Balanced distribution method and system for computing power network service requests

Abstract

The invention relates to a computing power network service request balanced distribution method and system, and the method comprises the steps: collecting the global information in a current preset period, including the request flow information of each access node, and the computing power information, network state information and historical load information of each service node, and carrying out the centralized convex optimization. Generating a global distribution strategy based on the global information; issuing the global distribution strategy to each access node, dynamically adjusting the distribution strategy through a negative feedback mechanism within a preset time sequence interval according to a local real-time request processing condition after the access node receives the global distribution strategy, and forwarding the distribution strategy to the service node according to the adjusted distribution strategy; and feedback information from the service node and the access node is received in real time, a global distribution strategy is updated based on the feedback information, and the distribution precision and the load balancing effect are improved in the next period. According to the invention, a global optimal strategy is realized through convex optimization, and the distribution probability is adjusted in real time in combination with a negative feedback mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of computing power networks, and particularly to a method and system for balanced distribution of computing power network service requests. Background Art

[0002] With the rapid development of cloud computing, edge computing, and distributed computing, the computing power network, as a new type of computing architecture, aims to meet the diverse service request requirements by efficiently scheduling and allocating computing resources. Regarding the service request distribution problem in the computing power network, some solutions have been proposed in related technical solutions, but there are still certain limitations.

[0003] For example, Chinese Patent CN118760529B discloses a method and system for determining a computing power scheduling strategy based on heterogeneous planning. This method obtains the set of computing power requests received by multiple network modules in the computing power network, determines the request pressure of each network module according to the request parameters, and calculates the operation tolerance by combining the historical operation records and module performance parameters, and finally generates a computing power scheduling strategy. This solution can improve the utilization rate and scheduling effect of computing power resources to a certain extent. However, this method has the following deficiencies: First, it relies on historical operation records and static performance parameters for scheduling decisions, and has insufficient adaptability to real-time network status and sudden traffic changes, which may lead to a lag in the scheduling strategy and is difficult to meet the requirements of dynamic traffic scenarios. Second, this solution does not explicitly include network latency in the optimization objective. In a cross-regional or cross-layer computing power network, requests may be allocated to nodes with higher latency, affecting the service response speed and user experience. In addition, this method adopts a centralized scheduling method and does not design a distributed cooperation mechanism, which may face performance bottlenecks or single-point failure risks in a large-scale distributed environment, limiting the scalability and robustness of the system.

[0004] Another Chinese patent CN116743664A discloses a request allocation method and device based on load balancing. This method monitors the connection numbers of multiple service instances, determines whether the balance condition is met according to a preset weight ratio, and adjusts the weights of abnormal service instances to achieve balanced allocation of connection requests. This solution can avoid the load skew problem caused by uneven request allocation to a certain extent. However, there are also obvious defects in its technical implementation: First, this method takes the connection number as the core indicator and does not consider the dynamic characteristics of requests (such as computing resource consumption or processing duration), which may cause resource-intensive requests to concentrate on some instances. Even if the connection numbers are balanced, resource overload or performance degradation may still occur. Second, this solution does not fully consider the network topology and latency factors. In a distributed computing power scenario, requests may be allocated to distant instances, increasing the response time and reducing the service quality. In addition, this method lacks a global optimization mechanism by locally adjusting the weights of abnormal instances. In a complex environment with multiple service instances and multiple access nodes, it may not be able to achieve overall load balancing, and may even cause system jitter or efficiency decline. Finally, this solution monitors and adjusts weights in a centralized manner, lacking distributed collaboration capabilities. It may face performance bottlenecks in large-scale distributed scenarios and is insufficient in reliability in the event of network partitioning or central node failure.

[0005] Therefore, the existing technologies still have the following common problems in the field of computing power network service request distribution: lacking the ability to quickly respond to dynamic traffic and real-time network status, and at the same time unable to achieve the coordination of global optimization and local adjustment in a distributed environment, resulting in low load balancing efficiency, insufficient network latency optimization, and limited system scalability. Therefore, there is an urgent need for a method for balanced distribution of computing power network service requests that can comprehensively consider real-time performance, network latency, global optimization, and distributed collaboration to improve the utilization efficiency of computing power resources, service response speed, and the robustness and scalability of the system. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for balanced distribution of computing power network service requests to solve the problems that the existing computing power network service request distribution technology lacks the collaborative scheduling ability across multiple layers of nodes and the comprehensive feedback mechanism of real-time load and latency, resulting in uneven request distribution, load imbalance, and low efficiency of the global distribution strategy, which limits the resource utilization efficiency and service quality.

[0007] To achieve one of the above-mentioned invention purposes, an embodiment of the present invention provides a method for balanced distribution of computing power network service requests, the method includes,

[0008] Collect global information within the current preset period, including request flow information of each access node, computing power information, network status information, and historical load information of each service node, and generate a global distribution strategy based on the global information through a centralized convex optimization method;

[0009] The global distribution policy is sent to each access node. After the access node receives the global distribution policy, according to the local real-time request processing situation, within a preset time sequence interval, the distribution policy is dynamically adjusted through a negative feedback mechanism, and is forwarded to the service node according to the adjusted distribution policy;

[0010] Feedback information from the service node and the access node is received in real time, and the global distribution policy is updated based on the feedback information to improve the distribution accuracy and load balancing effect in the next cycle.

[0011] As a further improvement of an embodiment of the present invention, the method further includes that the step of "generating a global distribution policy based on global information through a centralized convex optimization method" includes:

[0012] Define the access node router as , where , r is the access node, is the total number of access nodes;

[0013] Collect the request flow information of each access node and the computing power information of each service node; among them, the request flow information includes the request quantity and type, and the computing power information includes the service type, the current number of requests to be processed, and the maximum number of parallel requests that can be processed;

[0014] According to the historical load information and network topology, predict the request quantity of each access node in the next cycle through a machine learning model;

[0015] Based on the centralized convex optimization method, with the goal of minimizing the variance of the comprehensive request quantity of the service node, establish an optimization problem, generate a global distribution policy, and dynamically adjust the distribution probability;

[0016] The comprehensive request quantity of the service node is expressed as:

[0017]

[0018] where is the comprehensive request quantity of service node i, is the maximum number of parallel requests that can be processed by service node i, is the current number of requests to be processed by service node i, is the request quantity of access node r in the next cycle predicted through the machine learning model, is the distribution probability from access node r to service node i, and satisfies and .

[0019] As a further improvement of an embodiment of the present invention, the method further includes, based on the centralized convex optimization method, taking the minimization of the variance of the comprehensive request quantity of the service nodes as the objective, and establishing an optimization problem with the distribution probability as the optimization variable, which is expressed as:

[0020]

[0021]

[0022]

[0023]

[0024] wherein, is the comprehensive request quantity of service node m, is the network delay from access node r to service node i, is the delay weighting coefficient, is the average value of the network delay, is the total number of service nodes.

[0025] As a further improvement of an embodiment of the present invention, the method further includes that the "dynamically adjusting the distribution strategy through a negative feedback mechanism according to the local real-time request processing situation within a preset time sequence interval" includes:

[0026] After the access node receives the global distribution strategy, it preliminarily distributes each arriving request according to the distribution probability in the global distribution strategy;

[0027] During the request forwarding process, collect the feedback information of the service nodes, including the load status and delay information, and dynamically adjust the preliminary distribution strategy based on the negative feedback mechanism according to the load ratio of the service nodes and the network delay;

[0028] wherein, the load ratio of service node i is , is the number of pending requests of service node i.

[0029] As a further improvement of an embodiment of the present invention, the method further includes that dynamically adjusting the preliminary distribution strategy through the negative feedback mechanism includes,

[0030] Simulate distributing the requests to each service node respectively, and calculate the change amount of the load variance of all service nodes in the network after the distribution;

[0031] Adjust the distribution probability of the preliminary distribution strategy according to the change amount of the load variance, the load ratio of the service nodes and the network delay , so that the distribution result tends to global load balancing; the adjustment strategy includes,

[0032] When the load variance decreases, the load ratio is lower than the threshold, and the time delay is shorter than the threshold, increase according to the predefined ; when the load variance increases, the load ratio is higher than the threshold, or the time delay is longer than the threshold, decrease according to the predefined ;

[0033] When a preset number of requests are processed or a preset time sequence interval is reached, perform the adjustment of the distribution probability to ensure the real-time performance and accuracy of the adjustment process.

[0034] As a further improvement of an embodiment of the present invention, the method further includes that the "updating the global distribution policy based on feedback information" includes:

[0035] Regularly receive the latest computing power information and network status information broadcast by the service nodes;

[0036] Collect the local request processing situation and distribution effect of the access nodes, including the average time delay and the load balancing situation;

[0037] Based on the feedback information, dynamically adjust the distribution parameters in the next cycle, including the number of requests and the time delay weighting coefficient to improve the load balancing of request distribution and the time delay optimization effect.

[0038] As a further improvement of an embodiment of the present invention, the method further includes that after the access node receives the global distribution policy, or when the local processing reaches a preset number of requests, or a preset time sequence interval, perform local information update, including,

[0039] Calculate the total number of requests that have been forwarded by all access nodes;

[0040] Calculate the average distribution probability based on the negative feedback mechanism, and update the computing power information and network status information of the service nodes stored locally by the access nodes;

[0041] Through periodic local information update, improve the response speed and adaptability of distributed decision-making.

[0042] To achieve one of the above-mentioned invention purposes, an embodiment of the present invention further provides a computing power network service request balanced distribution system, which includes a global policy generation module, a local policy adjustment module, and a feedback optimization module;

[0043] The global policy generation module is used to collect the global information in the current preset cycle, including the request flow information of each access node, as well as the computing power information, network status information, and historical load information of each service node, and generate a global distribution policy based on the global information through a centralized convex optimization method;

[0044] The local policy adjustment module is used to send the global distribution policy to each access node. After the access node receives the global distribution policy, according to the local real-time request processing situation, within a preset time sequence interval, the distribution policy is dynamically adjusted through a negative feedback mechanism, and is forwarded to the service node according to the adjusted distribution policy;

[0045] The feedback optimization module is used to receive feedback information from the service node and the access node in real time, and update the global distribution policy based on the feedback information to improve the distribution accuracy and load balancing effect in the next cycle.

[0046] To achieve one of the above-mentioned invention purposes, an embodiment of the present invention further provides an electronic device, including a memory and a processor, characterized in that a computer program that can run on the processor is stored in the memory, and when the program is executed on the processor, the steps in the above-mentioned method for balanced distribution of computing power network service requests are implemented.

[0047] To achieve one of the above-mentioned invention purposes, an embodiment of the present invention further provides a storage medium, the storage medium stores a computer program, characterized in that when the computer program is executed by a processor, the steps in the above-mentioned method for balanced distribution of computing power network service requests are implemented.

[0048] Compared with the prior art, the present invention provides a method and system for balanced distribution of computing power network service requests, which supports centralized and distributed dual-mode scheduling, realizes the global optimal policy through convex optimization, combines a negative feedback mechanism to adjust the distribution probability in real time, and improves the cross-layer resource scheduling efficiency and dynamic adaptability. Multi-objective optimization is introduced to flexibly balance load balancing and network delay, and different scenarios are adapted by adjusting the delay coefficient. In the distributed mode, the negative feedback regulation enhances the ability to handle burst traffic, and the local estimation mechanism reduces the information synchronization overhead, alleviates the impact of delay, and significantly improves the resource utilization rate, service quality and system robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is the overall flowchart of the method for balanced distribution of computing power network service requests described in the present invention.

[0050] Figure 2 is the schematic architecture diagram of the system for balanced distribution of computing power network service requests described in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The present invention will be described in detail below in conjunction with the specific embodiments shown in the drawings. However, these embodiments do not limit the present invention, and any structural, method, or functional transformation made by those of ordinary skill in the art based on these embodiments is included in the protection scope of the present invention.

[0052] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where like or similar reference numerals denote like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0053] In the first embodiment of the present invention, the present invention provides a method for balanced distribution of computing power network service requests, as Figure 1 shown, the method includes,

[0054] S1: Collect global information within the current preset period, including request flow information of each access node, computing power information, network status information, and historical load information of each service node, and generate a global distribution strategy based on the global information through a centralized convex optimization method;

[0055] S2: Send the global distribution strategy to each access node, and after the access node receives the global distribution strategy, according to the local real-time request processing situation, within a preset time sequence interval, dynamically adjust the distribution strategy through a negative feedback mechanism, and forward it to the service node according to the adjusted distribution strategy;

[0056] S3: Real-time receive feedback information from service nodes and access nodes, and update the global distribution strategy based on the feedback information to improve the distribution accuracy and load balancing effect in the next cycle.

[0057] In a specific embodiment of the present invention, a global distribution strategy is generated based on global information through a centralized convex optimization method, specifically,

[0058] Define the access node router as , where , r is the access node, is the total number of access nodes;

[0059] Collect the request flow information of each access node and the computing power information of each service node; among them, the request flow information includes the request quantity and type, and the computing power information includes the service type, the current number of requests to be processed, and the maximum number of parallel requests that can be processed;

[0060] According to the historical load information and network topology, predict the request quantity of each access node in the next cycle through a machine learning model;

[0061] Based on the centralized convex optimization method, with the goal of minimizing the variance of the comprehensive request quantity of service nodes, establish an optimization problem, generate a global distribution strategy, and dynamically adjust the distribution probability;

[0062] The comprehensive request quantity of the service node is expressed as:

[0063]

[0064] Among them, is the comprehensive request quantity of service node i, is the maximum parallel processing request quantity of service node i, is the current pending request quantity of service node i, is the request quantity of access node r in the next cycle predicted by the machine learning model, is the distribution probability from access node r to service node i, and satisfies and .

[0065] Furthermore, based on the above centralized convex optimization method, with the goal of minimizing the variance of the comprehensive request quantity of service nodes, an optimization problem with the distribution probability as the optimization variable is established and expressed as:

[0066]

[0067]

[0068]

[0069]

[0070] Among them, is the comprehensive request quantity of service node m, is the network delay from access node r to service node i, is the delay weighting coefficient, is the average value of network delays, is the total number of service nodes.

[0071] It should be noted that in the present invention, the access node router needs to decide which specific computing power service node (such as an edge cloud or a central cloud computing node) the requests arriving locally are distributed to for response. Let be the number of service nodes, be the number of access nodes. Let the maximum parallel processing request quantity of the computing power service node be , is the maximum parallel processing request quantity of the th service node. Using the distribution probability as the optimization variable, the distribution probability that access node schedules requests to service node is expressed as , and there is:

[0072]

[0073]

[0074]

[0075] The sum of the probabilities of distributing the same access node to different service nodes is 1. Therefore, there is .

[0076] Furthermore, the distribution probability matrix X is expressed as:

[0077]

[0078] Access node to service node network delay can be obtained by network probing, and its matrix representation is:

[0079]

[0080] Suppose the centralized mechanism estimates the number of requests arriving within the next time slot period (broadcast period or request distribution policy update period) for the access router as . Therefore, the estimated vector of the number of requests arriving in the next time slot is:

[0081]

[0082] Then the total network delay of all requests after the next time slot ends is:

[0083]

[0084] Suppose the current number of requests being queued and processed detected by the centralized method is . Then the weighted number of requests is . The weighted number of requests eliminates the difference in processing capabilities of different computing power service nodes. Then, the combined number of requests that the service node has completed processing and is queuing and processing after the next time slot ends is estimated as:

[0085]

[0086] It can also be expressed as the sum of the remaining requests to be processed at the current moment and the requests to be distributed in the next time slot. After the next time slot ends, the vector containing the combined number of requests of each service node is expressed as:

[0087] ​

[0088] To achieve the balanced distribution of requests in the computing power network, this method takes the variance of the comprehensive request quantity of service nodes as the optimization objective, and the objective function is represented by as follows:

[0089]

[0090] In addition to the optimization objective, this method also takes the distribution probability from access nodes to service nodes as a variable, and constructs a quadratic convex optimization mathematical model by combining the network delay requirement constraint and the single access node distribution probability constraint, as follows:

[0091]

[0092]

[0093]

[0094]

[0095] Objective function: Minimize the weighted load variance of computing power service nodes

[0096] Constraint 1: The average transmission delay of requests is less than or equal to the product of the average network delay and the weighting coefficient of

[0097] Constraint 2: The value range of the probability variable is between 0 and 1

[0098] Constraint 3: The sum of the probability variables of a single access node is 1

[0099] Weighting coefficient is used to balance the optimization of the distributed service load balancing of the computing power network or the optimization of network delay.

[0100] In a specific implementation scenario of the present invention, a mature operations research tool gurobipy is used to solve the quadratic objective function, and the pseudo-code of the solution process is as follows: Algorithm1: Method for Balanced Distribution of Computing Power Network Service Requests Based on Centralized Quadratic Convex Optimization Input: Number of access nodes Number of service nodes , service node information , transmission delay matrix , vector of the number of requests arriving in the next time slot Return: Request distribution probability matrix 1: / / Initialize the request distribution probability variable 2: = 0 / / Initialize the optimization objective 3: / / Establish the model 4: / / Update the model environment 5: for ( = 1 to ) / / Define the optimization objective 6: 7: for ( = 1 to ) 8: 9: for ( = 1 to ) 10: 11: end for 12: end for 13: 14: 15: 16: end for 17: / / Set the model optimization objective, with the direction of minimizing 18: / / Add the delay constraint 19: / / Add the variable range constraint 20: / / Add the variable constraint 21: / / Execute the quadratic form solution algorithm to obtain the optimal solution

[0101] In a specific embodiment of the present invention, according to the local real-time request processing situation, within a preset time sequence interval, the distribution strategy is dynamically adjusted through a negative feedback mechanism. Specifically,

[0102] After the access node receives the global distribution strategy, each arriving request is preliminarily distributed according to the distribution probability in the global distribution strategy;

[0103] During the request forwarding process, feedback information of service nodes is collected, including load status and latency information, and based on the negative feedback mechanism, the preliminary distribution strategy is dynamically adjusted according to the load ratio of service nodes and network latency;

[0104] Among them, the load ratio of service node i is , where

[0105] Dynamically adjusting the preliminary distribution strategy through the negative feedback mechanism includes,

[0106] Simulating the distribution of requests to each service node respectively, and calculating the change amount of the load variance of all service nodes in the network after distribution;

[0107] Adjust the distribution probability of the preliminary distribution strategy according to the change amount of the load variance, the load ratio of the service node and the network latency , so that the distribution result tends to global load balancing; the adjustment strategy includes,

[0108] When the load variance decreases, the load ratio is lower than the threshold and the latency is shorter than the threshold, increase according to the predefined ; when the load variance increases, the load ratio is higher than the threshold or the latency is longer than the threshold, decrease according to the predefined ;

[0109] When processing a preset number of requests or reaching a preset time interval, execute the adjustment of the distribution probability to ensure the real-time and accuracy of the adjustment process.

[0110] It should be noted that in the distributed distribution scheme, service nodes regularly broadcast the computing power information, network status information, etc. of their own nodes. The access node adopts the principle of processing each arriving request immediately, and independently calculates the request distribution strategy based on the collected computing power service, network topology, and historical service request information of each node, and distributes and forwards the request to the corresponding service node according to the decision result. If the number of requests arriving at each access node is uneven, for example, 100 requests arrive at access node 1 within 1 second, while 1000 tasks arrive at node 2, the distributed request distribution algorithm proposed in this patent also requires the access node to broadcast the historical request arrival information of its own node to achieve a better estimate of the number of requests.

[0111] Furthermore, in the specific implementation scenario of the present invention, three feedback update methods for the distribution probability are adopted:

[0112] 1) Since the goal of balanced distribution is to minimize the variance of the load values of heterogeneous service nodes in the network, we can simulate distributing requests to each service node, calculate the change in the variance of the loads of all service nodes after distributing to each service node respectively, and update the distribution probability of the service node according to the change in the load variance: the more the load variance decreases after distributing to a certain service node, the more it indicates that increasing the distribution probability of this service node can reduce the load variance.

[0113] 2) The distribution algorithm should also minimize network latency as much as possible. The smaller the network latency from an access node to a certain service node, the distribution probability of this service node should be appropriately increased.

[0114] 3) For the inconsistent processing performance of each service node, assume that the maximum number of parallel request processing of service node i is , and at this time the number of requests to be processed by node A is , then the ratio of the number of requests of node i to the maximum number of parallel processing can more accurately reflect the load situation of the service node. If this ratio is close to 1, it can be considered that the load of this service node is heavy, and the distribution probability of this service node should be reduced.

[0115] By comprehensively considering the above three methods, the feedback update setting of the distribution probability is determined by the comprehensive consideration of three kinds of information: the change in variance, latency, and the ratio of the current service node task quantity to the maximum parallel quantity of the service node.

[0116] In a specific embodiment of the present invention, the global distribution policy is updated based on feedback information. Specifically,

[0117] regularly receive the latest computing power information and network status information broadcast by the service node;

[0118] collect the local request processing situation and distribution effect of the access node, including average latency and load balancing situation;

[0119] Based on the feedback information, dynamically adjust the distribution parameters in the next cycle, including the number of requests and the latency weighting coefficient , to improve the load balancing and latency optimization effect of request distribution.

[0120] Further, after the access node receives the global distribution policy, or when the local processing reaches a preset number of requests, or at a preset time interval, local information is updated. Specifically,

[0121] calculate the total number of requests that all access nodes have completed forwarding;

[0122] calculate the average distribution probability based on the negative feedback mechanism, and update the computing power information and network status information of the service node stored locally by the access node.

[0123] Improve the response speed and adaptability of distributed decision-making through periodic local information updates.

[0124] It should be noted that the specific method for estimating and updating the service node information stored locally is as follows: Every time an access node finishes forwarding 100 requests (this quantity can be adjusted according to the actual situation), or when seconds have passed since the access node last received the service node broadcast information packet, it is estimated that the number of requests that all access nodes have completed forwarding at this time is (If the number of tasks reaching each access node is relatively uniform, then ; if it is not uniform, then The estimation requires analyzing the comparison characteristics of the number of requests reaching different access nodes to better estimate the number of requests. This algorithm assumes that it is relatively uniform). Then, each access node can independently estimate the average distribution probability of all access nodes to different service nodes according to the above three negative feedback calculation factors, and use this average distribution probability to update the service node information stored locally. Finally, update the average distribution probability.

[0125] In a specific implementation scenario of the present invention, use to represent the current number of tasks of service node , use the maximum parallel task number of service node , use to represent the available bandwidth of service node . The pseudo-code of the request distribution algorithm based on negative feedback and with local information estimation and update is as follows: Algorithm2: Request Distribution Algorithm Based on Negative Feedback and with Local Information Estimation and Update Input: Environmental parameters: Number of access nodes ; Number of service nodes ; Request distribution information of other nodes that have been completed when the request distribution is uneven Computing power parameters: Processing power index; Service node information Network parameters: Request bandwidth requirement ; Transmission delay ; Service node information update flag Request number I Algorithm parameters: Average distribution probability ; Service node information stored locally ; Weighting factor Return: (Service node information stored locally), , the target distribution service node A requested 1: if ( then / / Received an updated broadcast information packet 2: = / / Update the service node information stored locally 3: / / Reset the request number 4: / / Reset the average distribution probability 5: end if 6: / / Initialize the distribution probability according to the processing capacity index 7:Calculate the load value of service node i = 8: Calculate the ratio of the number of requests of service node i to the maximum parallel processing number 9:Calculate the load variance 10:for ( = 1 to ) 11: = 12: According to Calculate the obtained load variance 13: / / Calculate the load variance change amount 14:end for 15:Perform a normalization operation on and initialize the distribution probability ), initialize as an empty list 16:for ( = 1 to ) 17: 18:end for 19:for ( = 1 to ) 20:if ( then 21: / / Eliminate service nodes that do not meet the request bandwidth requirements 22: end for 23: / / Recalculate the distribution probability 24: Calculate a random number between 0 and 1 According to the probability Make a decision 25: A = 26: if ( then 27: / / Estimate and update the task quantity of the service node 28: Utilize Update 29: for ( = 1 to ) 30: / / Update the average distribution probability 31: end for 32: , 33: end if 34:

[0126] In the second embodiment of the present invention, the present invention provides a computing power network service request balanced distribution system, as Figure 2 shown, the system includes a global policy generation module 1, a local policy adjustment module 2, and a feedback optimization module 3;

[0127] The global policy generation module 1 is used to collect global information within the current preset period, including request flow information of each access node, computing power information, network status information, and historical load information of each service node, and generate a global distribution policy based on the global information through a centralized convex optimization method;

[0128] The local policy adjustment module 2 is used to send the global distribution policy to each access node, and after the access node receives the global distribution policy, according to the local real-time request processing situation, within a preset time sequence interval, dynamically adjust the distribution policy through a negative feedback mechanism, and forward it to the service node according to the adjusted distribution policy;

[0129] The feedback optimization module 3 is used to receive feedback information from the service node and the access node in real time, and update the global distribution policy based on the feedback information to improve the distribution accuracy and load balancing effect in the next cycle.

[0130] In the third embodiment of the present invention, an electronic device is provided, including a memory and a processor. It is characterized in that a computer program that can run on the processor is stored in the memory, and when the program is executed on the processor, the steps in the above-mentioned computing power network service request balanced distribution method are implemented.

[0131] In the fourth embodiment of the present invention, a storage medium is provided. The storage medium stores a computer program, and it is characterized in that when the computer program is executed by a processor, the steps in the above-mentioned computing power network service request balanced distribution method are implemented.

[0132] In summary, a computing power network service request balanced distribution method and system provided by the present invention support centralized and distributed dual-mode scheduling, achieve a global optimal strategy through convex optimization, combine a negative feedback mechanism to adjust the distribution probability in real time, and improve the cross-layer resource scheduling efficiency and dynamic adaptability. Multi-objective optimization is introduced to flexibly balance load balancing and network delay, and different scenarios are adapted by adjusting the delay coefficient. In the distributed mode, the negative feedback control enhances the ability to handle burst traffic, the local estimation mechanism reduces the information synchronization overhead, alleviates the delay impact, and significantly improves the resource utilization rate, service quality, and system robustness.

[0133] It should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0134] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described modules can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0135] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0136] In addition, in each embodiment of the present application, the various functional modules can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of a combination of hardware and software functional modules.

[0137] The integrated module implemented in the form of software functional modules can be stored in a computer-readable storage medium. The above-mentioned software functional modules are stored in a storage medium and include several instructions for causing a computer system (which may be a personal computer, a server, or a network system, etc.) or a processor to execute some steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. However, such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present application.

Claims

1. A method for balanced distribution of computing power network service requests, characterized in that: include, Collect global information within the current preset period, including request flow information of each access node and computing power information, network status information and historical load information of each service node, and generate a global distribution strategy based on the global information through a centralized convex optimization method; The global distribution strategy is sent to each access node. After the access node receives the global distribution strategy, it dynamically adjusts the distribution strategy through the negative feedback mechanism within the preset time interval according to the local real-time request processing situation, and forwards it to the service node according to the adjusted distribution strategy; Receive feedback information from service nodes and access nodes in real time, and update the global distribution strategy based on the feedback information to improve distribution accuracy and load balancing effect in the next cycle.

2. The method for balanced distribution of computing power network service requests according to claim 1, characterized in that: The step of "generating a global distribution strategy based on global information through a centralized convex optimization method" includes: Define the access node router as ,in, , r is the access node, is the total number of access nodes; Collect request flow information of each access node and computing power information of each service node; wherein the request flow information includes the number and type of requests, and the computing power information includes the service type, the current number of pending requests, and the maximum number of parallel processing requests; Based on historical load information and network topology, the machine learning model is used to predict the number of requests for each access node in the next cycle; Based on the centralized convex optimization method, the optimization problem is established with the goal of minimizing the variance of the number of comprehensive requests of service nodes, generating a global distribution strategy and dynamically adjusting the distribution probability; The comprehensive request quantity of the service node is expressed as: in, is the total number of requests for service node i, is the maximum number of parallel processing requests for service node i, is the number of pending requests currently being processed by service node i, is the number of requests to access node r in the next cycle predicted by the machine learning model, is the distribution probability from access node r to service node i, and satisfies and .

3. The method for balanced distribution of computing power network service requests according to claim 2, characterized in that: Based on the centralized convex optimization method, the goal is to minimize the variance of the number of comprehensive requests of service nodes, and the optimization problem with distribution probability as the optimization variable is expressed as: in, is the total number of requests for service node m, is the network delay from access node r to service node i, is the delay weighting coefficient, is the average value of network delay, is the total number of service nodes.

4. The method for balanced distribution of computing power network service requests according to claim 3 is characterized in that: The “dynamically adjusting the distribution strategy through a negative feedback mechanism within a preset time interval according to the local real-time request processing situation” includes: After receiving the global distribution strategy, the access node performs preliminary distribution on each arriving request according to the distribution probability in the global distribution strategy; During the request forwarding process, feedback information from service nodes is collected, including load status and delay information, and the initial distribution strategy is dynamically adjusted based on the service node load ratio and network delay based on the negative feedback mechanism; Among them, the load ratio of service node i is , is the number of pending requests for service node i.

5. The method for balanced distribution of computing power network service requests according to claim 4 is characterized in that: Dynamically adjusting the preliminary distribution strategy through the negative feedback mechanism includes: Simulate the distribution of requests to each service node and calculate the change in the load variance of the service nodes in the entire network after distribution; The distribution probability of the preliminary distribution strategy is adjusted according to the change in the load variance, the service node load ratio and the network delay. , so that the distribution results tend to global load balancing; adjustment strategies include, When the load variance decreases, the load ratio is lower than the threshold, and the delay is shorter than the threshold, the predefined increase ; When the load variance increases, the load ratio is higher than the threshold, or the duration is longer than the threshold, the load is reduced according to the predefined ; When processing a preset number of requests or reaching a preset time interval, the distribution probability is executed adjustments to ensure the real-time and accuracy of the adjustment process.

6. The method for balanced distribution of computing power network service requests according to claim 1, characterized in that: The “updating the global distribution strategy based on feedback information” includes: Regularly receive the latest computing power information and network status information broadcast by service nodes; Collect local request processing and distribution effects of access nodes, including average latency and load balancing; Based on the feedback information, dynamically adjust the distribution parameters in the next cycle, including the number of requests and delay weighting factor , in order to improve the load balancing and latency optimization effects of request distribution.

7. The method for balanced distribution of computing power network service requests according to claim 1, characterized in that: Also includes, After the access node receives the global distribution strategy, or when the local processing reaches a preset number of requests, or a preset time interval, the local information update is performed, including: Calculate the total number of requests that have been forwarded by all access nodes; Calculate the average distribution probability based on the negative feedback mechanism, and update the service node computing power information and network status information stored locally on the access node; Improve the responsiveness and adaptability of distributed decision-making through periodic local information updates.

8. A computing power network service request balanced distribution system, characterized by: It includes global strategy generation module, local strategy adjustment module and feedback optimization module; The global strategy generation module is used to collect global information within the current preset period, including request flow information of each access node and computing power information, network status information and historical load information of each service node, and generate a global distribution strategy based on the global information through a centralized convex optimization method; The local policy adjustment module is used to send the global distribution policy to each access node, and after the access node receives the global distribution policy, dynamically adjust the distribution policy through a negative feedback mechanism within a preset timing interval according to the local real-time request processing situation, and forward it to the service node according to the adjusted distribution policy; The feedback optimization module is used to receive feedback information from the service node and the access node in real time, and update the global distribution strategy based on the feedback information to improve the distribution accuracy and load balancing effect in the next cycle.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory stores a computer program that can be run on the processor, and when the program is executed on the processor, the steps in the method for balanced distribution of computing power network service requests as described in any one of claims 1-7 are implemented.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps in the method for balanced distribution of computing power network service requests as described in any one of claims 1 to 7 are implemented.

Images