Method and system for realizing load balancing in data center network

By dividing image flow and mouse flow in the data center network according to the transmission bandwidth, and using improved ant colony algorithm and detection packet information to select the optimal path, the problem that load balancing methods in the prior art fail to effectively distinguish and process different traffic is solved. , realizes efficient traffic transmission and network resource utilization.

CN118945116APending Publication Date: 2024-11-12SHANDONG COMP SCI CENTNAT SUPERCOMP CENT IN JINAN +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411212329.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The load balancing method in the existing data center network fails to effectively distinguish and process mouse flow from elephant flow, resulting in unbalanced resource utilization and network performance degradation, and cannot meet the transmission needs of different traffic.

Method used

By dividing the data flow into an elephant stream and a mouse flow according to the transmission bandwidth, using improved ant colony algorithm and detection packets to obtain bandwidth factors, elephant stream key factor and queue length index, selecting the optimal transmission path for elephant stream and mouse flow respectively. Implement intelligent scheduling of traffic.

Benefits of technology

It improves link bandwidth utilization, reduces delay, optimizes the utilization rate of network resources, ensures efficient transmission of large-stream elephant streams and small-stream mouse streams, and improves the overall performance and stability of the data center network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118945116A_ABST
    Figure CN118945116A_ABST
Patent Text Reader

Abstract

The invention provides a method and system for realizing load balancing in a data center network, and the method comprises the steps: dividing a data flow into an image flow and a mouse flow according to a transmission bandwidth; packaging the network load information of all reachable path nodes into a detection packet, and broadcasting the detection packet to the whole network; based on the probe packet information, acquiring a bandwidth factor, an image stream criticality factor and a queue length index, and distributing transmission paths for the image stream and the mouse stream; the method specifically comprises the following steps: carrying out improved ant colony algorithm iteration on an object flow, taking a bandwidth factor and an elephant flow criticality factor as the basis of ant colony path finding, obtaining and comparing the overhead of different paths of an optimal solution set, and forwarding the elephant flow to a path with the minimum overhead; and calculating the reachable path time delay of the mouse stream, taking the queue length index and the time delay as the basis of mouse stream routing, screening the path meeting the time delay limitation, and forwarding the mouse stream to the path with the minimum overhead. The flow is distributed to different paths according to the transmission requirements of the flow, so that the link bandwidth utilization rate is improved, and the time delay is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a method and system for implementing load balancing in a data center network. Background Art

[0002] With the rapid development of technologies such as cloud computing, big data, and artificial intelligence, the scale and complexity of data center networks (DCNs) continue to increase.

[0003] There are multiple paths in the data center network, and there are flows with various characteristics. The typical characteristic is the distribution of flow size. The data volume of small flows is usually less than 100KB, and the large flows are usually more than 10MB. More than 80% of small flows only occupy about 10% of the flow in the data center network, while less than 20% of large flows contribute more than 90% of the flow. Among them, large and small flows are often also called elephant and mouse flows. The size of the flow also determines different transmission requirements. This extremely unbalanced flow distribution brings challenges to the flow transmission of data centers.

[0004] Spine-leaf is the most common network architecture in data centers, which includes three layers: spine, leaf, and ToR, providing multiple transmission paths from end to end. Traditional data center load balancing methods usually do not distinguish between rat flows and elephant flows, resulting in unbalanced resource utilization and degraded network performance. Rat flows are numerous and sensitive to latency. When mixed with elephant flows for transmission, they are often affected by elephant flows, resulting in increased response time. However, elephant flows are prone to network bandwidth congestion due to their huge data volume, further affecting the performance and stability of the entire data center. The load balancing methods in the prior art have the following limitations:

[0005] (1) Without considering the characteristics of the traffic itself, a unified load balancing strategy is adopted for the traffic, which cannot meet the transmission requirements of different traffic.

[0006] (2) The network load condition is considered in a single way and various indicators cannot be fully considered to select appropriate forwarding paths for rat flows and elephant flows.

[0007] Therefore, how to effectively distinguish and handle rat flows and elephant flows to achieve efficient load balancing has become an important issue in current data center network optimization. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a method and system for implementing load balancing in a data center network, which effectively distributes traffic to different paths according to its own transmission requirements, improves link bandwidth utilization and reduces latency, and meets the transmission requirements of different traffic.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a method for implementing load balancing in a data center network, comprising:

[0011] Divide data streams into elephant streams and mouse streams according to transmission bandwidth;

[0012] Pack the network load information of all reachable path nodes into a detection packet and broadcast it to the entire network;

[0013] Based on the detection packet information, the bandwidth factor, elephant flow criticality factor and queue length index are obtained, and transmission paths are allocated for elephant flow and mouse flow. Specifically, the improved ant colony algorithm is iterated for the object flow, and the bandwidth factor and elephant flow criticality factor are used as the basis for ant colony pathfinding. The overheads of different paths in the optimal solution set are obtained and compared, and the elephant flow is forwarded to the path with the smallest overhead. The reachable path delay of the mouse flow is calculated, and the queue length index and delay are used as the basis for mouse flow routing. The path that meets the delay limit is screened, and the mouse flow is forwarded to the path with the smallest overhead.

[0014] Preferably, the data flow is divided into elephant flow and mouse flow according to the transmission bandwidth, specifically: the data flow exceeding 5% of the average remaining bandwidth of the link is marked as elephant flow, otherwise it is marked as mouse flow.

[0015] Preferably, the broadcasting rule of the detection packet is:

[0016] When a leaf switch receives a probe packet from a spine switch, it forwards the probe packet to all downstream ports. When it receives a probe packet from an edge switch, it forwards it to all ports except the receiving port.

[0017] The spine switch forwards the received probe packets to all ports except the receiving port;

[0018] After receiving the detection packet, the edge switch no longer forwards it.

[0019] Preferably, the improved ant colony algorithm is:

[0020]

[0021] Among them, η ij (t) is the bandwidth factor of the link between nodes i and j at time t, μ ij (t) is the criticality factor of the link between i and j, τ ij (t) is the pheromone factor of the link between nodes i and j, α, β, and γ are the weights of the bandwidth factor, the image flow criticality factor, and the pheromone factor, respectively, where α>β.

[0022] Preferably, the object flow is iterated by an improved ant colony algorithm, specifically including:

[0023] Setting a master ant colony and a slave ant colony; the master ant colony and the slave ant colony have the same number of ants and path finding probability mechanism;

[0024] In the main ant colony, the bandwidth factor and the criticality factor of the image flow The calculation method is:

[0025]

[0026] Among them, B bottle is the bottleneck bandwidth of the path, queue is the queue length of the switch, and U ij is the link bandwidth utilization;

[0027]

[0028] in, is the total bandwidth occupied by the elephant flow, E is the number of elephant flows, and BW is the capacity of the link. In the data center network, the link capacity BW is a constant. The larger it is, the lighter the link load is, and the greater the probability that the ant will choose j as the next node;

[0029] From the bandwidth factor of the ant colony and the criticality factor of the image flow The calculation method is:

[0030]

[0031]

[0032] Among them, X max is the maximum number of iterations, X is the current number of iterations, and E represents the number of elephant flows.

[0033] Preferably, the queue length index is:

[0034]

[0035] Among them, queue is the queue length of the switch when the detection packet leaves, and C is the state trend of the queue length. When the queue length trend is increasing, C<0; when the queue length trend is decreasing, C>0. Considering the change of queue length can enhance Q i Ability to adapt to changes in network status.

[0036] Preferably, the reachable path delay is calculated for the rat flow, the queue length index and the delay are used as the basis for the rat flow routing, the path that meets the delay limit is screened, and the rat flow is forwarded to the path with the minimum cost, specifically:

[0037] cost2(p)=α3·delay(p)+α4·Q(p)

[0038]

[0039] Among them, the average link delay avg As the delay limit, delay(p) is the node delay, N is the total number of nodes; Q(p) is the total number of nodes Q on path p. i , α3 and α4 are the coefficients of path delay and path queue index respectively.

[0040] In a second aspect, the present invention provides a system for implementing load balancing in a data center network, comprising:

[0041] Mouse-elephant flow division module, used for dividing the data flow into elephant flow and mouse flow according to the transmission bandwidth;

[0042] The load information acquisition module is used to package the network load information of all reachable path nodes into a detection packet and broadcast it to the entire network;

[0043] The load balancing module is used to obtain the bandwidth factor, elephant flow criticality factor and queue length index based on the detection packet information, and allocate transmission paths for elephant flow and mouse flow; specifically: the object flow is iterated by the improved ant colony algorithm, the bandwidth factor and elephant flow criticality factor are used as the basis for ant colony path finding, the overhead of different paths in the optimal solution set is obtained and compared, and the elephant flow is forwarded to the path with the smallest overhead; the reachable path delay is calculated for the mouse flow, the queue length index and delay are used as the basis for mouse flow routing, the path that meets the delay limit is screened, and the mouse flow is forwarded to the path with the smallest overhead.

[0044] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a method for implementing load balancing in a data center network described in the first aspect.

[0045] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for implementing load balancing in a data center network described in the first aspect are implemented.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. The implementation mechanism of load balancing in the data center network proposed in the present invention effectively improves the transmission quality and efficiency of the data center network through customized detection data packets and intelligent routing strategies. Different load balancing strategies are adopted according to the traffic characteristics (rat flow and elephant flow): the object flow uses the improved ant colony algorithm method, comprehensively considers bandwidth and delay, and selects the optimal path; for rat flow, the path that meets the delay limit is screened, and the queue length and delay are comprehensively considered, and the rat flow is forwarded to the path in the path set in sequence according to the size of the overhead. It can not only optimize the transmission path of large-flow elephant flow, but also ensure the low delay requirement of small-flow rat flow. It significantly improves the utilization rate of network resources, reduces the risk of network congestion, and provides a strong guarantee for the high-quality transmission of the data center network.

[0048] 2. The present invention forwards the detection data packet by broadcasting and follows specific rules to ensure effective coverage of all reachable paths in the network while avoiding path overlap. This design improves the efficiency and accuracy of network detection and provides a reliable data basis for subsequent traffic management and optimization. By finely controlling the forwarding path of the detection data packet, the network can more accurately grasp the actual status of each link, thereby making more reasonable resource allocation and path selection decisions, and improving overall network performance and user experience.

[0049] 3. In the elephant flow routing, the present invention significantly improves the intelligence and efficiency of network routing by integrating bandwidth factors, elephant flow criticality factors and master-slave ant colony strategies. By effectively utilizing global network status information (such as bandwidth utilization, elephant flow density) and combining the parallel search capability of the ant colony algorithm, adaptive optimization of complex network environments is achieved. The master ant colony responds quickly to network changes, while the slave ant colony balances exploration and utilization through iterative strategies to avoid premature convergence. The application of Tabu tables prevents circular paths and ensures the effectiveness of the paths. Overall, this module improves the real-time, reliability and efficiency of elephant flow transmission, and provides strong support for large-scale network traffic management.

[0050] 4. The present invention comprehensively considers the link bandwidth factor and the elephant flow criticality factor, so that the improved ant colony algorithm can allocate network resources more reasonably, avoid the performance degradation of some links due to overuse, and thus improve the utilization rate of overall network resources. By selecting links with large remaining bandwidth and light load as transmission paths, it helps to reduce the risk of network failures caused by link congestion or overload, and improve the stability and reliability of the network.

[0051] 5. In the mouse flow routing, the present invention provides an efficient and flexible multi-path transmission solution for the mouse flow by screening the paths that meet the delay constraints and comprehensively considering the queue length index and path delay. It not only ensures the delay requirements of the mouse flow transmission, but also realizes the intelligent scheduling of traffic by comparing the cost functions of different paths. Multi-path transmission not only improves the utilization rate of network resources, but also enhances the fault tolerance of the network, making the mouse flow transmission more reliable and efficient, which is of great significance for improving the overall performance of the network and user experience.

[0052] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.

[0054] Figure 1 A main flow chart of a method for implementing load balancing in a data center network provided by an embodiment of the present invention;

[0055] Figure 2 A detailed flow chart of a method for implementing load balancing in a data center network provided by an embodiment of the present invention;

[0056] Figure 3 A schematic diagram of the format of a detection data packet provided by an embodiment of the present invention;

[0057] Figure 4 A schematic diagram of a specific process of the improved ant colony algorithm provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0059] Embodiment 1

[0060] like Figure 1 As shown, this embodiment discloses a method for implementing load balancing in a data center network, comprising the following steps:

[0061] S1: Divide the data stream into elephant stream and mouse stream according to the transmission bandwidth;

[0062] S2: Pack the network load information of all reachable path nodes into a detection packet and broadcast it to the entire network;

[0063] S3: Based on the detection packet information, the bandwidth factor, elephant flow criticality factor and queue length index are obtained to allocate transmission paths for elephant flow and mouse flow. Specifically, the improved ant colony algorithm is iterated for the object flow, and the bandwidth factor and elephant flow criticality factor are used as the basis for ant colony path finding. The overheads of different paths in the optimal solution set are obtained and compared, and the elephant flow is forwarded to the path with the lowest overhead. The reachable path delay is calculated for the mouse flow, and the queue length index and delay are used as the basis for mouse flow routing. The path that meets the delay limit is screened, and the mouse flow is forwarded to the path with the lowest overhead.

[0064] Next, combine Figure 2 , a method for implementing load balancing in a data center network disclosed in this embodiment is described in detail.

[0065] In S1, the edge switch periodically calculates the data flow transmission bandwidth and uploads it to the controller. The controller calculates the average remaining bandwidth of the link. For flows that exceed 5% of the average remaining bandwidth of the link, they are marked as elephant flows, otherwise they are rat flows. Data flow transmission bandwidth B t The calculation formula is:

[0066]

[0067] Among them, Byte(t) is the total number of bytes of flow table entries at time t, Byte(tT s ) is tT s Total number of bytes in the flow table at time T s is the sampling duration.

[0068] In S2, the edge switch periodically sends a detection data packet, which can collect network load information of each link and switch it passes through. The detection data packet consists of IP, Ethernet header and a custom telemetry layer, including switch ID, link bandwidth factor η ij , the criticality factor μ ij and the queue length index Q i ,like Figure 3 shown.

[0069] The probe data packet is forwarded in the network in a broadcast manner. In order to cover all reachable paths and avoid path overlap, the forwarding rules of the probe data packet are formulated: when the leaf switch receives the probe data packet from the spine switch, it forwards the probe data packet to all downstream ports. When receiving the probe data packet from the edge switch, it forwards it to all ports except the receiving port; the spine switch forwards the received probe data packet to all ports except the receiving port; after receiving the probe data packet, the edge switch no longer forwards it.

[0070] In this embodiment, by formulating forwarding rules for detection data packets, it is ensured that all possible paths are detected and each path is detected only once, thereby avoiding redundancy and unnecessary network load. When a leaf switch receives a detection data packet from a spine switch, it forwards it to all downstream ports to cover all paths from the spine to the leaf; when it receives it from an edge switch, it forwards it to all ports except the receiving port, thereby avoiding loops and repeated detection of the same path. The spine switch forwards the received detection data packet to all ports except the receiving port, thereby covering the connection between the spine switches and avoiding repeated forwarding through the same spine switch. After receiving the detection data packet, the edge switch no longer forwards it and terminates the detection process in time to prevent the data packet from continuing to propagate outside the edge of the network, which helps to avoid unnecessary path overlap.

[0071] As a specific implementation method, after receiving a detection data packet, the non-edge switch first parses the telemetry layer, updates the link information between the previous hop switch and the local path information table, then adds the local data to the detection data packet, and adds the switch ID and input and output port numbers, and finally forwards the detection data packet according to the forwarding rules.

[0072] As a specific implementation method, after receiving the detection data packet, the edge switch parses the telemetry layer and obtains the bandwidth factor η according to the switch ID. ij and the criticality factor μ ij , update to the local path information table, and obtain a complete path based on the switch ID. After completing a round of forwarding of detection data packets, the edge switch will obtain all reachable paths to the remaining edge switches.

[0073] In this embodiment, all edge switches in the data center network will construct a detection data packet at a certain frequency. In order to reduce the network burden, the data packet is simplified and carries a customized telemetry layer. In addition to the basic second and third layers, the link bandwidth factor, elephant flow criticality factor and queue length index of each node are also collected. The edge switch sends the detection data packet to the leaf switch directly connected to the switch, and then according to the forwarding rules, the switch writes the information carried by the detection data packet into the local data table. These data are the main basis for the elephant flow and mouse flow routing, and the local data is added to the telemetry layer of the detection data packet until the detection data packet reaches another edge switch. In this way, the edge switch can obtain a complete path after parsing the switch_id field of the detection data packet. The improved ant colony algorithm for elephant flow and the delay screening algorithm for mouse flow will select a path set among these paths.

[0074] This embodiment collects network load information required for routing decisions by sending detection data packets into the network, and stores the information on the switch to provide necessary information for subsequent improved ant colony algorithm and rat flow routing algorithm.

[0075] In S3, based on the detection packet information, the bandwidth factor, the elephant flow criticality factor and the queue length index are obtained, and the transmission paths are allocated for the elephant flow and the mouse flow.

[0076] (1) Elephant flow routing

[0077] like Figure 4 As shown in the figure, an improved ant colony algorithm is used to select paths for elephant flows. The main feature is to set up two ant colonies, the master and the slave, to avoid premature convergence and local optimal solutions.

[0078] Considering that elephant flows have large data volumes and long durations, the main transmission requirement is high available bandwidth. Therefore, this embodiment designs two heuristic factors, using bandwidth factor and elephant flow criticality factor as the basis for ant colony path finding and for elephant flow path selection decision.

[0079] As a specific implementation method, considering that the link can reflect the carrying capacity of the path through the bandwidth, this embodiment sets the link bandwidth factor η ij As the main basis for elephant flow routing decisions; it is also considered that in a link, the more elephant flows there are, the less remaining available bandwidth, so the ratio of elephant flow load to link carrying capacity is used as the secondary basis for elephant flow routing decisions, which is called the elephant flow criticality factor μ ij , bandwidth factor and the criticality factor of the image flow Calculation method:

[0080]

[0081] Among them, B bottle is the bottleneck bandwidth of the path, queue is the queue length of the switch, and U ij is the link bandwidth utilization. The bottleneck link bandwidth and the switch queue length are proportional to the heuristic factor, and the link bandwidth utilization is inversely proportional to the heuristic factor.

[0082] In this embodiment, by introducing the link bandwidth factor in the path-finding process of the ant colony algorithm, ants are more inclined to select links with larger remaining bandwidth when selecting paths, thereby effectively avoiding selecting paths that are already congested or about to be congested.

[0083]

[0084] in, is the total bandwidth occupied by the elephant flow, E is the number of elephant flows, and BW is the capacity of the link. In data center networks, the link capacity BW is usually a constant. The larger it is, the lighter the link load is, and the greater the probability that the ant will choose j as the next node.

[0085] In this embodiment, when selecting a path, combining the elephant flow criticality factor can make the algorithm more inclined to select a link with a smaller number of elephant flows and a lighter load, which helps to achieve a balanced distribution of network traffic.

[0086] Before using the improved ant colony algorithm for iterative calculation, in order to reduce the algorithm calculation time, the links in the reachable path are distinguished from other links. First, initialize the link pheromone:

[0087]

[0088] Among them, τ(l) is the initial pheromone concentration of link l, M is the pheromone carried by each ant, Len(p) is the length of path p, and feasible_paths is the set of accessible paths. When a link does not belong to a accessible path, the initial pheromone concentration is set to 0.

[0089] Place N ants at the source edge switch and add the switch to the ant's taboo Tabu table. When selecting the next node, the ant can only choose a node that is not in the taboo Tabu table to prevent the ant from falling into a circular path. When selecting a path for the elephant flow, consider the elephant flow's demand for high bandwidth. Based on S2, the ant begins to iterate the path. The probability of ant a choosing node j at node i is as follows:

[0090]

[0091] Among them, η ij (t) is the bandwidth factor of the link between nodes i and j at time t, including the bandwidth factor of the main ant colony and the bandwidth factor from the ant colony μ ij (t) is the elephant flow criticality factor of the link between i and j, including the elephant flow criticality factor of the main ant colony and the criticality factor of the elephant flow from the ant colony τ ij (t) is the pheromone factor of the link between nodes i and j, α, β, and γ are the weights of the bandwidth factor, the image flow criticality factor, and the pheromone factor, respectively, where α>β.

[0092] Due to the limitations of the traditional ant colony algorithm, it is easy to fall into the local optimal solution. In order to avoid premature convergence, in addition to the main ant colony, a slave ant colony is also set up to prevent the main ant colony from falling into the local optimal solution.

[0093] The slave ant colony and the master ant colony have the same number of ants. The master and slave ant colonies have basically the same pathfinding probability mechanism and maintain their own independent network topology. The main difference is that the slave ant colony reduces the influence of the bandwidth factor and the elephant flow criticality factor at the beginning of the iteration, and emphasizes the influence of the pheromone concentration; as the number of iterations increases, the influence of the two factors gradually increases. Bandwidth factor of the slave ant colony and the criticality factor of the image flow The calculation method is as follows:

[0094]

[0095]

[0096] Among them, X max is the maximum number of iterations, X is the current number of iterations, and E represents the number of elephant flows.

[0097] In the routing process of the elephant flow, the master and slave ant colonies start to find the path at the same time as two independent pathfinding systems. The introduction of the slave ant colony can reduce the impact of the master ant colony falling into the local optimal solution to a certain extent. The reason is that the pathfinding mechanism of the slave ant colony can avoid falling into the local optimal solution, and the present invention finally obtains a unique solution by comparing the optimal solutions of the slave ant colony and the master ant colony. Therefore, if the master ant colony obtains the local optimal solution after iteration, the solution of the slave ant colony is selected.

[0098] This embodiment considers the link bandwidth factor and the elephant flow criticality factor when selecting a path, which helps to select a path with a smaller transmission delay, thereby improving the efficiency of data transmission. By optimizing the path selection mechanism, the improved ant colony algorithm can improve the throughput of the network, so that more data can be transmitted in a shorter time.

[0099] In this embodiment, all ants completing a path search are considered as one iteration. After each iteration, the ants add the node they are at to the taboo Tabu table. At the same time, all ants passing by will affect the pheromone concentration of the link, and the link pheromone needs to be updated. The pheromone formula of link ij at time t+1 is:

[0100] τ ij (t+1)=τ ij (t)·(1-ρ)+Δτ ij (t)

[0101] Among them, ρ is the pheromone volatilization factor, which determines the speed of pheromone volatilization, Δτ ij (t) is the pheromone increment brought to the link by all ants in this iteration.

[0102] The pheromone increment Δτ brought by the ants in this iteration ij The calculation formula of (t) is:

[0103]

[0104] Among them, Len a Indicates the length of the path the ant has walked in this iteration, cost ij represents the cost of link ij, Indicates the total cost of all links.

[0105] It can be seen from the formula that the pheromone increment Δτ brought by the ants in this iteration ij (t) is related not only to the pheromones carried by the ant itself and the length of the path it has walked, but also to the ratio of the link cost to the sum of all link costs. The smaller the ratio of link cost to total cost, the greater Δτ ij The larger (t) is, the greater the probability that the ant will choose this link in the subsequent iterations.

[0106] In this embodiment, the maximum number of hops of a single ant is set to X. max , X max It depends on the actual topology scale. Due to the special structure of the spine-leaf topology, the number of switches that the path between the same pod or different pods passes through is fixed. Paths exceeding this number are unlikely to be optimal solutions. After each iteration, determine whether the target switch is reached. If so, add the path to the optimal solution set. If not, determine whether the current iteration number of the ant colony has reached the maximum iteration number X. max .

[0107] After reaching the maximum number of iterations, the ant colony algorithm will obtain an optimal solution set that comprehensively considers the bandwidth factor, the criticality factor of the elephant flow, and the pheromone concentration. In order to select the final path, it is necessary to define the cost function of the path. The higher the path delay and the higher the bandwidth utilization, the lower the available bandwidth of the path, and the higher the path-finding cost of the ants. Therefore, the definition of the cost function is:

[0108] cost1(p)=α1·delay(p)+α2·U avg (p)

[0109] Where cost(p) is the cost function of path p, delay(p) is the path delay, and U avg (p) is the average link utilization of the path, and α1 and α2 are the coefficients of path delay and link utilization, respectively.

[0110] After obtaining the optimal solution set, first eliminate the paths that do not meet the maximum hop count limit in the optimal solution set, then compare the costs of the remaining paths, and use the path with the smallest cost as the forwarding path of the elephant flow.

[0111] (2) Rat flow routing

[0112] Since the characteristics of rat flow are high burstiness and sensitivity to delay, and elephant flow has a large amount of data and is sensitive to bandwidth, elephant flow has a higher tolerance to delay than rat flow. Therefore, this embodiment sacrifices part of the transmission performance of elephant flow and gives priority to rat flow. Find all paths that meet the delay limit among all reachable paths. This embodiment uses the average link delay as the delay limit:

[0113]

[0114] Where delay(p) is the node delay and N is the total number of nodes.

[0115] Taking advantage of the multi-path transmission of the spine-leaf topology, different paths are selected for different mouse flows. For all paths that meet the delay constraint, the switch selects the path as the forwarding path for the mouse flow in order of cost, and repeats the process. In this embodiment, the switch queue length and delay are used as the main basis for mouse flow routing, and the cost function is defined as:

[0116] cost2(p)=α3·delay(p))+α4·Q(p)

[0117] Where Q(p) is the number of nodes Q on path p. i , α3 and α4 are the coefficients of path delay and path queue index respectively.

[0118] This embodiment uses queue length as one of the bases for rat flow routing. Queue length reflects the number of data packets waiting to be processed in the switch, which directly affects the network delay. The queue length index Q i Definition:

[0119]

[0120] Among them, queue is the queue length of the switch when the detection packet leaves, and C is the state trend of the queue length. When the queue length trend is increasing, C<0; when the queue length trend is decreasing, C>0. Considering the change of queue length can enhance Q i Ability to adapt to changes in network status.

[0121] In this embodiment, the switch queue length index is directly related to the waiting time of the data packet in the switch and is an important indicator for evaluating the degree of network congestion. By incorporating it into the path cost function of the mouse flow, the algorithm can perceive the congestion hotspots in the network more real-time and accurately. In the path selection process, considering the switch queue length index can prompt the algorithm to select those paths with shorter queue lengths and stable change trends, thereby effectively reducing the waiting time of data packets in the network and reducing network latency.

[0122] This specific embodiment uses a custom detection data packet to be constructed and propagated in the network by broadcasting, providing the necessary network load status information for the switches in the network. The edge switch divides the flow into mouse flow and elephant flow, and adopts different load balancing strategies. The object flow uses the improved ant colony algorithm method to comprehensively consider bandwidth and delay to select the optimal path; for mouse flow, the path that meets the delay limit is screened, and the queue length and delay are comprehensively considered. The mouse flow is forwarded to the path in the path set in sequence according to the size of the overhead. By adopting different routing strategies according to the characteristics of different traffic, high-quality transmission is achieved.

[0123] Embodiment 2

[0124] This embodiment provides a system for implementing load balancing in a data center network, including:

[0125] Mouse-elephant flow division module, used for dividing the data flow into elephant flow and mouse flow according to the transmission bandwidth;

[0126] The load information acquisition module is used to package the network load information of all reachable path nodes into a detection packet and broadcast it to the entire network;

[0127] The load balancing module is used to obtain the bandwidth factor, elephant flow criticality factor and queue length index based on the detection packet information, and allocate transmission paths for elephant flow and mouse flow; specifically: the object flow is iterated by the improved ant colony algorithm, the bandwidth factor and elephant flow criticality factor are used as the basis for ant colony path finding, the overhead of different paths in the optimal solution set is obtained and compared, and the elephant flow is forwarded to the path with the smallest overhead; the reachable path delay is calculated for the mouse flow, the queue length index and delay are used as the basis for mouse flow routing, the path that meets the delay limit is screened, and the mouse flow is forwarded to the path with the smallest overhead.

[0128] Embodiment 3

[0129] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for implementing load balancing in a data center network as described in the first embodiment above are implemented.

[0130] Embodiment 4

[0131] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in a method for implementing load balancing in a data center network as described in the first embodiment above are implemented.

[0132] The steps or modules involved in the above embodiments 2 to 4 correspond to those in embodiment 1. For the specific implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for implementing load balancing in a data center network, characterized in that: include: Divide data streams into elephant streams and mouse streams according to transmission bandwidth; Pack the network load information of all reachable path nodes into a detection packet and broadcast it to the entire network; Based on the detection packet information, the bandwidth factor, elephant flow criticality factor and queue length index are obtained to allocate transmission paths for elephant flow and mouse flow; Specifically, the improved ant colony algorithm is iterated for the object flow, and the bandwidth factor and elephant flow criticality factor are used as the basis for ant colony path finding. The overhead of different paths in the optimal solution set is obtained and compared, and the elephant flow is forwarded to the path with the lowest overhead. The reachable path delay is calculated for the mouse flow, and the queue length index and delay are used as the basis for mouse flow routing. The path that meets the delay limit is screened, and the mouse flow is forwarded to the path with the lowest overhead.

2. A method for implementing load balancing in a data center network according to claim 1, characterized in that: The data flow is divided into elephant flow and mouse flow according to the transmission bandwidth, specifically: the data flow exceeding 5% of the average remaining bandwidth of the link is marked as elephant flow, otherwise it is marked as mouse flow.

3. The method for implementing load balancing in a data center network according to claim 1, characterized in that: The broadcasting rules of the detection packet are: When a leaf switch receives a probe packet from a spine switch, it forwards the probe packet to all downstream ports. When it receives a probe packet from an edge switch, it forwards it to all ports except the receiving port. The spine switch forwards the received probe packets to all ports except the receiving port; After receiving the detection packet, the edge switch no longer forwards it.

4. The method for implementing load balancing in a data center network according to claim 1, characterized in that: The improved ant colony algorithm is: Among them, η ij (t) is the bandwidth factor of the link between nodes i and j at time t, μ ij (t) is the criticality factor of the link between i and j, τ ij (t) is the pheromone factor of the link between nodes i and j, α, β, and γ are the weights of the bandwidth factor, the image flow criticality factor, and the pheromone factor, respectively, where α>β.

5. The method for implementing load balancing in a data center network according to claim 1, characterized in that: The object flow is iterated by the improved ant colony algorithm, specifically including: Setting a master ant colony and a slave ant colony; the master ant colony and the slave ant colony have the same number of ants and path finding probability mechanism; In the main ant colony, the bandwidth factor and the criticality factor of the image flow The calculation method is: Among them, B bottle is the bottleneck bandwidth of the path, queue is the queue length of the switch, and U ij is the link bandwidth utilization; in, is the total bandwidth occupied by the elephant flow, E is the number of elephant flows, and BW is the capacity of the link. In the data center network, the link capacity BW is a constant. The larger it is, the lighter the link load is, and the greater the probability that the ant will choose j as the next node; From the bandwidth factor of the ant colony and the criticality factor of the image flow The calculation method is: Among them, X max is the maximum number of iterations, X is the current number of iterations, and E represents the number of elephant flows.

6. The method for implementing load balancing in a data center network according to claim 1, characterized in that: The queue length index is: Among them, queue is the queue length of the switch when the detection packet leaves, and C is the state trend of the queue length. When the queue length trend is increasing, C<0; when the queue length trend is decreasing, C>0. Considering the change of queue length can enhance Q i Ability to adapt to changes in network status.

7. The method for implementing load balancing in a data center network according to claim 1, characterized in that: The method of calculating the reachable path delay for the rat flow, taking the queue length index and the delay as the basis for the rat flow routing, screening the path that meets the delay limit, and forwarding the rat flow to the path with the smallest cost is as follows: cost2(p)=α3·delay(p)+α4·Q(p) Among them, the average link delay avg As the delay limit, delay(p) is the node delay, N is the total number of nodes; Q(p) is the total number of nodes Q on path p. i , α3 and α4 are the coefficients of path delay and path queue index respectively.

8. A system for implementing load balancing in a data center network, characterized in that: include: Mouse-elephant flow division module, used for dividing the data flow into elephant flow and mouse flow according to the transmission bandwidth; The load information acquisition module is used to package the network load information of all reachable path nodes into a detection packet and broadcast it to the entire network; A load balancing module is used to obtain bandwidth factor, elephant flow criticality factor and queue length index based on the detection packet information, and allocate transmission paths for elephant flow and mouse flow; Specifically, the improved ant colony algorithm is iterated for the object flow, and the bandwidth factor and elephant flow criticality factor are used as the basis for ant colony path finding. The overhead of different paths in the optimal solution set is obtained and compared, and the elephant flow is forwarded to the path with the lowest overhead. The reachable path delay is calculated for the mouse flow, and the queue length index and delay are used as the basis for mouse flow routing. The path that meets the delay limit is screened, and the mouse flow is forwarded to the path with the lowest overhead.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a method for implementing load balancing in a data center network as described in any one of claims 1-7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the processor implements the steps in a method for implementing load balancing in a data center network as described in any one of claims 1-7.

Citation Information

Cited By

  • Dynamic path optimization method and system for cross-device file transmission

    CN120434239A

  • Dynamic path optimization method and system for cross-device file transfer

    CN120434239B