A dynamic load balancing method suitable for a weak network environment

Abstract

This invention discloses a dynamic load balancing method suitable for weak network environments, belonging to the field of network load balancing technology. The method includes the following steps: Step 1: Perform multi-dimensional link status collection in the network, monitor the key performance indicators of the physical and transport layers of each candidate path in real time, and calculate a comprehensive metric for each path; Step 2: Dynamically filter out a set of reliable paths for load balancing; Step 3: Dynamically calculate and adjust the traffic allocation weights of each path, including dynamic weight calculation based on inverse performance ratio, adaptive adjustment of sensitivity index, and weight fine-tuning based on real-time congestion; Step 4: Perform stability control and oscillation suppression; Step 5: Identify different upper-layer applications, and implement traffic classification and differentiated policy execution for different upper-layer application requirements.

Description

Technical Field

[0001] This invention belongs to the field of network load balancing technology, specifically relating to a dynamic load balancing method suitable for weak network environments. Background Technology

[0002] With the ubiquitous development of information and communication technologies, network application scenarios have expanded from stable, high-speed wired environments to complex, weak network environments characterized by high latency, high packet loss, bandwidth asymmetry, and dynamic topology changes, such as satellite communication, mobile ad hoc networks, remote wireless access, and emergency communication. In these environments, traditional multipath transmission and load balancing technologies face severe challenges.

[0003] Existing routing protocols that support multi-path load balancing, such as the Enhanced Interior Gateway Routing Protocol (EINS), support both equi-cost and unequal-cost multi-path load balancing modes. Their core mechanism is based on a composite metric (considering bandwidth, latency, etc.) and controls the traffic allocation ratio of unequal-cost paths through parameters. EINS guarantees a loop-free topology through a distributed update algorithm, and its load balancing decisions are essentially based on relatively static link metrics.

[0004] Existing load balancing mechanisms, especially equivalent or multi-path routing based on traditional routing protocols, were originally designed primarily for relatively stable enterprise or data center networks. They typically rely on static or slowly changing link metrics (such as bandwidth and hop count) for path selection and traffic allocation. In weak network environments, these mechanisms exhibit significant limitations: First, their static metrics cannot accurately reflect the real-time transmission efficiency of the links (such as instantaneous throughput, queuing latency, and packet loss rate), leading to traffic potentially being continuously injected into congested or degraded paths; second, their fixed allocation strategies or switching thresholds are difficult to adapt to drastic link fluctuations, easily causing routing oscillations and traffic jitter, which in turn reduces overall performance; finally, they lack awareness of the differentiated needs of upper-layer applications and cannot provide differentiated path guarantees for latency-sensitive services such as voice and video versus bandwidth-sensitive services such as file transfer. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the present invention aims to provide a dynamic load balancing method suitable for weak network environments, in order to solve the technical problems of existing load balancing technologies in weak network environments, such as lagging link quality perception, rigid traffic allocation strategies, frequent path switching causing oscillations, and lack of service differentiation guarantees.

[0006] To achieve the above objectives, the present invention may adopt the following specific technical solutions: The aforementioned dynamic load balancing method suitable for weak network environments includes the following steps: Step 1: Collect multi-dimensional link status data in the network, monitor the key performance indicators of the physical layer and transport layer of each candidate path in real time, and calculate a comprehensive metric for each path. Step 2: Dynamically filter out a set of reliable paths for load balancing; Step 3: Dynamically calculate and adjust the traffic allocation weights for each path, including dynamic weight calculation based on inverse performance ratio, adaptive adjustment of sensitivity index, and weight fine-tuning based on real-time congestion. Step 4: Perform stability control and oscillation suppression; Step 5: Identify different upper-layer applications and implement traffic classification and differentiated strategy execution based on the different needs of upper-layer applications.

[0007] Furthermore, in step 2, the steps for dynamically selecting reliable paths are as follows: Step 2.1: Maintain a state machine for each candidate path. For link i, it includes the following core states: Dormant state: The path has not been evaluated or has been moved out of the working set and does not carry user data traffic; Probe Status: The path is being actively or passively probed, collecting performance data to calculate initial integrated metrics. ; Qualified status: Path It remained stable at an acceptable level and was officially included in the dynamic qualified path set to participate in traffic allocation. Removal status: The path performance continues to deteriorate and it is being removed from the working set; Step 2.2: Initial comprehensive metric value for a dormant path. A path must consistently and stably outperform the admission criteria for a period of time before it can migrate to a qualified state; a path in a qualified state, when the initial comprehensive metric value... If a value momentarily falls short of the admission criteria, it will not be immediately removed; it will only be removed if the initial comprehensive metric value is lower than the admission criteria. A path will only be downgraded or removed if it consistently performs worse than the admission criteria for a period of time. Downgrading is more conservative than admission, preventing high-quality paths from being mistakenly removed due to momentary fluctuations.

[0008] Furthermore, in step 3, the dynamic weight calculation based on inverse performance ratio specifically involves: first, calculating basic weights based on the dynamic transmission efficiency metric of the path; then, designing a metric dominance using an inverse proportional relationship. And introduce a dynamic sensitivity index. To amplify or reduce the impact of quality differences.

[0009] Furthermore, in step 3, the specific calculation of the adaptive adjustment of the sensitivity index is as follows: ① Calculate the dispersion index using the coefficient of variation The larger the value, the greater the difference in quality between the paths; ② Regarding the dynamic sensitivity index The mapping rule uses a smooth function with a saturation region to... Mapped to Between intervals; among which, This is the lower limit of the dynamic sensitivity index, for example, 0.2; This is the upper limit of the dynamic sensitivity index, for example, 4; ③ Two threshold values ​​are set: when At that time, it was assumed that the path differences were small. Approaching The distribution tends to be uniform; when At that time, it was believed that the paths differed greatly. Approaching The allocation is highly concentrated on the optimal path; among them, The value range is 0.15 ~ 0.25. The value range is 0.60 ~ 0.80. The setting standard should take into account the statistical data of actual measurement in weak network environment to ensure that the threshold covers the quality distribution range of typical weak network scenarios. ④ Choose to use piecewise linear mapping. At the same time, assigning weights and measuring dominance. Proportional to the standard proportion, ensuring fair distribution; At the same time, amplify the advantages of high-quality paths; This weakens the advantage of high-quality paths.

[0010] Furthermore, in step 3, the specific calculation for weight fine-tuning based on real-time congestion is as follows: The exit queue for each path defines two thresholds: a warning threshold and a warning threshold. With overload threshold Real-time measurement of queue occupancy rate Calculate the penalty factor : , in, As the starting point for response, when the queue occupancy rate is lower than this value, the link is considered to be in an idle or healthy state. As a severely congested line, when the queue occupancy rate reaches or exceeds this value, the link is considered to be in a dangerous state, and the maximum penalty is applied. ; Defined as a penalty intensity coefficient; finally, the weights are applied. .

[0011] Furthermore, in step 4, an adaptive exponentially weighted moving average filter is introduced, utilizing the path metric coefficient of variation. The rate of change of weights is used to characterize network volatility; a hard limit is imposed on the smoothed rate of change of weights, and the maximum allowable weight change in a single period is set to approximate the target weight through multiple subsequent periods; in the selection of migration flows, the most suitable connection for migration is selected from the path where the weight decreases; the selection criteria include, but are not limited to, connection age, traffic size, or application type.

[0012] Furthermore, in step 5, the corresponding strategy parameter set is invoked based on the business type to which the data stream belongs, to perform path filtering and weight calculation: For real-time interactive services, low-latency, low-jitter paths should be adopted, and multi-path routing and frequent switching should be restricted. (Sensitivity coefficient) Larger, for example, 2.5; For streaming media services, using high-bandwidth paths and allowing limited parallel transmission enables traffic to be more freely split and aggregated across multiple paths to maximize overall bandwidth utilization. This approach offers higher tolerance for path switching and a lower sensitivity coefficient. Smaller, for example, 0.8; For reliable transmission services, a low packet loss rate path is used.

[0013] Compared with the prior art, the present invention has the following advantages: (1) This invention proposes dynamic intelligent weight allocation. Through dynamic weights with inverse efficiency, adaptive adjustment of sensitivity index and weight fine-tuning based on real-time congestion, it can perceive and adapt to the drastic fluctuations in the quality of weak network links in real time, intelligently guide traffic to the optimal path, and significantly improve the overall throughput and transmission efficiency of multi-link networks in weak network environments.

[0014] (2) The present invention can effectively suppress frequent path switching and traffic fluctuation caused by network jitter through weighted smoothing filtering, thus ensuring the smoothness of decision-making and overall stability in weak network environment.

[0015] (3) This invention can provide differentiated and customized path selection and resource guarantee strategies for different upper-layer services such as real-time interaction and streaming media, and has great potential for promotion. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Example 1 like Figure 1 As shown, a dynamic load balancing method suitable for weak network environments includes the following steps: (I) Step 1: Collect multi-dimensional link status data in the network, monitor the key performance indicators of the physical and transport layers of each candidate path in real time, and calculate a comprehensive metric for each path. Where bandwith represents bandwidth and Delay represents latency.

[0019] Key performance indicators include, but are not limited to, effective throughput, end-to-end latency and jitter, packet loss rate, and egress queue congestion.

[0020] (ii) Step 2: Dynamically select a set of reliable paths for load balancing.

[0021] Maintain a state machine for each candidate path. For link i, it includes a sleep state, a probe state, a qualified state, and a removal state.

[0022] Dormant state: The path has not been evaluated or has been moved out of the working set and does not carry user data traffic.

[0023] Probe Status: The path is being actively or passively probed, collecting performance data to calculate initial integrated metrics. .

[0024] Acceptable status: Initial comprehensive metric value of the path It remained stable at an acceptable level and was officially included in the dynamic qualified path set to participate in traffic allocation.

[0025] Removal status: The path performance continues to deteriorate and it is being removed from the working set.

[0026] The initial comprehensive metric of a dormant path. A path must consistently and stably outperform the admission criteria for a period of time before it can migrate to a qualified state; a path in a qualified state, when its initial comprehensive metric value... If a value momentarily falls short of the admission criteria, it will not be immediately removed; it will only be removed if the initial comprehensive metric value is lower than the admission criteria. Only after consistently falling below the admission threshold for a period of time will an application be downgraded or removed.

[0027] (III) Step 3: Dynamically calculate and adjust the traffic allocation weights of each path, including dynamic weight calculation based on inverse efficiency ratio, adaptive adjustment of sensitivity index, and weight fine-tuning based on real-time congestion.

[0028] (1) Calculation of dynamic weights based on inverse performance ratio Basic weights are calculated based on a dynamic transmission efficiency metric for the path (smaller values ​​indicate better path quality). A dominance metric is designed using an inverse proportional relationship. And introduce a dynamic sensitivity index. To amplify or reduce the impact of quality differences: , , in, For the basic weights of the path, For path At any moment The dynamic transmission efficiency metric value is calculated by step 1 for this path. As a smoothing factor, set it to a very small positive constant (e.g., ), mainly used to prevent the metric value of a certain path from being... When the value approaches zero, the weight calculation may result in a division by zero error or unstable value; however, the choice of its value does not affect the relative proportion of the weight. It is a dynamic sensitivity index that determines how sensitive traffic allocation is to performance differences between paths.

[0029] (2) Dynamic sensitivity index Adaptive adjustment calculation The dispersion index is calculated using the coefficient of variation. , , , in, It is a dimensionless indicator; the larger the value, the greater the quality difference between the different paths.

[0030] For dynamic sensitivity index The mapping rule uses a smooth function with a saturation region to... Mapped to Between intervals.

[0031] , in, It is a saturation function of the Sigmoid type. Two thresholds are set: when... At that time, it was assumed that the path differences were small. Approaching The distribution tends to be uniform; when At that time, it was believed that the paths differed greatly. Approaching The allocation is highly concentrated on the optimal path. Specifically, a piecewise linear mapping can be used: , The value determines how sensitive traffic allocation is to performance differences between paths. When assigning weights and It is proportional to the standard proportion for fair distribution. At this time, the advantages of high-quality paths are amplified. When there are significant performance differences between paths, the system will highly concentrate traffic on a few optimal paths, which avoids the use of inferior links in weak networks. This approach weakens the advantage of high-quality paths, making the allocation more even, and is suitable for scenarios where all paths are of high quality and where maximizing parallelism is required.

[0032] (3) Calculation of weight fine-tuning based on real-time congestion The exit queue for each path defines two thresholds: a warning threshold and a warning threshold. With overload threshold Real-time measurement of queue occupancy rate Punishment factor The specific calculations are as follows: , in, Defined as the penalty intensity coefficient, As the starting point for response, when the queue occupancy rate is lower than this value, the link is considered to be in an idle or healthy state. As a severely congested link, when the queue occupancy rate reaches or exceeds this value, the link is considered to be in a dangerous state of high congestion, and the maximum penalty is applied at this time. .

[0033] The final application weights are calculated as follows: ; This fine-tuning mechanism can react to impending congestion by rapidly reducing the traffic weight of the corresponding path and redirecting some traffic to paths with idle queues, thus achieving proactive and preventative congestion avoidance.

[0034] (iv) Step 4: Perform stability control and oscillation suppression to ensure the smoothness of decision-making and system stability of the above dynamic mechanism under the severe fluctuations of weak network.

[0035] Weights calculated directly from the adaptive algorithm It contains instantaneous information about path quality, as well as measurement noise and short-term bursts of interference. Direct application of this information would lead to frequent changes in scheduling strategies. This application introduces an adaptive exponentially weighted moving average filter: , in, It is an adaptive smoothing factor, and its value is in between. The closer it is to 1, the greater the inertia, the stronger the dependence on historical weights, and the slower the response to the current system.

[0036] The smoothing factor should not be a fixed value, but should be dynamically adjusted according to the stability of the network state; therefore, the path metric coefficient of variation is used. The rate of change is used to characterize network volatility: , , in, The base smoothing factor (e.g., 0.3) is the lower bound when the network is extremely unstable, ensuring that the system still retains a minimum level of responsiveness. This is the maximum smoothing factor (e.g., 0.8). It is used when the network is very stable, giving the system strong inertia and filtering out small fluctuations. The attenuation coefficient controls the rate at which α(t) decreases as ΔCV(t) increases.

[0037] To further prevent excessively large adjustments in a single instance, a hard limit is imposed on the smoothed weight change rate, setting a maximum allowable weight change per cycle, and the target weight is approximated over multiple subsequent cycles. In selecting migration flows, the most suitable connection for migration is chosen from the paths where weights decrease. Selection criteria can include connection age, traffic volume, or application type.

[0038] (v) Step 5: Identify different upper-layer applications and implement traffic classification and differentiated strategy execution based on the different needs of upper-layer applications.

[0039] Based on the business type of the data stream, the corresponding set of strategy parameters is invoked to perform path filtering and weight calculation. For example, for real-time interactive businesses, low-latency, low-jitter paths are prioritized, and multi-path routing and frequent switching are restricted, with a sensitivity coefficient... The sensitivity coefficient is relatively high; for streaming media services, high-bandwidth paths are prioritized and limited parallel transmission is allowed. Traffic can be more freely split and aggregated across multiple paths to maximize overall bandwidth utilization. This approach has a higher tolerance for path switching and a lower sensitivity coefficient. For services requiring reliable transmission, the path with the lowest packet loss rate should be selected first.

[0040] Example 2 In this embodiment, the radio is equipped with a satellite link (high latency, fluctuating bandwidth) and a 4G / 5G wireless public network link (high bandwidth but severe latency jitter and intermittent congestion). The two links serve as backups for each other and are activated simultaneously, forming a heterogeneous weak network multipath environment. The radio simultaneously runs two types of services: Service A, real-time video conferencing, which is extremely sensitive to latency and jitter and requires moderate bandwidth; Service B, sensor data backhaul, which requires high reliability (low packet loss) and allows for a certain degree of latency.

[0041] Step 1: Conduct multi-dimensional link status collection in the network, and monitor key performance indicators of the physical and transport layers of each candidate path in real time, including but not limited to effective throughput, end-to-end latency and jitter, packet loss rate, and egress queue congestion level, and calculate M. A =0.5, M B =0.7.

[0042] Step 2: Dynamically select a set of reliable paths for load balancing. Both paths have health scores higher than the preset minimum qualification threshold of 0.4 and are selected into the qualified path set, but traffic allocation is expected to be biased towards 4G / 5G.

[0043] Step 3: Dynamically calculate and adjust the traffic allocation weights for each path, including dynamic weight calculation based on inverse performance ratio, adaptive adjustment of sensitivity index, and weight fine-tuning based on real-time congestion.

[0044] (1) Calculation of dynamic weights based on inverse performance ratio Basic weights are calculated based on a dynamic transmission efficiency metric for the path (smaller values ​​indicate better path quality). A dominance metric is designed using an inverse proportional relationship. And introduce a dynamic sensitivity index. To amplify or reduce the impact of quality differences: , , in, For the basic weights of the path, For path At any moment The dynamic transmission efficiency metric value is calculated by step 1 for this path. As a smoothing factor, set it to a very small positive constant (e.g., ), mainly used to prevent the metric value of a certain path from being... When the value approaches zero, the weight calculation may result in a division by zero error or unstable value; however, the choice of its value does not affect the relative proportion of the weight. It is a dynamic sensitivity index that determines how sensitive traffic allocation is to performance differences between paths.

[0045] (2) Dynamic sensitivity index Adaptive adjustment calculation The dispersion index is calculated using the coefficient of variation. , , , in, It is a dimensionless indicator; the larger the value, the greater the quality difference between the different paths.

[0046] For dynamic sensitivity index The mapping rule uses a smooth function with a saturation region to... Mapped to Between intervals.

[0047] , in, It is a saturation function of the Sigmoid type. Two thresholds are set: when... At that time, it was assumed that the path differences were small. Approaching The distribution tends to be uniform; when At that time, it was believed that the paths differed greatly. Approaching The allocation is highly concentrated on the optimal path. Among them, The lower limit of the dynamic sensitivity index is 0.2; The upper limit of the dynamic sensitivity index is 4; specifically, a piecewise linear mapping can be used: , The value determines how sensitive traffic allocation is to performance differences between paths. The coefficient of variation (COP) of the health scores of the two paths is calculated; assuming the COP is 0.5, the current COP is obtained through a Sigmoid mapping. =2, which means the system determines that the two paths have a significant quality difference and tends to use a less balanced allocation to avoid using links with poor dynamic stability. Substituting this back into the basic weight calculation, the basic weight W for business A is... A =14%, Business B's basic weight W B =86%.

[0048] (3) Calculation of weight fine-tuning based on real-time congestion The exit queue for each path defines two thresholds: a warning threshold and a warning threshold. With overload threshold Real-time measurement of queue occupancy rate Punishment factor The specific calculations are as follows: , The monitoring showed that the instantaneous queue occupancy rate of the 4G / 5G path was 85%, which is greater than... This triggers a congestion penalty, applying a penalty factor. The final weight calculation is as follows: ; =45%, Due to momentary congestion, the traffic allocation of 4G / 5G was significantly reduced, and satellite links took on more traffic, achieving rapid congestion avoidance.

[0049] (iv) Step 4: Perform stability control and oscillation suppression to ensure the smoothness of decision-making and system stability of the above dynamic mechanism under the severe fluctuations of weak network.

[0050] Weights calculated directly from the adaptive algorithm It contains instantaneous information about path quality, as well as measurement noise and short-term bursts of interference. Direct application of this information would lead to frequent changes in scheduling strategies. This application introduces an adaptive exponentially weighted moving average filter: , in, It is an adaptive smoothing factor, and its value is in between. The closer it is to 1, the greater the inertia, the stronger the dependence on historical weights, and the slower the response to the current system.

[0051] The smoothing factor should not be a fixed value, but should be dynamically adjusted according to the stability of the network state; therefore, the path metric coefficient of variation is used. The rate of change is used to characterize network volatility: , , in, The base smoothing factor (e.g., 0.3) is the lower bound when the network is extremely unstable, ensuring that the system still retains a minimum level of responsiveness. This is the maximum smoothing factor (e.g., 0.8). It is used when the network is very stable, giving the system strong inertia and filtering out small fluctuations. The attenuation coefficient controls the rate at which α(t) decreases as ΔCV(t) increases.

[0052] To further prevent excessively large single adjustments, a hard limit is imposed on the smoothed weight change rate, setting a maximum allowable weight change per cycle, and approximating the target weight over multiple subsequent cycles. Traffic switching from 4G / 5G to satellite does not exceed 20% of its total traffic. This avoids connection resets or packet out-of-order issues caused by sudden weight changes.

[0053] (V) Step 5: Identify different upper-layer applications and implement traffic classification and differentiated policy execution based on the different needs of these applications. Service A (Video Conferencing): Enable a high-stability policy. Even though the final weight of the satellite link is 45%, to avoid out-of-order transmission across paths, the system will still bind all traffic of Service A to the relatively lower latency and uncongested 4G / 5G path (although its weight is penalized, there is still sufficient bandwidth to carry high-priority video streams). Service B (Data Backhaul): Enable a high-reliability policy and allocate its traffic according to the final weight (satellite 45%, 4G / 5G 55%).

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic load balancing method suitable for a weak network environment, characterized by, Includes the following steps: Step 1: Collect multi-dimensional link status data in the network, monitor the key performance indicators of the physical layer and transport layer of each candidate path in real time, and calculate a comprehensive metric M for each path. Step 2: Dynamically filter out a set of reliable paths for load balancing; Step 3: Dynamically calculate and adjust the traffic allocation weights for each path, including dynamic weight calculation based on inverse performance ratio, adaptive adjustment of sensitivity index, and weight fine-tuning based on real-time congestion. Step 4: Perform stability control and oscillation suppression; Step 5: Identify different upper-layer applications and implement traffic classification and differentiated strategy execution based on the different needs of upper-layer applications.

2. The dynamic load balancing method for weak network environment according to claim 1, wherein, In step 2, the steps for dynamically selecting reliable paths are as follows: Step 2.1: Maintain a state machine for each candidate path. For link i, the state machine includes the following states: Dormant state: The path has not been evaluated or has been moved out of the working set and does not carry user data traffic; Probe status: Path is being actively or passively probed, performance data is being collected to compute initial composite metric values ; Qualified state: the path is Continuously stable at an acceptable level, officially included in the dynamic qualified path set, and participates in traffic distribution; Removal status: The path performance continues to deteriorate and it is being removed from the working set; Step 2.2, a path in dormant state, initial integrated metric value Must be continuously and steadily better than the admission 300ms~2.5s to migrate to eligible state; a path in eligible state, when the initial integrated metric value Is transiently worse than the admission, it will not be removed immediately; only when the initial integrated metric value Is continuously worse than the admission 500ms~5s, it will be degraded or removed.

3. The dynamic load balancing method suitable for weak network environments according to claim 1, characterized in that, In step 3, the dynamic weight calculation based on inverse performance ratio specifically involves: first, calculating the basic weights based on the dynamic transmission efficiency metric of the path; then, designing the metric dominance using the inverse proportional relationship. And introduce a dynamic sensitivity index. To amplify or reduce the impact of quality differences, and to calculate the basic path weights. .

4. The dynamic load balancing method suitable for weak network environments according to claim 3, characterized in that, In step 3, the specific calculation for the adaptive adjustment of the sensitivity index is as follows: ① Calculate the dispersion index using the coefficient of variation The larger the value, the greater the difference in quality between the paths; ② Regarding the dynamic sensitivity index The mapping rule uses a smooth function with a saturation region to map the dispersion index. Mapped to Between intervals, among which, It is the lower limit of the dynamic sensitivity index; This is the upper limit of the dynamic sensitivity index; ③ Two threshold values ​​are set: when At that time, it was assumed that the path differences were small. Approaching The distribution tends to be uniform; when At that time, it was believed that the paths differed greatly. Approaching The allocation is highly concentrated on the optimal path; ④ Choose to use piecewise linear mapping. At the same time, assigning weights and measuring dominance. Proportional to the standard proportion, ensuring fair distribution; At the same time, amplify the advantages of high-quality paths; This weakens the advantage of high-quality paths.

5. A dynamic load balancing method suitable for weak network environments according to claim 4, characterized in that, In step 3, the specific calculation for weight fine-tuning based on real-time congestion is as follows: The exit queue for each path defines two thresholds: a warning threshold and a warning threshold. With overload threshold Real-time measurement of queue occupancy rate Calculate the penalty factor : ， in, As the starting point for response, when the queue occupancy rate is lower than this value, the link is considered to be in an idle or healthy state. As a severely congested line, when the queue occupancy rate reaches or exceeds this value, the link is considered to be in a dangerous state, and the maximum penalty is applied. ; Defined as a penalty intensity coefficient; finally, the weights are applied. .

6. The dynamic load balancing method suitable for weak network environments according to claim 1, characterized in that, In step 4, an adaptive exponentially weighted moving average filter is introduced, and the path metric coefficient of variation is used. The rate of change of weights is used to characterize network volatility; a hard limit is imposed on the smoothed rate of change of weights, and the maximum allowable weight change in a single period is set to approximate the target weight through multiple subsequent periods; in the selection of migration flows, the most suitable connection for migration is selected from the path where the weight decreases; the selection criteria include, but are not limited to, connection age, traffic size, or application type.

7. A dynamic load balancing method suitable for weak network environments according to claim 1, characterized in that, In step 5, the corresponding strategy parameter set is called according to the service type of the data stream to perform path filtering and weight calculation: for real-time interactive services, low-latency and low-jitter paths are used and multi-path splitting and frequent switching are restricted; for streaming media services, high-bandwidth paths are used and limited parallel transmission is allowed, allowing traffic to be split and aggregated more freely on multiple paths to maximize the total bandwidth utilization; for reliable transmission services, low packet loss paths are used.

Images