A server load balancing method based on distributed computing

Abstract

The application discloses a kind of server load balancing methods based on distributed computing, comprising the following steps: in each entrance load balancer, the instantaneous or short window statistical data of server is collected to generate state vector and divide service category;Based on state vector, under the improved distributed average field game model, the objective function including congestion cost, migration cost and risk sensitive cost is constructed;Average field distribution is established, and the feasible routing set is determined under the capacity, quota and affinity constraint;Each entrance is based on optimal response and uses mirror outside ladder two-step update;In the updating process, introduce time attenuation and maximum protection, combine temperature decreasing disturbance and cold start release mechanism, finally according to the change rate and default rate, determine convergence and output routing distribution.The application introduces improved distributed average field game model, realizes the load balancing of high-efficiency and stable service-oriented level agreement in large-scale cluster.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of network scheduling, and in particular to a server load balancing method based on distributed computing. BACKGROUND

[0002] Existing server load balancing methods are mostly based on static rules or simple dynamic scheduling strategies, such as round robin, least connection, consistent hashing, etc. Such methods are simple to implement, but often only focus on a single indicator, and cannot effectively take into account multiple dimensional constraints such as latency, resource utilization and service level agreement (SLA), making it difficult to cope with the challenges of burst traffic and business heterogeneity in large-scale distributed systems.

[0003] The centralized optimization methods and deep learning driven strategies that have emerged in recent years have improved scheduling results to some extent, but centralized methods have the problems of insufficient scalability and complex computation, and deep learning models face the limitations of poor convergence stability, weak interpretability and difficulty in cross-scene migration, making it difficult to run stably in large-scale data center environments for a long time.

[0004] Mean field game models have attracted attention because they can describe the group interaction behavior of large-scale participants, but traditional mean field games still have deficiencies when applied to server load balancing, including that the cost function only depends on the average latency and lacks tail risk characterization, business categories cannot be prioritized, distributed computing is sensitive to delay and packet loss, and communication overhead is large.

[0005] Therefore, how to provide a server load balancing method based on distributed computing is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a server load balancing method based on distributed computing. The present application introduces an improved distributed mean field game model in a distributed computing environment, each ingress load balancer performs optimal response calculation under local state and global mean field estimation constraints, and combines the differences in the weights of congestion cost, migration cost and risk-sensitive cost to build, achieving priority control for different business categories; by using mirror outer ladder double-step update, importance Gossip and delay-robust correction mechanisms, the convergence stability and adaptability to network uncertainty are improved while reducing communication overhead, thereby achieving efficient, stable and service level agreement-oriented load balancing in large-scale clusters.

[0007] A server load balancing method based on distributed computing according to an embodiment of the present application comprises the following steps:

[0008] In each ingress load balancer, instantaneous or short window statistical data of the server are collected and aggregated into a state vector, and requests are divided into different business categories;

[0009] constructing an objective function for each service category under an improved distributed mean field game model based on the state vector;

[0010] In the improved distributed mean field game, a mean field distribution is established for different service categories, and a feasible routing set is determined under capacity, quota and affinity constraints;

[0011] According to the objective function and the state vector, each ingress load balancer calculates the routing distribution as an optimal response to the mean field variable, solves it using a mirror outer ladder double-step update method, exchanges mean field estimates with neighboring nodes through an importance Gossip mechanism, and only performs communication on adjacent edges that are congested and active;

[0012] When updating the mean field estimate, time decay weights are applied to lagging or lost information, and a maximum protection strategy is introduced;

[0013] In the exploration mechanism of the improved distributed mean field game, a temperature-decreasing exploration disturbance is injected into the routing distribution, and a gradually increasing upper limit of the amount is set when a new server node is added;

[0014] According to the moving average results of the routing change rate, the mean field estimate change rate and the service level agreement violation rate, it is determined whether the mean field consistency condition is met or the risk response is triggered, and the final routing distribution is output when the convergence standard is met.

[0015] Further, the state vector generation step includes:

[0016] In the sampling period, the queue length, request delay, active connection number, processor and memory utilization, bandwidth occupancy and service level agreement violation of each server are obtained;

[0017] The data is short-window aggregated and normalized to obtain a state vector containing quantile statistics and moving average;

[0018] The service category division rules include:

[0019] Based on the service level label, port number or application identifier in the request identifier, and the preset service level strategy, the explicit label is determined first, then the strategy table mapping is determined, and finally the default rule is classified;

[0020] The request is divided into at least one of high priority class, normal priority class and batch processing class.

[0021] Further, the construction of the objective function includes:

[0022] Based on the delay quantile value and the queue length in the state vector, the congestion cost is set;

[0023] The migration cost is set based on data locality or migration distance between the request and the server;

[0024] The risk-sensitive cost is set based on high quantile latency probability exceeding a preset service level agreement threshold;

[0025] Different service categories combine the above cost items by setting differentiated weights to obtain respective target functions.

[0026] Further, the establishment of the average field distribution and the set of feasible routes comprises:

[0027] Each ingress load balancer uploads the local route distribution obtained based on the state vector, and the local route distribution is aggregated to form the average field variable corresponding to the service category;

[0028] The average field variable is used to depict the average traffic proportion of the service category in the global range, and serves as an input parameter of the subsequent optimal response;

[0029] When constructing the set of feasible routes, the route probability space corresponding to the average field variable is projected by combining the capacity threshold of the server, the cross-tenant quota limit, and the affinity strategy of the rack or available area;

[0030] The consistency of the network-wide aggregated route distribution and the average field variable is used as the convergence condition of the improved distributed average field game, and the set of feasible routes is updated.

[0031] Further, the optimal response solution and the average field estimation exchange comprises:

[0032] Each ingress load balancer updates the route distribution as an optimal response problem for the current average field variable, and the optimal response problem takes the route distribution as the decision variable under the constraint of the target function;

[0033] To solve the optimal response problem, a mirror outer gradient two-step update method is adopted, the first step calculates a predicted route distribution based on the current gradient, and the second step corrects the gradient using the prediction result and outputs the updated route distribution;

[0034] After completing the local optimal response calculation, each ingress load balancer exchanges information with adjacent nodes through the importance Gossip mechanism, the information includes the local average field estimation value and the dual variable, and only the adjacent edges represented by the state vector as congested active are selected to participate in the communication;

[0035] In the execution of the importance Gossip process, a weighted strategy based on the edge congestion degree is adopted to aggregate the received adjacent estimates to form an updated local average field variable, and the update result is used to replace the original estimation.

[0036] Furthermore, the delay robustness correction when updating the mean field estimate includes:

[0037] The timestamp attached to the received adjacency information is compared with the local clock to calculate the information delay.

[0038] When the delay exceeds a preset threshold, a time decay weight is applied to this information to reduce its weight in the mean field estimation update.

[0039] When delays or loss cause the adjacency index to be abnormally low, a maximum value protection strategy is introduced to limit the value of the adjacency index to a preset range that is not lower than the historical maximum value.

[0040] The local mean field estimate is updated based on the attenuated and protected adjacency information.

[0041] Furthermore, the exploration of disturbances and cold start surges includes:

[0042] Within the set of feasible routes, a random disturbance with temperature parameter control is applied to the route distribution, and the disturbance amplitude is gradually reduced according to a preset temperature decrease function.

[0043] When a new server node is detected to be added, an initial upper limit for the volume of the node is set, and the upper limit is gradually increased in a preset step over multiple consecutive sampling periods.

[0044] In the process of increasing the upper limit of the volume, the speed of increase is adjusted by combining the resource utilization and response latency indicators in the state vector of the node.

[0045] After the aforementioned increase in the upper limit is completed, the node will be included in the regular route distribution update.

[0046] Furthermore, the temperature decrease function includes:

[0047] In the initial exploration phase, the perturbation amplitude is set to a preset initial level and remains unchanged for several consecutive sampling periods;

[0048] During the exploration transition phase, the perturbation amplitude is gradually reduced after each sampling period according to a fixed decreasing step size until it approaches the preset lower limit;

[0049] During the stabilization phase, when the disturbance amplitude reaches the lower limit, it stops decreasing and instead maintains the lower limit value to sustain a small amount of exploratory disturbance in subsequent cycles.

[0050] Furthermore, the convergence determination and risk response triggering include:

[0051] Based on the rate of change of route distribution, the rate of change of mean field estimation and the rate of change of dual variables calculated by each ingress load balancer within a continuous sampling period, it is determined whether it is lower than the preset threshold.

[0052] The moving average of the service level agreement default rate recorded in the state vector is used to determine whether it is higher than a preset risk threshold.

[0053] When the rate of change meets the threshold condition, it is determined that the mean field uniformity convergence has been achieved.

[0054] When the default rate exceeds the risk threshold, a risk response process is triggered, and adjustments are made to the routing distribution within the constraints.

[0055] The beneficial effects of this invention are:

[0056] This invention introduces an improved distributed mean field game model into a distributed computing environment, enabling each ingress load balancer to make routing decisions under the dual constraints of local state information and global mean field estimation. This overcomes the scalability limitations of traditional centralized optimization methods and achieves distributed and scalable solutions in large-scale clusters.

[0057] This invention introduces congestion cost, migration cost, and risk-sensitive cost into the objective function, and combines them with differentiated weight configurations for service categories, enabling route allocation to simultaneously consider resource utilization efficiency and service level agreement (SLA) guarantee requirements, thereby improving the service quality control capability for high-priority services.

[0058] This invention employs a mirrored external ladder two-step update in the optimal response solution, and combines mechanisms such as importance Gossip, time decay, and maximum value protection to improve the convergence stability of distributed computing and its robustness to network latency and packet loss, thereby achieving stable load balancing with low communication overhead. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Fig. 1 This is a flowchart of a server load balancing method based on distributed computing proposed in this invention;

[0061] Fig. 2 This is a Gossip closed-loop diagram for a server load balancing method based on distributed computing proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figs. 1-2 A server load balancing method based on distributed computing includes the following steps:

[0064] The load balancer at each entry point collects instantaneous or short-window statistical data from the server and aggregates it into a state vector, which then divides requests into different business categories.

[0065] Based on the state vector, an objective function is constructed for each business category under the improved distributed average field game model;

[0066] In the improved distributed mean-field game, a mean-field distribution is established for different business categories, and a set of feasible routes is determined under constraints of capacity, quota, and affinity.

[0067] Each ingress load balancer uses the calculation of the route distribution as the optimal response to the mean field variable based on the objective function and state vector. It solves the problem using the mirrored external ladder two-step update method and exchanges the mean field estimate with neighboring nodes through the importance Gossip mechanism. Communication is only performed on congested and active adjacent edges.

[0068] When updating the mean field estimate, a time decay weight is applied to lagged or missing information and a maximum value protection strategy is introduced.

[0069] In the improved distributed average field game exploration mechanism, a temperature-decreasing exploration perturbation is injected into the routing distribution, and a gradually increasing release limit is set when a new server node is added.

[0070] Based on the moving average results of the route change rate, mean field estimated change rate, and service level agreement default rate, determine whether the mean field consistency condition is met or a risk response is triggered. If the convergence criteria are met, output the final route distribution.

[0071] In this embodiment, the step of generating the state vector includes:

[0072] Within each sampling period, each ingress load balancer acquires real-time server performance metrics, including at least: queue length, request response latency, number of active connections, processor utilization, memory utilization, bandwidth utilization, and service level agreement (SLA) breach status. This data is synchronized using a unified timestamp to ensure time-series consistency across different metrics.

[0073] After acquiring the data, the raw data undergoes short-window aggregation, which calculates the moving average and moving quantile using the most recent sampling periods as the time window to characterize the stability and fluctuation trend of the indicators. Subsequently, the aggregated data is normalized to map indicators of different dimensions to a unified interval, forming a multi-dimensional state vector containing quantile statistics and moving averages. This state vector serves as the input to the subsequent game theory model, representing the real-time operating state of the local nodes.

[0074] The detailed rules for classifying the business categories include: When the load balancer receives a request, it parses the identification information carried in the request. The identification information includes at least a service level label, port number, and application identifier. It matches the explicit service level label with a preset service level policy table. If a match is found, the request is directly classified into the corresponding category. If an explicit label is missing, it searches the policy table for a mapping relationship between the port number and the application identifier to determine the category. If no match is found, the request is classified into the default category.

[0075] After classification, all requests are divided into at least one of three categories: high-priority, normal-priority, and batch processing. High-priority requests correspond to stricter service level agreement requirements, normal-priority requests meet regular service needs, and batch processing requests are used for low-priority background tasks. The classification results are then bound to the state vector, providing input for subsequent objective function construction and mean-field game solving.

[0076] In this embodiment, the construction of the objective function includes:

[0077] After obtaining the state vector bound to the business category information, the ingress load balancer constructs corresponding objective functions for different categories. The objective function consists of three parts: congestion cost, migration cost, and risk-sensitive cost.

[0078] Congestion cost: Calculate the latency P99 and queue length quantile for each server within a sliding window; perform segmented linear mapping of latency according to the Service Level Agreement threshold: when P99 does not exceed the threshold, obtain the latency sub-item based on the ratio of P99 to the threshold; when P99 exceeds the threshold, obtain the latency sub-item by linearly amplifying the excess ratio; obtain the queue sub-item based on the ratio of the queue quantile to the preset capacity limit; the congestion cost is taken as the "weighted sum of the latency sub-item and the queue sub-item", where the weight is a fixed constant and set during deployment.

[0079] Migration cost: Based on the data locality and topological distance between the request and the target server, a hierarchical mapping is performed: the same rack is recorded as low-level cost, the same availability zone across racks is recorded as medium-level cost, and the different availability zones are recorded as high-level cost; when the request requires a cold start (including cache misses or loading dependencies), a fixed additional item is added to the corresponding level cost; the migration cost is the sum of the level cost and the additional item for whether it is a cold start.

[0080] Risk-sensitive cost: Within a sliding window, the percentage of samples with latency exceeding the service level agreement threshold is counted as the excess probability, and the ratio of the average magnitude of the excess to the threshold is calculated as the excess intensity; the risk-sensitive cost is the weighted sum of the excess probability and the excess intensity, where the weight is a fixed constant and is set during deployment.

[0081] For different business categories such as high priority, normal, and batch processing, a set of fixed weights is set for each category. The congestion cost, migration cost, and risk-sensitive cost are linearly combined to obtain the objective function for that category. The weights are non-negative, configured during deployment, and satisfy the constraint that the sum of all weights is 1.

[0082] The objective function constructed in the above manner can reflect both the local operating state of the server and the priority differences between business categories, providing a quantitative basis for calculating the optimal response in the improved distributed average field game.

[0083] In this embodiment, the establishment of the average field distribution and feasible route set includes:

[0084] After completing the local objective function calculation, each ingress load balancer obtains the local route distribution for the current sampling period. This local route distribution represents the request allocation ratio for that ingress point across different servers.

[0085] The local route distribution of all entry nodes is aggregated according to business category to form a global average field variable for the corresponding category. The average field variable is used to describe the average distribution of requests of this category across all servers nationwide.

[0086] After generating the mean field variables, a set of feasible routes needs to be constructed under constraints. These constraints include:

[0087] Capacity constraints: When a server's processor utilization, memory utilization, or queue length approaches a preset capacity threshold, the server's allocation ratio in the route set is limited; when the state vector shows that a server's processor utilization, memory utilization, or network bandwidth utilization exceeds a preset safety threshold, the server's maximum allocation ratio in the route distribution is limited; when a server's queue length or latency percentile exceeds a set threshold, the server is marked as an overloaded node, and its allocation ratio in the feasible route set is reduced; when the server's state recovers to below the threshold, the restrictions are gradually lifted, allowing the server to accept requests at a normal ratio again.

[0088] Quota constraints: Set maximum traffic quotas or minimum service guarantee ratios for specific tenants or business categories to ensure fairness among different businesses; set maximum traffic quotas or minimum guarantee ratios for different business categories according to service level agreements; when the allocation of a certain business category on a certain server exceeds the quota limit, the excess portion will be forcibly reallocated to other servers; when a certain business category does not reach the guarantee ratio, resources will be allocated to that category first until the minimum guarantee is met; in multi-tenant scenarios, business requests from different tenants will be routed according to the quota ratios agreed upon in the tenant contracts.

[0089] Affinity Constraints: Based on rack topology, availability zone location, or historical scheduling policies, requests of the same type are preferentially restricted to a set of servers with higher affinity. When a request originates from a server with topological affinity (e.g., located in the same rack, availability zone, or edge node), the priority allocation ratio of that server in the feasible route set is increased. When a request crosses different racks or availability zones, the allocation ratio is limited due to the long migration distance. When a request involves servers with dedicated caches, storage binding, or application dependencies, the priority selection of this type of server is maintained to avoid additional migration costs. When a node failure or rack isolation occurs, the node and its corresponding affinity mapping are removed from the feasible route set.

[0090] When applying the above constraints, the mean field variable is checked. If there are cases where the allocation ratio exceeds the capacity threshold, violates the quota limit, or does not meet the affinity policy, then constraint projection is performed on that allocation ratio to adjust it to the range that meets the conditions. The set obtained after constraint projection processing is the feasible route set.

[0091] This set of feasible routes serves as the solution space for subsequent optimal response calculations, ensuring that the route distribution reflects the trend of the global mean field variables and meets capacity and policy requirements during the update process.

[0092] In this embodiment, the exchange between the optimal response solution and the mean-field estimation includes:

[0093] Each ingress load balancer uses the objective function as the optimization objective and the feasible route set as the constraint space to describe the route distribution update as the optimal response problem for the current average field variable;

[0094] The decision variables for the optimal response problem are the routing probability vectors corresponding to different servers, and the optimization objective is to minimize the weighted sum of congestion cost, migration cost, and risk-sensitive cost.

[0095] In the first update step, each entry point calculates the gradient of the objective function based on the local state vector and mean field variables, and performs a predictive update of the route distribution based on the gradient to obtain the intermediate route distribution.

[0096] In the second update step, each entry point recalculates the correction gradient using the prediction results, and updates the route distribution again based on the correction gradient to obtain a new local route distribution;

[0097] After completing the local route distribution update, each ingress load balancer exchanges information with its neighboring nodes. The information exchanged includes the local mean field estimate and dual variables.

[0098] The selection of adjacent edges is based on the determination of congestion activity in the state vector, and information exchange is only performed on adjacent edges marked as congestion active.

[0099] During the information exchange process, each entry point assigns weights to the received neighbor mean field estimates according to the congestion level of adjacent edges, and performs weighted aggregation to obtain the updated local mean field variables.

[0100] The congestion metric for the server corresponding to the adjacent edge is calculated by adding the "ratio of queue length to maximum queue capacity" and the "ratio of latency P99 to the Service Level Agreement (SLA) threshold". For example, if the queue length of an adjacent edge is 0.6 times the maximum capacity and the latency P99 is 0.4 times the threshold, then the congestion metric for that adjacent edge is 1.0.

[0101] After obtaining the congestion indices for all adjacent edges, a reciprocal normalization process is performed. This involves first taking the reciprocal of each congestion index, then normalizing all the reciprocals until their sum equals 1, which is then used as the final allocation weight. This method ensures that adjacent edges with higher congestion levels receive lower allocation weights, and adjacent edges with lower congestion levels receive higher allocation weights.

[0102] Each ingress load balancer multiplies the received neighbor mean field estimate with its corresponding allocation weight and sums the results to obtain a weighted aggregation result. This weighted aggregation result is the updated local mean field variable and is used in the next sampling period to calculate the optimal response.

[0103] The updated mean field variable replaces the original estimate and serves as the input for calculating the optimal response in the next sampling period.

[0104] In this embodiment, the delay robustness correction when updating the mean field estimate includes:

[0105] When the ingress load balancer receives mean field estimation information uploaded by neighboring nodes, it compares the timestamp attached to the information with the local clock to detect any delay. If the delay exceeds a preset threshold, the information is assigned a gradually decreasing attenuation weight, so that its impact on the update process weakens as the delay increases; if the delay exceeds the maximum allowable range, the information is discarded and no longer participates in the update.

[0106] In cases where some adjacency information is lost or abnormally low, a protection lower limit is set for the relevant indicators. This protection lower limit is determined based on historical observations to ensure that the indicator does not fall below a reasonable range when participating in aggregation. The adjacency information after attenuation and protection processing is weighted and aggregated together with the normal information to obtain the corrected local mean field variable, which is then used as the input for calculating the optimal response in the next sampling period.

[0107] In this embodiment, the exploration disturbance and cold start surge include:

[0108] Within the set of feasible routes, a random perturbation based on a temperature-decreasing function is applied to the route distribution. The temperature-decreasing function remains high in the initial stage to ensure sufficient randomness in the route distribution; in the intermediate stage, the perturbation amplitude is gradually reduced according to a preset step size to minimize unnecessary fluctuations; and in subsequent stages, the perturbation amplitude remains constant when it drops to a set lower limit, thus maintaining a weak exploration during the convergence phase.

[0109] When a new server node is detected joining, an initial load factor cap is set for that node to prevent resource overload caused by a sudden surge in requests. In subsequent sampling periods, the load factor cap for that node is increased gradually, with the increase rate dynamically adjusted based on metrics such as processor utilization, memory usage, and response latency, until the node has stable processing capacity. After the load factor cap has been increased, the node participates in regular route distribution updates along with other servers.

[0110] In this embodiment, the temperature decrease function includes:

[0111] The temperature decrease function takes the sampling period number as input and the disturbance amplitude as output, and is implemented in a piecewise linear decrease manner.

[0112] In the initial exploration phase, the function output is fixed at 1.0 as the perturbation amplitude and remains unchanged for the first 10 sampling periods.

[0113] During the transition phase, starting from the 11th sampling period, the function reduces the perturbation amplitude by 0.05 after each sampling period until it drops to 0.2. That is, the output is 0.95 in the 11th period, 0.90 in the 12th period, and so on, until the output is 0.20 in the 26th period.

[0114] During the stabilization phase, starting from the 26th sampling period, the function output is fixed at 0.20, no longer decreasing, and remains at this value in all subsequent sampling periods.

[0115] The temperature decrease function corresponds to three distinct logics at different stages: constant maintenance, linear decrease, and constant maintenance, which ensures the control process in which the disturbance amplitude gradually transitions from an initial high value to a stable low value.

[0116] In this embodiment, the process of increasing the upper limit of cold start volume is implemented by taking the sampling period number as input and the upper limit of the allocation ratio as output, and adopting a segmented linear increment method.

[0117] During the initial cold start phase, the maximum number of newly added server nodes is set to 10%, and remains unchanged for the first 5 sampling periods.

[0118] During the ramp-up transition phase, starting from the 6th sampling period, the system increases the ramp-up limit of that node by 10% after each sampling period, until it reaches 50%. That is, it is 20% in the 6th period, 30% in the 7th period, 40% in the 8th period, and 50% in the 9th period.

[0119] During the stable volume ramp-up phase, starting from the 9th sampling period, the upper limit of volume ramp-up is fixed at 50%, and this value is maintained in all subsequent sampling periods.

[0120] Through the above method, the cold start node's capacity gradually increases linearly from an initial low proportion to a stable level, avoiding node overload caused by a sudden surge in requests, while ensuring that new nodes can gradually participate in normal routing allocation.

[0121] In this embodiment, the convergence determination and risk response triggering include:

[0122] After each ingress load balancer completes the route distribution update, the route change rate, mean field estimate change rate, and dual variable change rate are continuously monitored over 10 sampling periods. When the maximum value of all three is below 1%, the current iteration process is determined to have reached the mean field consistent convergence state, and the route distribution at that moment is output as a stable result.

[0123] Simultaneously, risk monitoring is conducted on the service level agreement (SLA) default rate. Specifically, the moving average of the default rate over the past 20 sampling periods is calculated, and a risk response process is immediately triggered when this value exceeds 2%. During the risk response process, the ingress load balancer reduces the proportion of traffic allocated to high-latency servers by 20% and reallocates this portion of traffic to servers with queue lengths in the lowest quartile, thereby correcting the routing distribution in the next sampling period.

[0124] Example 1:

[0125] To verify the feasibility of this invention in practice, it was applied to a data center load balancing scenario of a cloud computing platform. This data center contains 300 servers, divided into three availability zones, each with 100 servers. The test scenario involved simultaneously handling three types of services: high-priority, normal-priority, and batch processing, with a total request volume of approximately 300,000 requests per second (QPS). Specifically, high-priority services accounted for 20%, normal-priority services for 50%, and batch processing services for 30%.

[0126] In the comparative experiment, load balancing was performed using both the traditional least connection method and the improved distributed average field game method of this invention.

[0127] Table 1 Test Results

[0128]

[0129] As shown in Table 1, under the same load conditions, after adopting the method of this invention, the P99 latency for high-priority services decreased from 185ms to 95ms, the P99 latency for ordinary services decreased from 210ms to 140ms, and the latency for batch processing services was slightly improved. The overall server utilization rate increased from 63.4% to 76.1%, indicating more efficient resource utilization. The proportion of requests exceeding the SLA decreased from 3.2% to 0.9%, demonstrating the significant effectiveness of the method of this invention in ensuring service level agreements.

[0130] The method of this invention generates only about 18.7KB / s of neighbor node communication overhead during the distributed solution process, which is far lower than the global synchronization bandwidth consumption required by the centralized method; it can converge to a stable route distribution within 7 iterations, showing good convergence speed.

[0131] The load balancing method implemented by the present invention through an improved distributed average field game can balance high-priority service protection, resource utilization improvement and communication overhead control in a large-scale cluster environment, and has strong engineering feasibility.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A server load balancing method based on distributed computing, characterized in that, Includes the following steps: The load balancer at each entry point collects instantaneous or short-window statistical data from the server and aggregates it into a state vector, which then divides requests into different business categories. Based on the state vector, an objective function is constructed for each business category under the improved distributed average field game model; In the improved distributed mean-field game, a mean-field distribution is established for different business categories, and a set of feasible routes is determined under constraints of capacity, quota, and affinity. Each ingress load balancer uses the calculation of the route distribution as the optimal response to the mean field variable based on the objective function and state vector. It solves the problem using the mirrored external ladder two-step update method and exchanges the mean field estimate with neighboring nodes through the importance Gossip mechanism. Communication is only performed on congested and active adjacent edges. When updating the mean field estimate, a time decay weight is applied to lagged or missing information and a maximum value protection strategy is introduced. In the exploration mechanism of the improved distributed average field game, a temperature-decreasing exploration perturbation is injected into the routing distribution, and a gradually increasing upper limit is set when a new server node is added. Based on the moving average results of the route change rate, mean field estimated change rate, and service level agreement default rate, determine whether the mean field consistency condition is met or a risk response is triggered. If the convergence criteria are met, output the final route distribution.

2. The server load balancing method based on distributed computing according to claim 1, characterized in that, The steps for generating the state vector include: During the sampling period, obtain the queue length, request latency, number of active connections, processor and memory utilization, bandwidth utilization, and service level agreement (SLA) breach status of each server. The data is processed by short-window aggregation and normalization to obtain a state vector containing quantile statistics and moving averages; The detailed rules for classifying the business categories include: The service level tag, port number, or application identifier in the request identifier is matched with the preset service level policy. Explicit tags are judged first, the policy table mapping is judged, and the system is classified according to the default rules. The requests are classified into at least one of the following categories: high priority, normal priority, and batch processing.

3. The server load balancing method based on distributed computing according to claim 1, characterized in that, The construction of the objective function includes: The congestion cost is set based on the delay quantile value in the state vector and the queue length. Set migration costs based on data locality or migration distance between the request and the server; Risk-sensitive costs are set based on the probability of high-quantile latency exceeding the preset service level agreement threshold; Different business categories obtain their own objective functions by setting differentiated weight combinations of cost terms.

4. The server load balancing method based on distributed computing according to claim 1, characterized in that, The establishment of the average field distribution and feasible route set includes: Each ingress load balancer will upload the local route distribution obtained based on the state vector, and aggregate them to form the average field variable corresponding to the service category; The mean field variable is used to characterize the average traffic share of the business category in the global scope and serves as an input parameter for the subsequent optimal response. When constructing a set of feasible routes, the route probability space corresponding to the mean field variable is constrained by combining server capacity thresholds, cross-tenant quota limits, and rack or availability zone affinity strategies. The consistency between the aggregated route distribution across the entire network and the average field variables is used as the convergence condition for the improved distributed average field game, and the feasible route set is updated accordingly.

5. A server load balancing method based on distributed computing according to claim 1, characterized in that, The exchange between the optimal response solution and the mean-field estimation includes: Each ingress load balancer describes the update of the route distribution as an optimal response problem for the current mean field variable, and the optimal response problem uses the route distribution as the decision variable under the constraint of the objective function; To solve the optimal response problem, a mirrored external ladder two-step update method is adopted. The first step is to calculate the predicted route distribution based on the current gradient, and the second step is to use the prediction results to correct the gradient and output the updated route distribution. After completing the local optimal response calculation, each ingress load balancer exchanges information with neighboring nodes through the importance Gossip mechanism. The information includes the local mean field estimate and dual variables, and only the adjacent edges represented by the state vector as congestion active are selected to participate in the communication. During the execution of the importance Gossip, a weighted strategy based on edge congestion is used to aggregate the received adjacency estimates to form an updated local mean field variable, and the updated result replaces the original estimate.

6. The server load balancing method based on distributed computing according to claim 1, characterized in that, The delay robustness correction for updating the mean field estimate includes: The timestamp attached to the received adjacency information is compared with the local clock to calculate the information delay. When the delay exceeds a preset threshold, a time decay weight is applied to the information, reducing its weight in the mean field estimation update; When delays or loss cause the adjacency index to be abnormally low, a maximum value protection strategy is introduced to limit the value of the adjacency index to a preset range that is not lower than the historical maximum value. The local mean field estimate is updated based on the attenuated and protected adjacency information.

7. A server load balancing method based on distributed computing according to claim 1, characterized in that, The exploration disturbances and cold start surges include: Within the set of feasible routes, a random disturbance with temperature parameter control is applied to the route distribution, and the disturbance amplitude is gradually reduced according to a preset temperature decrease function. When a new server node is detected to be added, an initial upper limit for the node is set, and the upper limit is gradually increased in a preset step over multiple consecutive sampling periods. In the process of increasing the upper limit of the volume, the speed of increase is adjusted by combining the resource utilization and response latency indicators in the node's state vector. After the aforementioned increase in the upper limit is completed, the nodes will be included in the regular route distribution update.

8. A server load balancing method based on distributed computing according to claim 1, characterized in that, The temperature decrease function includes: In the initial exploration phase, the perturbation amplitude is set to a preset initial level and remains unchanged for several consecutive sampling periods; During the exploration transition phase, the perturbation amplitude is gradually reduced after each sampling period according to a fixed decreasing step size until it approaches the preset lower limit; During the stabilization phase, when the disturbance amplitude reaches the lower limit, it stops decreasing and instead remains at the lower limit value to maintain a small amount of exploratory disturbance in subsequent cycles.

9. A server load balancing method based on distributed computing according to claim 1, characterized in that, The convergence determination and risk response triggering include: Based on the rate of change of route distribution, the rate of change of mean field estimation and the rate of change of dual variables calculated by each ingress load balancer within a continuous sampling period, it is determined whether it is lower than the preset threshold. The moving average of the service level agreement default rate recorded in the state vector is used to determine whether it is higher than a preset risk threshold. When the rate of change meets the threshold condition, it is determined that the mean field uniformity convergence has been achieved. When the default rate exceeds the risk threshold, a risk response process is triggered, and adjustments are made to the routing distribution within the constraints.

Images