communication between the two through a southbound protocol. The control plane is programmatically controlled by a centralized controller, while the data plane becomes more generalized, and its main responsibility is to receive instructions issued by the control plane and complete the data forwarding task. Through this architecture, the controller can centrally manage the network, monitor the network status in real time, obtain key information such as link load and traffic bandwidth requirements, so as to provide a decision-making basis for the optimized management of network traffic. These decision-making information will then be sent to the switch, and the switch will execute corresponding operations according to the instructions of the controller to achieve efficient management and scheduling of data streams. Summary of the Invention

[0004] To solve the problems of link congestion, latency degradation, and parameter rigidity in traditional traffic scheduling methods under high-load scenarios, the present invention proposes a data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm. The method includes: the switch receives a data stream and determines whether the destination address of the data stream belongs to a locally directly connected host. If so, the data stream is forwarded; otherwise, the size of the data stream is judged. If it is a small data stream, the data stream is scheduled by the ECMP algorithm, the path is directly allocated, and the flow table is issued; if it is a large data stream, the PSODE algorithm is used to calculate the optimal path, the link load and latency are optimized according to the optimal path, and the path information is sent to the switch to execute the scheduling.

[0005] Advantages of the present invention:

[0006] The present invention aims at the problem of network performance optimization in high concurrent traffic scenarios, realizes visual monitoring of the traffic status of the whole network through the SDN centralized control plane, and combines the improved meta-heuristic algorithm to perform optimal path planning under multiple constraints, thereby overcoming the local optimal problem of traditional methods under dynamic loads. The present invention schedules data streams through the PSODE algorithm, and optimizes the weight parameters in the algorithm through a multi-objective game model, thereby improving the accuracy of the model. The present invention combines the advantages of SDN networks and heuristic algorithms, and realizes link load balancing optimization and delay control in high-load scenarios of data center network traffic, thereby improving the overall performance of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is the flow chart of PSODE algorithm of the present invention;

[0008] Figure 2 It is the PSODE system framework diagram of the present invention;

[0009] Figure 3 It is the overall flow chart of the present invention;

[0010] Figure 4 This is the experimental topology diagram of the present invention. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] Aiming at the problems of unbalanced link load in data centers, differential impact of large and small flows, and lack of dynamic scheduling capabilities, a data center network traffic scheduling strategy based on differential evolution fusion particle swarm algorithm (PSODE) is proposed. This strategy is based on the SDN global topology and real-time link status to improve the premature convergence problem of the particle swarm algorithm. By utilizing the complementary characteristics of the particle swarm algorithm (PSO) and the differential evolution algorithm (DE), a serial hybrid strategy is designed. With PSO as the leading framework, DE operations are triggered dynamically and selectively in iterations to solve the premature convergence problem of the particle swarm algorithm. At the same time, adaptive inertia weights and crossover probability parameters are designed to balance the breadth and depth of path search, and achieve a balance between convergence speed and global search performance. In addition, a fitness function with maximum link utilization, end-to-end delay and path diversity as the goals is constructed to find the best transmission path for the large flow to be scheduled.

[0013] A data center traffic scheduling method based on differential evolution fusion particle swarm algorithm, such asFigure 3 and Figure 4 As shown in Figure 4 , the method includes: the switch receives a data stream and determines whether the destination address of the data stream belongs to a locally directly connected host. If so, the data stream is forwarded; otherwise, the size of the data stream is judged. If it is a small data stream, the data stream is scheduled by the ECMP algorithm, the path is directly allocated, and the flow table is issued; if it is a large data stream, the PSODE algorithm is used to calculate the optimal path, the link load and delay are optimized according to the optimal path, and the path information is sent to the switch for scheduling.

[0014] In this embodiment, scheduling the small data stream by the ECMP algorithm includes:

[0015] S101. Obtain multiple equivalent paths to the target address through the SDN controller, and the link metric values of each equivalent path are equal;

[0016] S102. Extract the five-tuple features of the data stream, including the source IP address, destination IP address, source port number, destination port number, and transport layer protocol type;

[0017] S103. Input the five-tuple features into a predefined hash function to generate a hash value, and determine the initial allocation path through modulo operation;

[0018] S104. Generate a flow table entry including a matching field and an action instruction based on the initial allocation path, and issue it to the flow table cache of the switch;

[0019] S105. Forward the data stream according to the flow table entry, and trigger the update of the flow table entry when it is detected that the path state deteriorates.

[0020] Based on the analysis of the principles, advantages and disadvantages of the two algorithms of Particle Swarm Optimization (PSO) and Differential Evolution (DE), it can be seen that the two have strong complementarity in search characteristics: PSO can quickly approach the optimal solution in the initial stage, but it is easy to fall into local optimum; DE is good at maintaining population diversity and has a strong ability to jump out of local optimum, but its initial convergence speed is usually not as fast as PSO. Therefore, this paper combines the two to propose a hybrid algorithm that combines Particle Swarm Optimization and Differential Evolution (Particle Swarm Optimization–Differential Evolution, PSODE) aiming to balance the fast convergence of the particle swarm algorithm and the global optimization ability of the differential evolution algorithm, and improve the solution efficiency and quality of the data center network traffic scheduling problem.

[0021] The idea of the PSODE fusion strategy is as follows: retain the advantage of the particle swarm optimization algorithm in quickly approaching the global optimal region in the initial stage, and dynamically and selectively introduce the differential mutation and crossover operations of the differential evolution algorithm during the PSO iteration process, so as to maintain population diversity and avoid the algorithm falling into local optima. In addition, this strategy realizes a reasonable balance between the convergence speed and the global search performance of the algorithm by adaptively adjusting parameters such as the inertia weight, mutation factor, and crossover probability. The following details the specific design of the PSODE fusion algorithm.

[0022] Considering the complementary characteristics of particle swarm optimization and differential evolution, the PSO algorithm and the DE algorithm are serially hybridized. Specifically, the PSODE algorithm uses the particle swarm optimization algorithm (PSO) as the main search framework, quickly approaches the solution space through the velocity and position iteration models in most of the iteration processes, and inserts DE operations in a dynamic and adaptive manner, adjusting the introduction frequency of the DE operator based on population diversity as the trigger basis. In addition, the algorithm only selectively performs DE mutation and crossover operations on the particles with fitness lower than the population mean, thus effectively reducing the computational cost while ensuring the search quality, and protecting the high-quality solutions with high fitness, and finally achieving an efficient balance between the convergence speed and the global optimization performance.

[0023] The PSODE fusion algorithm first initializes the particle swarm, calculates the initial fitness, and determines the global optimal solution; then enters the iteration process dominated by the particle swarm optimization algorithm. In each iteration, it updates the positions and velocities of the particles, and evaluates and updates the individual and global optimal solutions in real time for the population fitness. In addition, after each iteration, the algorithm dynamically adjusts the trigger frequency of the differential evolution algorithm (DE) according to the variance change of the population fitness. When the trigger condition is met, only the particles with fitness lower than the population mean in the population are subjected to DE mutation and crossover operations to improve the global search ability of the algorithm while ensuring the computational efficiency; then, adaptively adjusts parameters such as the inertia weight, mutation factor, and crossover probability according to the current iteration number to maintain the balance between exploration and exploitation of the algorithm; finally, when the predetermined termination condition is reached, the algorithm ends and outputs the global optimal solution.

[0024] Specifically, as Figure 1 shown, scheduling the data stream through the PSODE algorithm includes:

[0025] S101. Initialize the particle swarm, and the particle population is the large data stream;

[0026] S102. Calculate the initial fitness of each particle in the particle swarm;

[0027] S103. Update the optimal value of the particle and the global optimal value according to the initial fitness;

[0028] S104. Update the velocity and position of the particle according to the optimal value and global optimal value of the particle;

[0029] S105. Calculate the fitness values of all particles in the population according to the velocity and position of the particles, and calculate the variance of the fitness values to determine whether the triggering condition of the DE algorithm is satisfied. If the triggering condition is not satisfied, determine whether the termination condition is satisfied. If not, return to step S102. If satisfied, output the global optimal solution; if the triggering condition is satisfied, execute step S106;

[0030] S106. Perform DE operation on the particles with poor fitness;

[0031] S107. Update the population and parameters of the particles, and return to step S102.

[0032] The fitness value of the particle is calculated by comprehensively considering minimizing the maximum link utilization, minimizing the end-to-end delay, and path diversity. Its expression is:

[0033] fitness = α·f(MLU) + β·f(Delay) + γ·f(Diversity)

[0034] Among them, α, β, and γ are weight factors, f(MLU) is the maximum link utilization, f(Delay) is the average end-to-end delay, and f(Diversity) is the path diversity.

[0035] The weight factors are dynamically generated through a multi-objective game model. To improve the calculation efficiency and real-time performance, a lightweight approximate game strategy is adopted for dynamic weight adjustment, specifically including:

[0036] S101. Define the maximum link utilization, end-to-end delay, and path diversity as three virtual game players, which are the multi-objective participants in the scheduling optimization;

[0037] S102. Based on the real-time network state, calculate the contribution scores of each game player respectively;

[0038] S103. Use the Nash bargaining solution with a penalty factor as the theoretical weight allocation basis. The calculation formula of the penalty factor is:

[0039]

[0040] Among them, S i is the score of the i-th game player, and S total is the sum of the scores of all game players.

[0041] S104. To achieve fast response and low computational overhead, a lightweight approximate game strategy is introduced to dynamically adjust the weights, specifically including:

[0042] S1041. Calculate the difference in the index scores between two adjacent generations:

[0043]

[0044] S1042. Based on the difference Δ i Use the Softmax function with temperature coefficient to generate weights:

[0045]

[0046] where τ = 0.7 is the temperature parameter and n is the total number of optimization objectives.

[0047] S105. Introduce a dynamic adjustment window mechanism to update the weights only when the variance of the current population fitness changes significantly. Its update period T is adaptively adjusted according to the variance of the population fitness, and the formula is:

[0048]

[0049] where is the variance of the current population fitness, is the historical maximum fitness variance.

[0050] S106. Introduce a mechanism for setting weight caching and reuse to avoid invalid calculations, specifically including:

[0051] S1061. Skip weight calculation when the change rate of the score vectors between two adjacent generations satisfies the following conditions:

[0052]

[0053] where S (t) is the target score vector of the current generation, ||.||2 is the L2 norm of the vector, and 0.05 is the change rate threshold.

[0054] S1062. Reuse the most recent valid weight value as the current weight;

[0055] S107. Finally, apply the dynamic weights to the multi-objective fitness function to guide the differential evolution fusion particle swarm algorithm to perform path scheduling search and output the optimal path allocation scheme.

[0056] Performing DE operations on particles includes: adaptively adjusting the mutation factor, inertia weight, and crossover probability of the particles; performing mutation, crossover, and selection operations on the particles according to the adjusted mutation factor, inertia weight, and crossover probability to obtain a new particle swarm.

[0057] To reasonably determine the introduction frequency of DE operations, a dynamic trigger period K is introduced, and this value is adaptively adjusted according to the variance of the population fitness. The specific calculation method is:

[0058]

[0059] Among them, K base is the basic trigger period, and K range is the trigger period adjustment range, and σ 2 is the variance of the fitness of the current population. is the historical maximum variance. The round function is used for rounding to ensure that the calculated trigger period is an integer. It is initially set to the variance of the first iteration. When the population diversity is low (σ 2 is small), the trigger period K is decreased and the DE trigger frequency is increased. When the population diversity is high, the trigger period K is increased and the DE trigger frequency is decreased.

[0060] To avoid large-scale perturbations to all particles, the DE operation is selectively performed only on the particles with fitness lower than the population mean. The following is a pseudo-code example of the specific steps.

[0061]

[0062] To achieve a good integration of PSO and DE, the balance between global search and local development is taken into account through adaptive parameters. The parameter adaptive strategy mainly involves the dynamic adjustment of the inertia weight ω, the mutation factor F, and the crossover probability CR. Specifically:

[0063] The inertia weight is:

[0064]

[0065] Among them, ω max is the maximum inertia weight, ω min is the minimum inertia weight, t is the current iteration number, and T max is the total number of iterations. A linear decreasing mode is adopted to strengthen the global exploration ability in the early stage of the algorithm and the fine search ability in the later stage.

[0066] The mutation factor is:

[0067]

[0068] Among them, F min is the minimum value of the mutation factor, F max is the maximum value of the mutation factor, f i is the fitness of the current particle, f best is the best fitness of the population, f worst is the worst fitness of the population, and ε is a very small value.

[0069] In this embodiment, the value of F min is 0.4, and the value of F max is 0.9.

[0070] The crossover probability is:

[0071]

[0072] Among them, CR base is the initial crossover probability, with a value outside 0.8; γ is the attenuation rate, with a value of 0.5. A higher crossover probability in the early stage increases the population diversity and gradually decreases in the later stage to protect high-quality solutions.

[0073] To effectively apply the traffic scheduling strategy based on the differential evolution fusion particle swarm optimization algorithm (PSODE) in the data center network, this paper implements the corresponding scheduling module on the SDN controller and builds a complete system framework in combination with the underlying topology and traffic monitoring tools. By differentiating the large flows and small flows, it can not only ensure the refined global path optimization of the large flows but also guarantee the fast forwarding and low control overhead of the small flows.

[0074] In this embodiment, the system framework design consists of core components such as a network information module, a traffic information module, a PSODE optimization engine module, a flow table management module, and a dynamic weight adjustment module. The system framework is as Figure 2 shown. As the brain of the SDN network, the controller is responsible for global management and scheduling. By collecting the whole network topology information, monitoring the link status, analyzing the traffic characteristics, and combining with the PSODE optimization engine to calculate the optimal path, it provides an intelligent scheduling scheme for the data flow. The core functions of the controller include topology discovery, traffic monitoring, intelligent scheduling, and flow table management. It interacts with the OpenFlow switch through the southbound interface to achieve the unified allocation and dynamic optimization of network resources. With the advantage of centralized control, the controller can real-time sense the network load, dynamically adjust the traffic path, improve the link utilization rate, reduce the delay, and effectively avoid network congestion, thus ensuring the efficient and stable operation of the data center network.

[0075] Network information module: This module realizes the automatic discovery of the network topology by sending Packet-out messages carrying LLDP (Link Layer Discovery Protocol) packets during the network initialization phase. When the switch returns a Packet-in message containing the LLDP packet, the link connection information between the switches can be recorded. By continuously traversing all switches in a loop, the network information module can construct the overall network topology view and pre-calculate K shortest paths for any two edge nodes based on the shortest path algorithm, and store them in the specified data structure. In addition, this module also periodically detects parameters such as the available bandwidth and link delay of each link to provide real-time status information and necessary decision-making basis for path evaluation and selection during the large flow scheduling process.

[0076] Traffic information module: This module periodically sends to the switch at the controller end

[0077] OFPFlowStatsRequest and OFPPortStatsRequest messages are used to obtain real-time statistical data of each switch port and flow table, so as to provide key information such as link load, data flow characteristics, and port utilization for the upper-layer scheduling strategy. Based on these statistical data, this module can also judge whether a flow is a large data flow according to the bandwidth demand threshold of the flow (set to be greater than or equal to 10% of the link bandwidth). If the threshold is met, the flow will be assigned to the PSODE optimization engine for refined routing and scheduling; for small data flows that do not reach the threshold, equal-cost multi-path (ECMP) is directly used for forwarding processing, thereby reducing the algorithm calculation overhead and improving the overall efficiency of the system.

[0078] PSODE optimization engine module: According to the obtained network topology information and traffic statistical data, generate an optimal path allocation scheme for large flows, and use the PSODE algorithm to calculate the scheduling path for large flows. The optimization goals mainly include minimizing the network maximum link utilization (MLU) and delay to improve the throughput and overall service quality of the data center network.

[0079] Flow table management module: The controller sends the path allocation information to the switch through the OFPT_FLOW_MOD message, and the OpenFlow switch executes the forwarding rules sent by the controller according to the scheduling information to complete the network traffic scheduling.

[0080] Dynamic weight adjustment module: By tracking the difference trend of each optimization target index in real time, adopting a fast weight allocation strategy based on exponential function and temperature control, only the incremental change of the target scores of adjacent generations needs to be calculated in each iteration. The normalized weight is generated through hardware-accelerated vectorized operations, and a dynamic adjustment window mechanism is introduced. The weight update period is adaptively extended according to the population diversity. When it is detected that the network state fluctuation is lower than the threshold, the cache is automatically enabled to reuse the historical weight. At the same time, the weight calculation is decoupled from the main optimization process by using multi-threaded asynchronous processing, achieving a balance between algorithm efficiency and resource consumption while ensuring the multi-objective optimization effect.

[0081] In this embodiment, the data flow interaction involves multiple links, mainly including data collection, traffic classification, optimization solution, rule downloading, and feedback update to ensure the accurate scheduling of large flows and the efficient forwarding of small flows.

[0082] Data collection: The traffic information module periodically collects data such as link utilization, queue delay, and packet loss rate from the switch. The controller obtains the switch port and flow table statistical information by sending OFPFlowStatsRequest and OFPPortStatsRequest messages, and constructs a real-time network state view by synthesizing multi-round sampling data.

[0083] Traffic Classification: The controller dynamically classifies large flows and small flows according to traffic statistics information; network flows with a traffic bandwidth request greater than or equal to 10% of the link bandwidth are determined as large flows, and all flows that do not reach the threshold are classified as small flows and forwarded through fast Equal-Cost Multi-Path (ECMP). When the access layer switch receives a network flow, the switch first classifies it into large flows and small flows according to the transmission rate of the network traffic, where the transmission rate is:

[0084]

[0085] In the formula, S represents the transmission rate of the network flow (unit: bytes / second), represents the cumulative data volume at time t2 (unit: bytes), represents the cumulative data volume at time t1 (unit: bytes), and t2 - t1 is the two sampling intervals (unit: seconds).

[0086] Optimization Solving: The PSODE engine receives large flow feature data, generates a path allocation scheme, and the optimization objective function minimizes the maximum link utilization, delay, and path diversity.

[0087] Rule Issuance: The optimized path is converted into OpenFlow flow table rules by the flow table management module and issued to the switch through the southbound interface. The controller sends an OFPT_FLOW_MOD message to the switch, sets the matching fields (such as source IP, destination IP, port number, etc.), and specifies the corresponding forwarding ports to ensure that large flows are transmitted according to the optimized path. At the same time, small flows maintain the ECMP algorithm rules and do not occupy additional computing resources.

[0088] Feedback Update: During the traffic scheduling process, the controller continuously obtains the link status through sFlow-RT and monitors the network congestion situation.

[0089] The fitness function is used to measure the quality of the large flow path allocation scheme of the PSODE algorithm. Considering network load balancing and quality of service, a multi-objective optimization function is designed, incorporating three indicators: the maximum link utilization (MLU), minimizing the end-to-end delay, and path diversity into the optimization scope. Among them, minimizing the maximum link utilization can reduce the risk of single-point congestion and improve the global anti-interference ability of the network, while minimizing the path delay can meet the quality of service requirements of real-time sensitive services such as video streams and financial transactions.

[0090] In this embodiment, minimizing the maximum link utilization includes: MLU represents the maximum value of the utilization of all links in the current network. If the utilization of a certain link is much higher than that of other links, it will cause serious congestion. The set of large flows to be scheduled includes:

[0091] f = {f1, f2, …, f s , …, f n}

[0092] In the formula, n represents the number of large flows, and f s is the sth large flow. Denote the total bandwidth capacity of the link as C ij . If a certain large flow passes through the link (i, j), it will occupy a part of its available bandwidth.

[0093] There are several feasible paths P = {P1, P2, …, P m} between the source host and the destination host, where P i represents the ith candidate path and is composed of a series of ordered links, and its expression is:

[0094] P i = {P i1 , P i2 , …, P ik}

[0095] In the formula, k represents the number of hops of the path.

[0096] If the currently used bandwidth of the jth hop link on the ith path is B ij , then the available bandwidth of this link is as shown in formula (3-15).

[0097] B′ ij = C ij - B ij

[0098] In the formula, B′ ij represents the available bandwidth of this link, and this information will provide a basic reference for subsequent path evaluation and selection.

[0099] MLU can be expressed as:

[0100]

[0101] In the formula, L is the set of all links in the network; is the currently used bandwidth of link l; represents the new bandwidth requirement for the large flow allocated to the path containing link l. If not allocated, it is 0; represents the total bandwidth capacity of link l.

[0102] Optimization objective: Minimize MLU to avoid single-point bottlenecks in high-load scenarios.

[0103] In this embodiment, minimizing the path delay includes: The path delay is the end-to-end transmission time of the data packet from the source node to the destination node, including the propagation delay and the queuing delay. The path delay can be expressed as:

[0104]

[0105] where P is the path for large flow distribution; is the propagation delay of link l; is the queuing delay of link l.

[0106] Optimization objective: Reduce the end-to-end delay for real-time sensitive services.

[0107] Path diversity is calculated as follows:

[0108]

[0109] where UniquePaths is the number of non-repeated paths in the particle swarm, TotalPaths is the total number of paths in the particle swarm, and σ 2 is the variance of the path length.

[0110] In the Fat-Tree network architecture, due to the inherent characteristics of multi-path connections, there are multiple alternative routing paths for communication between any two hosts. When the network controller obtains the complete topology information, for the data flow f s to be scheduled, the system randomly selects several candidate paths from its set of all reachable paths to form the initial solution space of the PSODE fusion algorithm. In this paper, the particle coding scheme is combined with path mapping: in a population composed of M particles, each particle represents a reachable transmission path, and its dimensional parameters correspond to the number of network forwarding nodes passed by this path. Through the dynamic optimization process of the PSODE algorithm, the controller continuously performs operations such as adjusting the particle velocity vector, updating the multi-dimensional space position, and differential evolution strategy, explores the global optimal solution of the solution space in the iterative calculation, and the finally obtained global optimal particle is the best scheduling path for the target flow f s .

[0111] The above-mentioned embodiments further elaborate on the purpose, technical solutions, and advantages of the present invention. It should be understood that the above-mentioned embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm, characterized in that, Including: The switch receives a data stream and determines whether the destination address of the data stream belongs to a local directly connected host. If so, the data stream is forwarded. Otherwise, it determines the size of the data stream. If it is a small data stream, the data stream is scheduled through the ECMP algorithm, the path is directly allocated, and the flow table is issued. If it is a large data stream, the PSODE algorithm is used to calculate the optimal path, the link load and delay are optimized according to the optimal path, and the path information is sent to the switch for scheduling.

2. The data center traffic scheduling method based on the differential evolution fusion particle swarm algorithm according to claim 1, characterized in that Determining the size of the data stream includes obtaining the link bandwidth. If the data stream size is greater than or equal to 10% of the link bandwidth, the data stream is a large data stream; otherwise, it is a small data stream.

3. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 1, characterized in that Scheduling the small data stream through the ECMP algorithm includes: S101: Obtain multiple equivalent paths to the destination address through the SDN controller, and the link metric values of each equivalent path are equal. S102: Extract the five-tuple features of the data stream, including the source IP address, destination IP address, source port number, destination port number, and transport layer protocol type. S103: Input the five-tuple features into a predefined hash function to generate a hash value, and determine the initial allocation path through modulo operation. S104: Generate a flow table entry including a matching field and an action instruction based on the initial allocation path, and send it to the flow table cache of the switch. S105: Forward the data stream according to the flow table entry, and trigger the update of the flow table entry when the path state deteriorates.

4. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 1, characterized in that Scheduling the large data stream through the PSODE algorithm includes: S101: Initialize the particle swarm, and the particle population is the large data stream. S102: Calculate the initial fitness of each particle in the particle swarm. S103: Update the optimal value of the particle and the global optimal value according to the initial fitness. S104: Update the velocity and position of the particle according to the optimal value of the particle and the global optimal value. S105: Calculate the fitness values of all particles in the population according to the velocity and position of the particle, and calculate the variance of the fitness values. Determine whether the DE algorithm trigger condition is satisfied according to the variance of the fitness values. If the trigger condition is not satisfied, determine whether the termination condition is satisfied. If not, return to step S102. If satisfied, output the global optimal solution. If the trigger condition is satisfied, execute step S106. S106: Perform DE operation on the particles with poor fitness. S107: Update the particle population and parameters, and return to step S102.

5. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 4, characterized in that The fitness value of the particle is calculated by comprehensively considering minimizing the maximum link utilization rate, minimizing the end-to-end delay, and path diversity. The expression is: fitness = α·f(MLU)+β·f(Delay)+γ·f(Diversity) Where α, β, and γ are weight factors, f(MLU) is the maximum link utilization rate, f(Delay) is the average end-to-end delay, and f(Diversity) is the path diversity.

6. The data center traffic scheduling method based on the differential evolution fusion particle swarm algorithm according to claim 5, wherein The weight factors are dynamically generated through a multi-objective game model. To improve the calculation efficiency and real-time performance, a lightweight approximate game strategy is used for dynamic weight adjustment. Specifically, it includes: S101. Define the maximum link utilization rate, end-to-end delay, and path diversity as three virtual players to participate in multi-objectives in scheduling optimization; S102. Calculate the contribution scores of each player based on the real-time network state; S103. Use the Nash bargaining solution with a penalty factor as the basis for theoretical weight allocation to obtain the initial weights; the formula for the penalty factor is: Among them, S i is the score of the i-th player, and S total is the total score of all players. S104. Dynamically adjust the weights using a lightweight approximate game strategy; S105. Set a threshold. When the variance of the fitness of the current population is greater than the set threshold, use the dynamic adjustment window mechanism to update the weights; otherwise, use the weight caching and reuse mechanism to process the weights to obtain the final weights.

7. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 6, characterized in that Dynamically adjusting the weights using a lightweight approximate game strategy includes: S1041. Calculate the difference in scores between adjacent generations of indicators, and its expression is: S1042. Based on the difference Δ i Generate weights using a Softmax function with a temperature coefficient, and its expression is: where τ = 0.7 is the temperature parameter, and n is the total number of optimization objectives.

8. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 6, characterized in that The period of the dynamic adjustment window mechanism is: Among them, is the variance of the fitness of the current population, is the historical maximum fitness variance, is for floor function.

9. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 4, characterized in that The DE algorithm triggers the condition to set the dynamic trigger period K, and this value is adaptively adjusted according to the variance of the population fitness, and its formula is: Among them, K base is the basic trigger period, K range is the trigger period adjustment range, σ 2 is the variance of the fitness of the current population, is the historical maximum variance, and round is the rounding function.

10. A data center traffic scheduling method based on a differential evolution fusion particle swarm algorithm according to claim 4, characterized in that, Performing DE operations on particles includes: adaptively adjusting the mutation factor, inertia weight, and crossover probability of the particles; performing mutation, crossover, and selection operations on the particles according to the adjusted mutation factor, inertia weight, and crossover probability to obtain a new particle swarm; where the inertia weight is: Among them, ω max is the maximum inertia weight, ω min is the minimum inertia weight, t is the current iteration number, and T max is the total number of iterations; The mutation factor is: Among them, F min is the minimum value of the mutation factor, F max is the maximum value of the mutation factor, f i is the current particle fitness, f best is the best fitness of the population, f worst is the worst fitness of the population, and ε is the minimum value; The crossover probability is: Among them, CR base is the initial crossover probability, and γ is the decay rate.

Images