Industrial internet-oriented edge resource scheduling method and system

Abstract

The invention discloses an industrial internet-oriented edge resource scheduling method and system, and relates to the field of industrial internet, and the method comprises the steps: predicting the user request amount in a future time period through a SparseTSF model; calculating the minimum number of servers on the premise of full service quality by using an M / M / c queuing model based on the user request quantity in the future time period; determining a predicted activated server set based on energy efficiency optimization; and generating a server start-stop instruction according to the predicted activated server set and the current activated server set, and carrying out resource allocation. According to the invention, the future load can be predicted, the system power consumption is reduced through resource planning, the resource utilization rate is optimized, and the service quality is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of edge technology in industrial internet, and more particularly to an edge resource scheduling method and system for industrial internet. BACKGROUND

[0002] The current industrial internet edge system is mainly composed of a three-level architecture of "terminal device-edge server cluster-cloud platform": the terminal device (such as industrial sensor, automatic driving terminal) generates request data, the edge server cluster is responsible for processing data nearby to reduce delay, and the cloud platform only processes non-real-time global data. With the rapid development of industrial internet and edge computing, edge server clusters are widely used in intelligent manufacturing, Internet of Vehicles, smart factories and other high real-time requirement scenarios.

[0003] However, the intensive deployment of edge nodes leads to a sharp increase in energy consumption, which has become a key problem restricting the sustainable development of the system. In the prior art, most resource scheduling methods rely on real-time load monitoring and dynamic adjustment strategies, such as online decision-making methods based on reinforcement learning or heuristic greedy algorithms. These methods can reduce energy consumption to a certain extent, but still have many problems. For example, the existing scheduling relies on real-time decision-making, lacks the ability to predict user requests, cannot predict load changes in advance, and is difficult to cope with the spatiotemporal dynamics of user requests, resulting in scheduling lag; at the same time, it also leads to the long-term high-load operation of some servers and the idle state of the rest, resulting in a sharp increase in energy consumption and waste of resources. Secondly, the existing methods lack adaptation to edge characteristics, such as user-server proximity constraints and resource multi-dimensionality, which are not fully modeled, resulting in large fluctuations in service quality (such as response time).

[0004] Therefore, how to predict future load, reduce system power consumption, optimize resource utilization and improve service quality is a problem that those skilled in the art need to solve. SUMMARY

[0005] Therefore, the present application provides an edge resource scheduling method and system for industrial internet to solve the above problems.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions: The present application discloses an edge resource scheduling method for industrial internet, comprising: predicting the user request amount in the future time period through a SparseTSF model; based on the user request amount in the future time period, using an M / M / c queuing model to calculate the minimum number of servers under the premise of full service quality; and determining a predicted active server set based on energy efficiency optimization; According to the predicted active server set and the current active server set, a server start-stop instruction is generated to perform resource allocation.

[0007] Further, the SparseTSF model is used to predict the user request amount in the future time period, specifically including: Collecting original request amount / load data of the industrial internet edge server; Performing multi-scale decomposition on the original request amount / load data to obtain k time domain features; Performing cross-cycle convolution on the decomposed time domain features to obtain high-order features; Performing gated fusion on the high-order feature map to obtain k weights of the features; Obtaining current request amount / load data of the industrial internet and performing multi-scale decomposition to obtain k current time domain features, and based on the weights of the features, weighted calculation is performed to obtain the user request amount in the future time period .

[0008] Further, the formula of the multi-scale decomposition is: ; Wherein, represents k time domain features, represents inverse Fourier transform, represents a frequency domain mask matrix, represents Fourier transform, Λ represents original request amount / load data of the edge server.

[0009] Further, the formula of the cross-cycle convolution is: ; Wherein, represents high-order features, represents input features, represents weights, is an activation function, represents the k layer bias vector, which shifts the feature distribution to make it more easily activated by ReLU, and the ReLU function filters effective features.

[0010] Further, the formula of the gated fusion is: ; Wherein, represents the weight of the k feature, represents global average pooling, U and v are learnable parameters, and All of these are activation functions.

[0011] Furthermore, the calculation of the minimum number of servers required to achieve full service quality using the M / M / c queuing model specifically includes: First, determine the user request volume for the future time period. Server speed Target average queuing delay ; Then, initialize the parameters, set c=1 ; Finally, iterative calculations are performed to determine the minimum number of servers. The iterative calculation steps are as follows: Step 1: Calculate system utilization ,like If yes, proceed to step 4; otherwise, proceed to step 2. Step 2: Calculate the Erlang C formula: ; Step 3: Calculate the average queuing time: ;like Then let =c Exit; otherwise proceed to step 4. Step 4: c = c + 1 Return to step 1 in, This represents an Erlang C formula used to calculate the probability that a new request will have to wait when all servers are busy in an M / M / c queuing model. This indicates the waiting time for the request in the queue.

[0012] Furthermore, the step of determining the predicted activation server set based on energy efficiency optimization specifically involves: sorting the server set S by energy efficiency ratio and selecting the top... The most energy-efficient servers constitute the activation set The formula for the energy efficiency ratio is: ; in, In server set S, the first... j The energy efficiency ratio of a server d This indicates the number of resource dimensions (such as CPU, memory, bandwidth, etc.). Indicates the first j The first server k Class resource capacity, This represents the operating energy consumption of the j-th server. .

[0013] Further, the generation server start-stop instruction, specifically: for each server : if and , send start instruction; if and , send shutdown instruction; wherein represents the current active set.

[0014] Further, the resource allocation, specifically includes: According to the proximity constraint guarantee formula: , wherein represents the server to which the user is allocated at time t; the user is allocated to the server set in its coverage range; Within each server set, the local resource allocation is performed by using the dominant resource fair allocation algorithm, specifically: First, calculate the dominant resource demand of each user request, the formula is: ; wherein represents the dominant resource demand of the user , and represents the demand of the user for the first k type of resource; and the user requests are sorted in descending order ; Then, the user is allocated to the server with the lowest current utilization rate and satisfying the capacity constraint: , the capacity constraint needs to satisfy: ; represents the total demand of all users t allocated to the server j for the resource type at time k . The constraint ensures that the resource utilization rate of the server does not exceed 90% of its capacity at any time, thereby avoiding performance degradation or service interruption due to high load, and reserving 10% of the resource redundancy, which helps to dynamically absorb the impact of transient request peaks or prediction errors. Wherein, represents the optimal server selection, represents the active server set, represents the demand of the user for the first t type of resource at time k .

[0015] The application also discloses an edge resource scheduling system for industrial internet, comprising: Prediction layer: predict the user request volume in the future time period through SparseTSF model; Resource planning layer: based on the future time period user request volume, using M / M / c queuing model, calculate the minimum server quantity under the premise of full service quality; and based on energy efficiency optimization to determine the predicted active server set; Execution layer: according to the predicted active server set and the current active server set, generate server start-stop instructions, and perform resource allocation.

[0016] Via the above technical solution, compared with the prior art, the application provides an edge resource scheduling method and system for industrial internet, and the specific effects are as follows: (1) The total energy consumption of the system is greatly reduced: experimental results show that in a 7-day continuous test covering different load periods on weekdays and weekends, the total energy consumption of the method is reduced by 78.6% compared with the benchmark strategy (AlwaysOn, all servers are always on). Compared with the current advanced online scheduling method based on deep reinforcement learning (pDRL), EAPA can still achieve about 60% energy consumption reduction. This great improvement is directly due to the accurate prediction ability of the application, which enables the system to turn off redundant servers in advance, avoiding a large amount of idle energy (IdleEnergy) and the cost of frequent start-stop.

[0017] (2) Resource utilization optimization: the method realizes "on-demand supply" of server resources through prediction-driven fine management. The average server utilization rate is measured to be 48.2%, which is much higher than the AlwaysOn strategy (always on but load is uneven, utilization rate is usually less than 20%) and the random start-stop strategy (utilization rate fluctuates greatly, with an average of only ~25%). This proves that the application effectively solves the resource fragmentation problem of "part of the server overload, part of the server idle" commonly existing in edge computing, and realizes the intensive use of resources.

[0018] (3) Reliable guarantee of service quality (QoS): during the entire test period, the average response time of all user requests is stably controlled within the service level agreement (SLA) threshold of 100ms, and the SLA satisfaction rate reaches 100%. Even during the request peak period (such as the 12th time slot, the request volume reaches 2500 / min), through prediction and early deployment of 5 servers, the response time can still be kept at 100ms. This completely overturns the traditional concept that energy saving must sacrifice performance, and proves that the pre-configuration mechanism based on high-precision prediction can achieve energy saving while guaranteeing extreme user experience.

[0019] (4) System stability is enhanced: compared with the strategy that completely relies on real-time reaction and leads to frequent switching of servers, the method based on periodic prediction for scheduling significantly reduces the frequency of server start-stop, enhances the working life of hardware devices and the stability of the system as a whole.

[0020] (5) Direct electricity expenditure decreases: Taking a small 20-server edge cluster as an example, assuming the electricity price is 1 yuan per degree, the annual electricity bill can reach tens of thousands of yuan by using the AlwaysOn strategy. After applying the method, a considerable proportion of annual electricity bill can be saved. For large industrial parks or nationwide deployment of edge node networks, the economic benefits will grow exponentially.

[0021] (6) Equipment investment return rate increases: Higher resource utilization means that the same computing task can be completed with fewer physical servers, or more business can be carried on existing servers, delaying the need to purchase new servers due to business growth, reducing the enterprise's fixed asset investment (CAPEX).

[0022] (7) Operation and maintenance costs decrease: The increase in system automation reduces the dependence on real-time monitoring and intervention by humans. At the same time, the optimization of server start-stop frequency also reduces the failure rate of equipment, indirectly reducing maintenance costs and downtime risks.

[0023] (8) Enable high-quality development of industrial Internet: The invention solves the energy consumption bottleneck problem, making it possible and sustainable to deploy large-scale low-latency, high-reliability edge computing services in industrial sites, providing a solid technical foundation for the digital transformation of key industries such as intelligent manufacturing and autonomous driving, and having wide social application value.

[0024] In summary, the invention can predict future loads, reduce system power consumption, optimize resource utilization, and improve service quality through resource planning. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0026] Figure 1 The overall flowchart provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0028] The embodiment of the application discloses an edge resource scheduling method for an industrial internet, comprising: predicting a user request amount in a future time period through a SparseTSF model; based on the user request amount in the future time period, calculating a minimum server number under a premise of full service quality by using an M / M / c queuing model, and determining a predicted active server set based on energy efficiency optimization; generating a server start-stop instruction according to the predicted active server set and a current active server set, and performing resource allocation.

[0029] Specifically, Figure 1 The core structure of the application is shown, wherein a data acquisition and preprocessing module is deployed on an edge node, is responsible for real-time collection of historical load data, and performs preprocessing such as standardization and denoising. After receiving the preprocessed historical data, multi-scale spatiotemporal feature extraction and fusion are performed through three units of multi-scale decomposition, cross-period convolution and gate fusion. In the resource planner, the prediction result is received, and the minimum server number satisfying the QoS delay requirement is calculated based on the M / M / c queuing model . According to the energy efficiency ratio of each server , the optimal server is selected from the server cluster to form an active set . The dynamic start-stop controller receives the active set At, generates specific server start-stop instructions (such as starting server , and putting server to sleep), and issues them to the hardware through the underlying interface. The resource allocation executor (DRF) allocates the real-time arriving user tasks to the active servers according to the dominant resource fairness algorithm, while meeting the proximity constraint and capacity constraint.

[0030] In a specific embodiment, the SparseTSF model is used to predict the user request amount in the future time period, specifically comprising: collecting original request amount / load data of the industrial internet edge server; performing multi-scale decomposition on the original request amount / load data to obtain k time domain features; performing cross-period convolution on the decomposed time domain features to obtain high-order features; performing gate fusion on the high-order feature maps to obtain k weights of the features; obtaining current request amount / load data of the industrial internet, performing multi-scale decomposition to obtain k current time domain features, and calculating the user request amount in the future time period based on the weights of the features .

[0031] Taking an intelligent manufacturing park edge computing platform as an example, a multi-access edge computing (MEC) system is deployed, which includes an edge server cluster, terminal equipment, network equipment, and an edge control node. All edge servers are connected to the SDN switch through the network interface, the edge control node is connected to the SDN switch and the out-of-band management interface of all edge servers through the network, used to issue server start and stop instructions and monitor server status, and the terminal equipment accesses the network through a wireless AP, and its data traffic is directed to the currently activated edge server via the SDN switch.

[0032] The system collects historical load data, inputs the SparseTSF model for training and prediction. Every 24 minutes is a scheduling period, and according to the prediction result, 5-8 servers are dynamically activated, and the computing tasks are allocated through the DRF strategy. In actual operation, the system fine-tunes the server state according to real-time load, and automatically hibernates part of the servers when the load is below the threshold. This method can be realized through software-defined network (SDN) and edge controller cooperation, and supports API interface and existing MEC platform integration.

[0033] In one specific embodiment, the formula of multi-scale decomposition is: ; Wherein, represents k a time domain feature, represents an inverse Fourier transform, represents a frequency domain mask matrix, represents a Fourier transform, Λ represents the original request volume / load data of the edge server. Multi-scale decomposition converts the original request volume / load data of the edge server into frequency domain features through Fourier transform, retains key frequency bands with a frequency domain mask matrix, the purpose is to retain key frequency bands and filter noise, and then restores the filtered frequency domain data to the time domain through inverse Fourier transform.

[0034] In one specific embodiment, the formula of cross-cycle convolution is: ; Wherein, represents a high-order feature, represents the k th time domain feature obtained by multi-scale decomposition, represents a weight, is an activation function, represents the k layer bias vector, which shifts the feature distribution to make it more easily activated by ReLU. Cross-cycle convolution extracts the periodicity of industrial requests (such as shift cycles of intelligent manufacturing) through a convolution layer (input features, weights), and filters effective features with a ReLU activation function.

[0035] In one specific embodiment, the formula of gated fusion is: ; wherein, represents the weight of the i-th feature, k represents global average pooling, and U are learnable parameters, v and are activation functions.

[0036] In one specific embodiment, the minimum number of servers under the premise of full quality of service is calculated by using M / M / c queuing model, specifically including: First, determine the user request volume in the future time period , server rate , and target average queuing delay ; Then, initialize parameters, set c=1 ; Finally, perform iterative calculation to determine the minimum number of servers ; the steps of iterative calculation are: Step 1: calculate system utilization , if , go to step 4; otherwise, go to step 2; Step 2: calculate the Erlang C formula: ; Step 3: calculate the average queuing time: ; if , let =c , exit; otherwise, go to step 4; Step 4: c = c + 1 , return to step 1 wherein, represents the Erlang C formula for calculating the probability that a new request needs to wait when all servers in the M / M / c queuing model are busy, represents the waiting time of the request in the queue.

[0037] In one specific embodiment, the predicted active server set is determined based on energy efficiency optimization, specifically: from the server set S, sort the servers according to the energy efficiency ratio, and select the top servers with the highest energy efficiency to form the active set ; the formula of energy efficiency ratio is: ; wherein, ​In server set S, the first... j The energy efficiency ratio of a server d This indicates the number of resource dimensions (such as CPU, memory, bandwidth, etc.). Indicates the first j The first server k Class resource capacity, Indicates the first j Power consumption of the server .

[0038] In one specific embodiment, generating server start / stop instructions involves: for each server :like and Then send a start command; if and If so, a shutdown command is sent; where This indicates the currently active set.

[0039] In one specific embodiment, resource allocation includes: According to the proximity constraint guarantee formula: ,in Indicates user In time t The assigned server; will the user The set of servers assigned to its coverage area middle; Within each server set, a dominant resource fairness allocation algorithm is used for local resource allocation, specifically: First, calculate the dominant resource requirement for each user request using the following formula: ;in Indicates user The dominant resource demand, Indicates user For the first k The demand for class resources; and according to Sort user requests in descending order; Then, users are greedily assigned to the server with the lowest current utilization that meets the capacity constraint: The capacity constraint must satisfy: This indicates that at time t, all servers assigned to the server... j users For resource types k The total demand. This constraint ensures that the server's resource utilization does not exceed 90% of its capacity at any given time, thereby avoiding performance degradation or service interruption due to excessive server load. Maintaining 10% resource redundancy helps dynamically absorb the impact of instantaneous request spikes or prediction errors. Among these, represents the optimal server selection, represents the active server set, represents the user at time t the first k resource of the class.

[0040] The resource allocation is performed by the SDN controller according to updating the server table, and routing the newly arrived user request to the active server.

[0041] The embodiments of the present application also disclose an edge resource scheduling system for the industrial internet, comprising: a prediction layer: predicting the user request amount in the future time period through a SparseTSF model; a resource planning layer: calculating the minimum number of servers under the premise of full service quality based on the user request amount in the future time period through an M / M / c queuing model, and determining the predicted active server set based on energy efficiency optimization; an execution layer: generating server start-stop instructions according to the predicted active server set and the current active server set, and performing resource allocation.

[0042] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other. For the device disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0043] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An edge resource scheduling method for the industrial internet, characterized in that, The method comprises the following steps: predicting user request volume in a future time period through a SparseTSF model; calculating minimum server quantity under the premise of meeting service quality based on the user request volume in the future time period by using an M / M / c queuing model; determining a predicted active server set based on energy efficiency optimization; generating a server start-stop instruction and performing resource allocation according to the predicted active server set and a current active server set. 2.The edge resource scheduling method for industrial internet of claim 1, wherein, The step of predicting user request volume in a future time period through a SparseTSF model specifically comprises the following steps: collecting original request volume / load data of an industrial internet edge server; performing multi-scale decomposition on the original request volume / load data to obtain k time domain features; performing cross-period convolution on the time domain features obtained through decomposition to obtain high-order features; performing gated fusion on the high-order feature map to obtain a feature k weight of the feature; Obtain the current request volume / load data of the industrial internet, and perform multi-scale decomposition to obtain the current k time domain features, and based on the weight of each feature, weighted calculation is performed to obtain the user request volume of the future time period . 3.The edge resource scheduling method for industrial internet of claim 2, wherein, the formula of the multi-scale decomposition is: ； wherein denotes k temporal features, denotes an inverse Fourier transform, denotes a frequency domain mask matrix, denotes a Fourier transform, Λ denotes raw request volume / load data for the edge server.

4. The edge resource scheduling method for industrial internet according to claim 2, wherein, the formula of the cross-period convolution is: ； wherein, represents a high-order feature, represents an input feature, represents a weight, is an activation function, represents the first k layer bias vector.

5. The edge resource scheduling method for industrial internet according to claim 2, wherein, the formula of the gated fusion is: ； wherein, denotes the weight of the k th feature, denotes global average pooling, U and v are learnable parameters, and are activation functions.

6. The edge resource scheduling method for industrial internet according to claim 1, wherein, The step of calculating minimum server quantity under the premise of meeting service quality based on the user request volume in the future time period by using an M / M / c queuing model specifically comprises the following steps: First, determine future time period user request volume , server rate , target average queue delay ; Then, the parameters are initialized, set c=1 ; Finally, an iterative calculation is performed to determine the minimum number of servers ; the step of iterative calculation is: Step 1 : Compute system utilization If then go to Step 4; else go to Step 2; Step 2: Calculate the Erlang C formula: ; Step 3: Calculate the average queue time: ; if , then let =c , exit; otherwise go to Step 4; Step 4: c = c + 1 , return to step 1 wherein, denotes the Erlang C formula for calculating the probability that a new request has to wait when all servers in the M / M / c queuing model are busy, denotes the waiting time of a request in the queue.

7. The edge resource scheduling method for industrial internet according to claim 6, wherein, The server set is determined based on energy efficiency optimization, specifically, servers with the highest energy efficiency ratio are selected from the server set S to form the active set The server set is determined based on energy efficiency optimization, specifically, servers with the highest energy efficiency ratio are selected from the server set S to form the active set The formula of the energy efficiency ratio is: ； in, In server set S, the first... j The energy efficiency ratio of a server d Indicates the number of resource dimensions. Indicates the first j The first server k Class resource capacity, Indicates the first j Energy consumption of the server.

8. The edge resource scheduling method for industrial internet according to claim 7, wherein, The generating server start-stop instruction, specifically, for each server : if and , a start instruction is sent; if and , a shutdown instruction is sent; wherein represents the current active set. 9.The edge resource scheduling method for industrial internet of claim 7, wherein, The step of performing resource allocation specifically comprises the following steps: According to the proximity constraint guarantee formula: where denotes the user assigned to the server at time t ; the user is assigned to the set of servers in its coverage ; performing local resource allocation in each server set by using a dominant resource fair allocation algorithm, specifically as follows: First, the dominant resource requirement of each user request is calculated as follows: wherein denotes the dominant resource requirement of user , denotes the requirement of user for the kth type of resource; and the user requests are ordered in decreasing order of ; Then, the users are greedily assigned to the server with the lowest current utilization that satisfies the capacity constraint: , the capacity constraint needs to be satisfied: ; denotes the sum of the demands of all users t assigned to server j for resource type at time k ; where, denotes the optimal server selection, denotes the active server set, denotes the demand of user for the t th resource type at time k .

10. An edge resource scheduling system for industrial internet, applying the edge resource scheduling method for industrial internet according to any one of claims 1 to 9. The method comprises the following steps: a prediction layer for predicting user request volume in a future time period through a SparseTSF model; a resource planning layer for calculating minimum server quantity under the premise of meeting service quality based on the user request volume in the future time period by using an M / M / c queuing model; determining a predicted active server set based on energy efficiency optimization; an execution layer for generating a server start-stop instruction and performing resource allocation according to the predicted active server set and a current active server set.

Images