Cloud-edge cooperative system computing power-electric power joint optimization scheduling method considering heterogeneous characteristics

Abstract

The invention relates to a computing power-electric power joint optimization scheduling method for a cloud-edge collaborative system considering heterogeneous characteristics, and the method comprises the steps: taking an upper cloud data center as a scheduling main body, and carrying out the scheduling of the computing power-electric power joint optimization scheduling of the cloud-edge collaborative system based on the number of delay tolerant tasks, the unloading amount of delay sensitive tasks and a local electric energy cost characteristic curve in each scheduling time period; by taking maximization of the total operating profit in the whole scheduling period as an optimization target, optimizing to obtain a calculation service unit cost signal of the scheduling period, and issuing the calculation service unit cost signal to each edge base station; and each edge base station at the lower layer receives the time-sharing electric energy cost signal from the power grid, independently optimizes the real-time task unloading amount by taking the maximization of the comprehensive utility as an optimization target in combination with the calculation service unit cost signal received from the upper layer, and uploads the real-time task unloading amount to the cloud data center. Compared with the prior art, the method has the advantages that the modeling heterogeneous characteristics of cloud and edge computing power facilities are considered, and computing power collaboration of a cloud-edge system is effectively realized.

Description

Technical Field

[0001] This invention relates to the fields of computing power scheduling and power scheduling technology, and in particular to a cloud-edge collaborative system computing power-power joint optimization scheduling method that takes into account heterogeneous characteristics. Background Technology

[0002] In recent years, the rapid development of the Internet of Things (IoT) and artificial intelligence (AI) has driven an exponential increase in computing power demand, prompting the expansion of computing infrastructure deployment, such as cloud data centers and edge base stations. To comprehensively improve service quality within a region, cloud data centers typically collaborate with multiple nearby edge base stations to provide high-precision, low-latency computing services to users within the area. This new computing service paradigm is known as cloud-edge collaborative computing. However, the operation of cloud-edge collaborative computing systems generates significant irregular power demand, exacerbating the pressure on the power system's supply-demand balance. Although operators have some ability to adjust computing task processing plans and change the power demand of computing infrastructure at different stages, possessing flexibility in time and space power scheduling, this is usually constrained by various factors such as computing service quality requirements, economic requirements, and technical conditions. Therefore, how to efficiently schedule computing power and power for computing resources hosted by heterogeneous computing infrastructure has become a hot research topic both domestically and internationally.

[0003] In terms of computing power scheduling, cloud-edge collaborative computing systems optimize resource allocation efficiency by designing task offloading and resource allocation strategies among different entities. One existing technology establishes mathematical models for users offloading tasks to both the cloud and edge sides and proposes an optimal economic operating scheme for the system. Based on this, another existing technology considers the mutual interference of task offloading signals, modifies the transmission rate model, and formulates a non-cooperative game strategy for task offloading by each user to reduce latency and total energy consumption costs. Yet another existing technology considers that cloud data centers cache some computing tasks, which are then offloaded to the edge side together with mobile users, proposing a minimum energy consumption strategy for the system to complete all task processing. However, the above works only consider cloud-edge systems handling a single type of computing task and use similar linear models to estimate server computing latency and energy consumption at the cloud and edge. In reality, cloud servers typically handle large-scale computing tasks with high complexity, such as big data analysis and offline rendering, while edge servers are mainly used for small-scale computing tasks that are highly sensitive to latency, such as data retrieval and online interaction. There are differences between the two in terms of task types, resource levels, and operational metrics. To differentiate between cloud and edge computing, one existing technique considers that while cloud data centers handle real-time offloaded tasks at the edge, they also process another set of computationally intensive tasks, establishing a queuing computing model in the cloud. However, it does not investigate the impact of resource constraints on the efficiency of cloud-edge computing services. Furthermore, the above work fixes the electricity cost signal from the power system, lacking consideration of the interaction between computing power and electricity.

[0004] In fact, in the field of power dispatching, many studies have already modeled flexible loads for single computing resources, typically data centers, demonstrating the impact of computing power dispatching on the power system. Some works further focus on the joint dispatching of computing power and power in data centers. Chinese patent CN110084444A discloses a method for dispatching the power load of cloud data centers considering the randomness of natural resources. This method includes the following steps: predicting the total power load to be dispatched and the output of renewable energy; dispatching time-shiftable power loads at temporal and spatial scales; if the power load processing volume under the current dispatching strategy is less than the carrying capacity, calculating the power demand and waste heat utilization of each cloud data center, and simultaneously determining whether the power system and thermal system are in balance. If in balance, the power load dispatching scheme and total operating cost are recorded, thus obtaining all possible dispatching schemes, and the dispatching scheme with the lowest total operating cost is taken as the final dispatching scheme. However, this method does not consider the joint dispatching of computing power and power resources by cloud-edge collaborative computing systems. With the large-scale expansion of edge base stations, the total power demand on the edge side has a significant impact on power dispatching, necessitating research on joint computing and power dispatching strategies for cloud-edge collaborative systems.

[0005] Based on the above analysis, the research on the collaborative optimization of computing power and electricity for heterogeneous computing infrastructure faces the following challenges: the mathematical model that takes into account the differences between cloud and edge computing service facilities and the heterogeneity of computing tasks is not yet complete; and there is a lack of comprehensive design architecture that can simultaneously optimize computing power scheduling within the cloud-edge collaborative system and power interaction outside the system. Summary of the Invention

[0006] The purpose of this invention is to provide a cloud-edge collaborative system computing power-power joint optimization scheduling method that takes into account heterogeneous characteristics.

[0007] The objective of this invention can be achieved through the following technical solutions: A cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics is disclosed. The cloud-edge collaborative system includes a centralized cloud data center and multiple geographically dispersed edge base stations. The upper-layer cloud data center acts as the scheduling entity. During each scheduling period, the following steps are performed: The cloud data center, as a significant factor influencing power costs, participates in the scheduling process. By predicting the local power cost characteristic curve, the coupling relationship between power demand and power cost is incorporated into the upper-layer optimization objective. Combining the time coupling constraints of the task queue, the number of delay-tolerant tasks to be processed in the corresponding scheduling period is planned. The offloading volume of delay-sensitive tasks from edge base stations in all sub-periods is summarized. Combining the number of delay-tolerant tasks and the offloading volume of delay-sensitive tasks, the total operating profit within the entire scheduling cycle is maximized as the optimization objective. The task volume constraints of delay-tolerant tasks, the first energy consumption constraint, and the consistency constraints of delay-sensitive tasks are used as constraints to optimize and obtain the unit cost signal of computing services for this scheduling period, which is then distributed to each edge base station. Each edge base station in the lower layer receives time-of-use energy cost signals from the power grid and combines them with computing service unit cost signals received from the upper layer. With maximizing its own comprehensive utility as the optimization goal, and with latency constraints, secondary energy consumption constraints, and latency-sensitive task consistency constraints as constraints, it independently optimizes its own real-time task offloading volume and uploads it to the cloud data center.

[0008] The optimization objective of the cloud data center is expressed as: , in, For the entire scheduling cycle T Total operating profit within, This is the hourly time index within the corresponding scheduling period. This represents the fixed reward obtained from handling delay-tolerant tasks. This represents the revenue gained from providing computing power services to the edge. It is the cost of purchasing electricity paid to the local power grid; , , in, For time period t The total number of latency-sensitive tasks received by the internal cloud data center. For time period t The unit cost of cloud data center computing services For time period t The unit cost of electricity settled by the local power grid. To purchase electricity, , , , , in, It is an index of the step function relating the settlement of electricity costs to the electricity purchase demand of cloud data centers. S This represents the total number of steps. and Corresponding to the steps s The upper and lower limits of the covered electricity quota; For section s The highest marginal cost; For section s Quota variables on; To determine whether the intersection point falls within the segment s 0-1 variables.

[0009] The specific workload constraints for the delay-tolerant task are as follows: All pending, latency-tolerant tasks form a task queue. The cumulative number of tasks processed in the queue is within a dynamic feasible region, represented as follows: , in, h From the initial time period to the time period t Time index; For time period h The lower bound of the workload for latency-tolerant tasks processed by the internal cloud data center. Indicates the end time t The cumulative amount of tasks that must be completed, upper limit The total number of tasks completed is represented as follows: , , , in, It is a binary indicator variable used to determine the time period. h Will the arriving task be during the time period? t Deadline reached; Time period t Arrival delay-tolerant task attributes include , , These represent the amount of task arrival and the maximum tolerable delay, respectively.

[0010] The first energy consumption constraint is specifically: Cloud data centers obtain electricity from the local power grid and purchase electricity. Equal to its total internal energy consumption: , in, This represents the total energy consumption of the cloud data center. For electrical energy utilization efficiency, This represents the total number of cloud servers. This refers to the static power of a single server. This refers to the dynamic power consumption coefficient of the cloud server. For cloud data centers during time periods t Total task processing volume, , For time period t The workload of delay-tolerant tasks. For time period t The total number of latency-sensitive tasks received by the cloud data center; To ensure service quality, the computing performance of cloud servers is evaluated based on the M / M / 1 queuing model, and the average service time does not exceed the upper limit. : , in, This indicates the service speed of the cloud server.

[0011] The specific consistency constraint for the latency-sensitive task is as follows: At different times t Within, the time is further divided into sub-time periods based on a preset time interval. For the time index of the sub-period, For time period t The number of sub-time periods is divided, and the offloading volume of all edge base stations is consistent with the receiving volume of the cloud data center: , in, Indicates edge base station i During the period t Sub-periods within t The amount of latency-sensitive tasks offloaded to cloud data centers. N For the number of edge base stations, For time period t The total number of latency-sensitive tasks received by the internal cloud data center.

[0012] The optimization objective of the edge base station is expressed as: , , , , in, For edge base stationsi Its own comprehensive utility For edge base stations i During the period t Fixed revenue earned from handling local latency-sensitive tasks; Represents edge base station i During the period t The cost of purchasing electricity paid to the power grid; It is an edge base station i During the period t The cost of computing services paid to cloud data centers; It is an edge base station i The cost weight of delay penalties; Indicates edge base station i During the period t Service delay penalty costs; This is the hourly time index within the corresponding scheduling period. T Indicates the scheduling period; For the time index of the sub-period, For time period t The number of sub-time periods; For time period t Time-of-use fixed electricity purchase cost, For edge base stations i During the period t Sub-periods within t Electricity purchased from the power grid, For time period t The unit cost of cloud data center computing services For edge base stations i During the period t Sub-periods within t The amount of latency-sensitive tasks offloaded to cloud data centers. This is the delay penalty coefficient. For edge base stations i During the period t Sub-periods within t Total delay.

[0013] The delay constraint is expressed as follows: , in, For edge base stations i During the period t Sub-periods within t Total delay, For edge base stations i During the period t Sub-periods within t Local computation latency, For edge base stationsi During the period t Sub-periods within t The transmission delay caused by task unloading Edge base station i During the period t Sub-periods within t Maximum tolerable delay; , , , in, For edge base stations i CPU cycle requirements per unit task For edge base stations i During the period t Sub-periods within t Latency-sensitive tasks processed locally Indicates edge base station i The calculation of the delay factor, For edge base stations i Number of edge servers For edge base stations i The service speed of the edge server For edge base stations i During the period t Sub-periods within t The amount of latency-sensitive tasks offloaded to cloud data centers. For edge base stations i Transmission distance, For edge base stations i The transmission rate.

[0014] The second energy consumption constraint is as follows: Edge base station i During the sub-period t The total energy consumption within the unit is equal to the electricity purchased from the grid. : , , , in, For edge base stations i During the period t Sub-periods within t Electricity purchased from the power grid, For edge base stations i During the period t Sub-periods within t Server operating energy consumption, For edge base stations i During the period t Sub-periods within t Task offloading and transmission energy consumption For edge base stations i The computing energy consumption coefficient of edge servers For edge base stations i Number of edge servers For edge base stations i The service speed of the edge server For edge base stations i CPU cycle requirements per unit task For edge base stations i During the period t Sub-periods within t Latency-sensitive tasks processed locally Indicates edge base station i Energy consumption of unloading unit computing tasks. For edge base stations i During the period t Sub-periods within t The amount of latency-sensitive tasks offloaded to the cloud data center.

[0015] The edge base station employs a second-order cone relaxation technique during the lower-layer optimization process. In the delay constraint Relaxation is a constraint based on an inequality: Each lower-level optimization problem is transformed into a standard second-order cone programming problem.

[0016] The proposed computing power-power joint optimization scheduling method employs a heterogeneous decomposition algorithm for distributed solution across upper and lower layers. This algorithm decomposes the original problem into independent master and slave problems by setting boundary variables. These boundary variables include a coordination cost vector and an edge-side aggregation unloading matrix. The coordination cost vector corresponds to the computing service price decided in the upper-layer problem, while the edge-side aggregation unloading matrix corresponds to the task unloading amount decided in the lower-layer problem. Thus, the master problem is: given the aggregation unloading matrix, solve a profit maximization problem covering the entire scheduling cycle and with cross-time-period coupling constraints to obtain a new coordination cost vector. The slave problem is: given the coordination cost vector, all edge base stations solve their respective hourly optimization problems in parallel. Within each hour, each edge base station progressively completes the second-order cone programming problem for each sub-time period, extracting the optimization results of the unloading variables from the solution set to collectively constitute the task unloading amount for that edge base station.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a cloud-edge collaborative system operation characteristic model that takes into account the heterogeneity of cloud and edge computing facilities. In cloud data centers, the time flexibility of cloud servers in handling large-scale computing tasks is quantified by task queuing; in edge base stations, the nonlinear relationship between edge computing performance and the amount of tasks processed is characterized, which more accurately reflects the performance saturation effect of edge servers due to limited computing resources.

[0018] (2) This invention designs a computing power-power joint scheduling strategy for a cloud-edge collaborative system. This strategy establishes a cloud-edge two-layer optimization model, with cloud optimization as the upper-layer problem and all edge-side optimizations as the lower-layer problem, clearly defining the independent decision-making behavior and interaction mechanism of each party. In addition, this strategy integrates the interaction constraints with the local power grid and considers the bidirectional impact between computing service scheduling and electricity settlement costs.

[0019] (3) For the non-convex and dual-layer coupled computing power-electricity collaborative scheduling model, a distributed solution method of first convexizing and then decomposing is proposed. First, the lower-level non-convex subproblems are made convex by using the second-order cone relaxation technique; then, the global problem is decomposed into the main problem and the subproblems are solved iteratively by applying the heterogeneous decomposition algorithm, which ensures the privacy of each stakeholder and the scalability of the model. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the server performance test curve of the present invention; Figure 2 This is a schematic diagram of the cloud-edge collaborative system of the present invention; Figure 3 This is the local electricity cost characteristic curve of the present invention; Figure 4 This is a schematic diagram of the k-th iteration interaction process in the solution process of this invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0022] With the increasing scale of cloud-edge collaborative system deployments, how to efficiently coordinate computing resource scheduling and power dispatch decisions has become an urgent problem to be solved. However, most existing studies have neglected the heterogeneous characteristics of cloud and edge computing facilities and lack a comprehensive consideration of the bidirectional coupling effect between computing power scheduling and power consumption. To this end, this paper proposes a computing power-power joint optimization scheduling strategy that takes into account the heterogeneous characteristics of cloud and edge. First, a cloud-edge collaborative system operation characteristic model considering the heterogeneity of cloud and edge computing facilities is established to characterize the time flexibility of cloud-side computing power scheduling and the saturation effect of edge-side server computing performance. Second, a cloud-edge two-layer optimization framework considering computing power-power collaborative scheduling is designed, clarifying the decision-making mechanisms of the cloud side and the edge side. Finally, for this non-convex two-layer coupling model, a distributed solution method of first convexizing and then decomposing is proposed to ensure the privacy of each subject and the scalability of the model.

[0023] (1) Mathematical model of cloud-edge collaborative system considering heterogeneous characteristics The cloud-edge collaborative system studied in this embodiment includes a centralized cloud data center and N The system comprises geographically dispersed edge base stations. The cloud data center serves as the core computing hub, centrally scheduling and managing all computing requests received from the cloud. Each edge base station operates independently, providing low-latency computing services to nearby users. When the edge computing load is overloaded or service quality degrades, the edge base stations strategically offload some tasks to the cloud data center. The cloud and edge are connected via high-speed fiber optic channels to ensure timely task processing and result return.

[0024] In this architecture, the system handles two different types of computing tasks: 1) Delay-tolerant tasks (DTTs): These tasks involve large amounts of computation and have relatively lenient completion deadlines. They are directly submitted to the cloud data center and added to the task queue for processing. 2) Delay-sensitive tasks (DSTs): These tasks involve smaller amounts of computation but have strict requirements for processing latency. To minimize task transmission latency, they are preferentially offloaded to nearby edge base stations for real-time processing.

[0025] To ensure the universality of cloud-edge system modeling, this embodiment makes the following assumptions: 1) Server type: Within the same infrastructure, all servers have the same physical attributes, i.e., all cloud servers are homogeneous, and all edge servers are homogeneous. 2) Edge server operating mode: To ensure high service quality at the edge, all servers within each edge base station remain operational. 3) User service agreement: This embodiment focuses on the scheduling optimization problem between cloud-edge service providers, assuming that end users and each service provider have already signed a fixed service cost agreement, and does not involve optimization on the user side.

[0026] Based on the above definitions and assumptions, this embodiment will perform mathematical modeling on the operational characteristics of cloud data centers and edge base stations, including task scheduling and energy consumption calculation.

[0027] Cloud data center task scheduling includes handling DTTs and DSTs offloaded from the edge. Let... For the hourly time index within the corresponding scheduling period, the time period t The reached DTT attributes include , These represent the number of arriving tasks and the maximum tolerable latency, respectively. All pending DTTs constitute a task queue, and the cumulative processing volume of tasks in the queue is within a dynamic feasible region, represented as: (1) in, h From the initial time period to the time period t Time index; For time period h The number of DTTs processed by the internal cloud data center. Lower bound Indicates the end time t The cumulative amount of tasks that must be completed, upper limit The total number of tasks completed is represented as follows: (2) (3) (4) In the formula, It is a binary indicator variable used to determine the time period. h Will the arriving task be during the time period? t Deadline has been reached.

[0028] For DSTs offloaded at the edge, the cloud data center will process them immediately. (Time period) t The total number of DSTs received by the internal cloud data center is Then the cloud data center during the time period t Total task processing volume Represented as: (5) In terms of energy consumption calculation, the energy consumption of a cloud data center mainly consists of two parts: the energy consumption of servers processing computing tasks and the energy consumption of auxiliary facilities maintaining the data center's operating environment. For ease of calculation, the power usage effectiveness (PUE) metric is typically introduced. PUE is defined as the ratio of total data center energy consumption to server energy consumption. Therefore, the total energy consumption of a cloud data center can be expressed as: (6) In the formula, This represents the total energy consumption of the cloud data center. This is the PUE value; This represents the total number of cloud servers. This refers to the static power of a single server. This refers to the dynamic power consumption coefficient of the cloud server. To ensure service quality, the computing performance of the cloud server is evaluated according to the M / M / 1 queuing model (see the detailed description in "A Review and Outlook on Energy Flow-Data Flow Collaborative Planning of Integrated Data Center's Comprehensive Energy System," which will not be repeated here). The average service time does not exceed the upper limit. : (7) In the formula, This indicates the service speed of the cloud server.

[0029] Cloud data centers obtain power from the local power grid. (Electricity purchased) Equal to its internal energy consumption: (8) To meet the real-time processing requirements of DSTs, task scheduling decisions at the edge must be made within more granular time slots. This embodiment addresses this in various time periods. t The time interval is further divided into 5-second sub-time periods, let's assume... The time index for the sub-period, sub-period t The reached DST attribute is , representing the number of arriving tasks and the maximum tolerable latency, respectively. The number of arriving tasks in the DST includes the local processing capacity. The amount of tasks offloaded to cloud data centers : (9) In the formula, For edge base stations i During the period t Sub-periods within t Latency-sensitive tasks processed locally For edge base stations i During the period t Sub-periods within t The amount of latency-sensitive tasks offloaded to cloud data centers. For edge base stations i During the period t Sub-periods within t The DST reached the task volume.

[0030] Compared to cloud data centers, edge base stations, limited by a smaller server configuration, are more likely to reach their processing limits during runtime. When service approaches saturation, computational efficiency typically decreases significantly. This embodiment uses test results from The Standard Performance Evaluation Corporation to demonstrate this point. Figure 1 As shown, server response time goes through two distinct phases as task throughput increases. In the first phase, response time grows in a near-linear fashion. When the load approaches saturation, response time increases sharply in a non-linear fashion. Furthermore, more abundant server resources (Test 3) make it easier to maintain computing performance under high task throughput. In contrast, Tests 1 and 2, with limited resources, quickly enter the non-linear region as task throughput increases.

[0031] Cloud data centers have abundant resources and can be considered to operate primarily in the linear region; however, edge base stations have limited resources, and their significant nonlinear response characteristics under high load need to be characterized in the model. Based on the above analysis, this embodiment uses a quadratic function approximation to characterize the relationship between local computing latency and processing workload at the edge: (10) in, For edge base stations i During the period t Sub-periods within t Local computation latency, For edge base stations i CPU cycle requirements per unit task Indicates edge base station i The calculation of the delay factor, For edge base stations i Number of edge servers For edge base stations i The service rate of the edge server.

[0032] Meanwhile, the amount of tasks that can be processed locally is limited by the maximum service capacity of the edge base station, as expressed as: (11) Furthermore, the transmission latency caused by task offloading is related to the amount of offloaded tasks, transmission distance, and transmission rate: (12) in, For edge base stations i During the period t Sub-periods within t The transmission delay caused by task unloading For edge base stations iTransmission distance, For edge base stations i The transmission rate.

[0033] Considering that DSTs offloaded from the edge to the cloud are typically lightweight tasks, and the cloud possesses powerful parallel processing capabilities, the computation time for processing DSTs in the cloud data center is negligible. Furthermore, the size of the computation result is much smaller than the input data, so we ignore the latency impact of data transmission back. Therefore, the total task processing latency includes the local computation latency at the edge and the offloading and transmission latency, and to ensure service quality, the total latency must not exceed the maximum time limit. (13) in, For edge base stations i During the period t Sub-periods within t Total delay, Edge base station i During the period t Sub-periods within t Maximum tolerable delay.

[0034] During each hour t Within the cloud data center, the offloading volume of all edge base stations must be consistent with the receiving volume: (14) In terms of energy consumption, edge base stations i During the sub-period t The total energy consumption within the unit is equal to the electricity purchased from the grid. Specifically, this includes server operating energy consumption. Energy consumption for task unloading and transmission : (15) (16) (17) In the formula, The computing energy consumption coefficient of the edge server. Indicates edge base station i Energy consumption of unloading unit computing tasks.

[0035] (2) Cloud-edge collaborative system computing-power joint scheduling strategy This embodiment first introduces the computing power-power coordination mechanism of the cloud-edge collaborative system, and then establishes computing power-power collaborative scheduling models for data centers and edge computing respectively, thereby constructing a complete joint scheduling strategy.

[0036] like Figure 2As shown, the cloud-edge collaborative system architecture is divided into two layers, which manage the cloud side and the edge side respectively.

[0037] At the upper layer, the cloud data center acts as the scheduling entity, in each scheduling period. t Within this process, the following three steps are completed: First, predict local electricity market information. Cloud data centers continuously process large volumes of computing tasks during the scheduling cycle, and their electricity consumption behavior is sufficient to affect local electricity settlement costs. Therefore, cloud data centers, as a significant factor influencing electricity costs, participate in scheduling. By predicting the local electricity cost characteristic curve, the coupling relationship between electricity demand and electricity cost is incorporated into the optimization model. Second, considering the time coupling constraints of the task queue, the number of DTTs processed during that period is planned. Finally, the data from all sub-periods is aggregated. N The DST task offloading volume of each edge base station. After completing the above information coordination, the cloud side will perform centralized optimization at the upper layer and send the hourly computing service unit cost signal to the edge side. .

[0038] At the lower level, each edge base station independently completes its own scheduling plan without interfering with each other. Since its power consumption is relatively small and insufficient to affect the grid settlement cost, each edge base station only receives time-of-use energy cost signals from the grid. Based on the unit cost of computing services received from the upper layer, each edge base station optimizes its own real-time task offloading volume. And upload it to the cloud data center.

[0039] In summary, the computing service cost decisions of cloud data centers and the task offloading plans of each edge base station are mutually influential and constrained, and are also affected by power grid cost signals, forming a typical two-layer coupled optimization problem.

[0040] The optimization goal of cloud data centers is to maximize total operating profit over the entire scheduling cycle. The corresponding model is as follows: (18) st (19) (20) (twenty one) in, For the entire scheduling cycle T Total operating profit within; This represents the fixed income obtained from processing DTT; This represents the revenue gained from providing computing power services to the edge. It is the cost of purchasing electricity paid to the local power grid; For time periodt The unit cost of cloud data center computing services For time period t The total number of latency-sensitive tasks received by the internal cloud data center. For time period t The unit cost of electricity settled by the local power grid. To purchase electricity.

[0041] according to Figure 3 The curve characteristics in the local power grid and the unit cost of electricity in the local power grid settlement This can be converted into a step function relating the power purchase demand of cloud data centers, expressed as: (twenty two) (twenty three) (twenty four) (25) In the formula, This is the index of the staircase segment in the diagram. S This represents the total number of steps. and Corresponding to the steps s The upper and lower limits of the covered electricity quota; For section s The highest marginal cost; For section s Quota variables on; To determine whether the intersection point falls within the segment s 0-1 variables.

[0042] The optimization goal of each edge base station is to maximize its overall utility. The corresponding model is as follows: (26) st (27) (28) (29) (30) In the formula, To process the fixed income obtained from local DST; This represents the cost of purchasing electricity paid to the power grid. It is the cost of computing services paid to cloud data centers; It is the cost weight of the delay penalty; This indicates the penalty cost for service delays; For time period tTime-of-use fixed electricity purchase cost, For edge base stations i During the period t Sub-periods within t Electricity purchased from the power grid, For time period t The unit cost of cloud data center computing services This represents the delay penalty coefficient.

[0043] (3) Model solution method To address the non-convexity of the underlying problem, this embodiment employs a second-order cone (SOC) relaxation technique. SOC is a convexification method that transforms a specific non-convex quadratic constraint into a standard second-order cone constraint. This relaxes equation (10) into an inequality constraint: (31) Note that this relaxation does not affect the accuracy of the model results. The theoretical basis is that the objective function (26) of each edge base station aims to maximize its own utility, since this function includes a negative delay penalty term. The optimization solver will seek [optimization] during the process of maximizing the objective function. The minimum value within the feasible region. Therefore, at the optimal solution, the above inequality constraints must be equal, and the problems before and after relaxation are equivalent.

[0044] After the SOC relaxation transformation, each lower-level problem is transformed into a standard second-order cone programming problem.

[0045] To protect the privacy of information from all parties involved, this embodiment employs a heterogeneous decomposition algorithm for distributed solution. This algorithm decomposes the original problem into two independent subproblems (a master problem and a secondary problem) by setting boundary variables, and then converges to the global optimum through ordered iteration of the boundary variables.

[0046] Specifically, two sets of boundary variables are introduced to replace the coupling variables in the original two-layer problem, thus decoupling the original problem: Coordination cost vector : This boundary vector corresponds to the computational service price for upper-level problem decision-making; the aggregate unloading matrix on the edge side. : Among them, the unloading vector This boundary vector corresponds to the task unloading amount in the lower-level problem decision.

[0047] Main problem: Given an aggregate unloading matrix Under these conditions, the main problem is to solve a profit maximization problem that covers the entire scheduling cycle and has cross-time period coupling constraints, thereby obtaining a new coordination cost vector. .

[0048] From the problem: Given a coordination cost vector Under these conditions, since the optimization problems of each edge base station are separable across hours, all edge base stations can solve the hourly optimization problem in parallel. Within each hour, each edge base station will progressively complete a second-order cone programming problem every 5 seconds, extracting the unloading variable optimization results from the solution set. Together, they constitute the task offloading volume of this edge base station. .

[0049] The iterative solution process of the algorithm is as follows: Figure 4 As shown, the process essentially involves alternating between solving the master and slave problems and continuously updating the boundary variables until the convergence condition is met. The specific steps are as follows: Step 1: Initialization. Set the initial coordination cost vector for the cloud data center. This vector is broadcast to all edge base stations. This vector can be set based on historical data. Set the iteration count. k =1.

[0050] Step 2: Parallel solution of the problem. In the... k In this iteration, all problems will As known parameters, the respective utility maximization problems are solved in parallel to obtain the optimal planned unloading volume for each hour. .

[0051] Step 3: Information Aggregation and Solving the Main Problem. All offloading plans for edge base stations are reported and aggregated into an offloading matrix. The main problem uses these parameters as known parameters to solve its own profit maximization problem. After solving, the dual variables related to constraint (14) together constitute a new coordination cost vector. .

[0052] Step 4: Convergence Test. Check the convergence of the boundary variables. If the changes in the cost vector and unloading vector are both less than the set threshold in two consecutive iterations... , : (32) If the algorithm converges, the solution to the master-slave problem satisfies the global optimality condition. Otherwise, let... k = k +1, return to step 2, and continue to the next iteration.

[0053] This distributed solution method can not only effectively solve complex two-layer coupled models, but also ensure the privacy of all stakeholders and has good scalability, making it able to adapt to the continuous expansion of the scale of future cloud-edge collaborative systems.

[0054] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics, wherein the cloud-edge collaborative system comprises a centralized cloud data center and multiple geographically dispersed edge base stations, characterized in that, The upper-layer cloud data center, acting as the scheduling entity, performs the following steps within each scheduling period: The cloud data center, as a significant factor influencing electricity costs, participates in the scheduling process. By predicting the local electricity cost characteristic curve, it incorporates the coupling relationship between electricity demand and electricity costs into the upper-layer optimization objective. Combining the time coupling constraints of the task queue, it plans the number of latency-tolerant tasks to be processed in the corresponding scheduling period. It aggregates the offloading volume of latency-sensitive tasks from edge base stations across all sub-periods. Combining the number of latency-tolerant tasks and the offloading volume of latency-sensitive tasks, with the optimization objective of maximizing the total operating profit throughout the entire scheduling cycle, and using the task volume constraints of latency-tolerant tasks, the first energy consumption constraint, and the consistency constraints of latency-sensitive tasks as constraints, it optimizes and obtains the computational service unit cost signal for this scheduling period, which is then distributed to each edge base station. Each edge base station in the lower layer receives time-of-use energy cost signals from the power grid and combines them with computing service unit cost signals received from the upper layer. With maximizing its own comprehensive utility as the optimization goal, and with latency constraints, secondary energy consumption constraints, and latency-sensitive task consistency constraints as constraints, it independently optimizes its own real-time task offloading volume and uploads it to the cloud data center.

2. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics as described in claim 1, characterized in that, The optimization objective of the cloud data center is expressed as: ， in, For the entire scheduling cycle T Total operating profit within, This is the hourly time index within the corresponding scheduling period. This represents the fixed reward obtained from handling delay-tolerant tasks. This represents the revenue gained from providing computing power services to the edge. It is the cost of purchasing electricity paid to the local power grid; ， ， in, For time period t The total number of latency-sensitive tasks received by the internal cloud data center. For time period t The unit cost of cloud data center computing services For time period t The unit cost of electricity settled by the local power grid. To purchase electricity, ， ， ， ， in, It is an index of the step function relating the settlement of electricity costs to the electricity purchase demand of cloud data centers. S This represents the total number of steps. and Corresponding to the steps s The upper and lower limits of the covered electricity quota; For section s The highest marginal cost; For section s Quota variables on; To determine whether the intersection point falls within the segment s 0-1 variables.

3. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics as described in claim 1, characterized in that, The specific workload constraints for the delay-tolerant task are as follows: All pending, latency-tolerant tasks form a task queue. The cumulative number of tasks processed in the queue is within a dynamic feasible region, represented as follows: ， in, h From the initial time period to the time period t Time index; For time period h The lower bound of the workload for latency-tolerant tasks processed by the internal cloud data center. Indicates the end time t The cumulative amount of tasks that must be completed, upper limit The total number of tasks completed is represented as follows: ， ， ， in, It is a binary indicator variable used to determine the time period. h Will the arriving task be during the time period? t Deadline reached; Time period t Arrival delay-tolerant task attributes include , , These represent the amount of task arrival and the maximum tolerable delay, respectively.

4. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics as described in claim 1, characterized in that, The first energy consumption constraint is specifically: Cloud data centers obtain electricity from the local power grid and purchase electricity. Equal to its total internal energy consumption: ， in, This represents the total energy consumption of the cloud data center. For electrical energy utilization efficiency, This represents the total number of cloud servers. This refers to the static power of a single server. This refers to the dynamic power consumption coefficient of the cloud server. For cloud data centers during time periods t Total task processing volume, , For time period t The workload of delay-tolerant tasks. For time period t The total number of latency-sensitive tasks received by the cloud data center; To ensure service quality, the computing performance of cloud servers is evaluated based on the M / M / 1 queuing model, and the average service time does not exceed the upper limit. : ， in, This indicates the service speed of the cloud server.

5. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics according to claim 1, characterized in that, The specific consistency constraint for the latency-sensitive task is as follows: At different times t Within, the time is further divided into sub-time periods based on a preset time interval. For the time index of the sub-period, For time period t The number of sub-time periods is divided, and the offloading volume of all edge base stations is consistent with the receiving volume of the cloud data center: ， in, Indicates edge base station i During the period t Sub-periods within τ The amount of latency-sensitive tasks offloaded to cloud data centers. N For the number of edge base stations, For time period t The total number of latency-sensitive tasks received by the internal cloud data center.

6. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics according to claim 1, characterized in that, The optimization objective of the edge base station is expressed as: ， ， ， ， in, For edge base stations i Its own comprehensive utility For edge base stations i During the period t Fixed revenue earned from handling local latency-sensitive tasks; Represents edge base station i During the period t The cost of purchasing electricity paid to the power grid; It is an edge base station i During the period t The cost of computing services paid to cloud data centers; It is an edge base station i The cost weight of delay penalties; Indicates edge base station i During the period t Service delay penalty costs; This is the hourly time index within the corresponding scheduling period. T Indicates the scheduling period; For the time index of the sub-period, For time period t The number of sub-time periods; For time period t Time-of-use fixed electricity purchase cost, For edge base stations i During the period t Sub-periods within τ Electricity purchased from the power grid, For time period t The unit cost of cloud data center computing services For edge base stations i During the period t Sub-periods within τ The amount of latency-sensitive tasks offloaded to cloud data centers. This is the delay penalty coefficient. For edge base stations i During the period t Sub-periods within τ Total delay.

7. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics according to claim 1, characterized in that, The delay constraint is expressed as follows: ， in, For edge base stations i During the period t Sub-periods within τ Total delay, For edge base stations i During the period t Sub-periods within τ Local computation latency, For edge base stations i During the period t Sub-periods within τ The transmission delay caused by task unloading Edge base station i During the period t Sub-periods within τ Maximum tolerable delay; ， ， ， in, For edge base stations i CPU cycle requirements per unit task For edge base stations i During the period t Sub-periods within τ Latency-sensitive tasks processed locally Indicates edge base station i The calculation of the delay factor, For edge base stations i Number of edge servers For edge base stations i The service speed of the edge server For edge base stations i During the period t Sub-periods within τ The amount of latency-sensitive tasks offloaded to cloud data centers. For edge base stations i Transmission distance, For edge base stations i The transmission rate.

8. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics according to claim 1, characterized in that, The second energy consumption constraint is as follows: Edge base station i During the sub-period τ The total energy consumption within the unit is equal to the electricity purchased from the grid. : ， ， ， in, For edge base stations i During the period t Sub-periods within τ Electricity purchased from the power grid, For edge base stations i During the period t Sub-periods within τ Server operating energy consumption, For edge base stations i During the period t Sub-periods within τ Task offloading and transmission energy consumption For edge base stations i The computing energy consumption coefficient of edge servers For edge base stations i Number of edge servers For edge base stations i The service speed of the edge server For edge base stations i CPU cycle requirements per unit task For edge base stations i During the period t Sub-periods within τ Latency-sensitive tasks processed locally Indicates edge base station i Energy consumption of unloading unit computing tasks. For edge base stations i During the period t Sub-periods within τ The amount of latency-sensitive tasks offloaded to the cloud data center.

9. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics as described in claim 7, characterized in that, The edge base station employs a second-order cone relaxation technique during the lower-layer optimization process. In the delay constraint Relaxation is a constraint based on an inequality: Each lower-level optimization problem is transformed into a standard second-order cone programming problem.

10. The cloud-edge collaborative system computing power-power joint optimization scheduling method considering heterogeneous characteristics according to claim 1, characterized in that, The proposed computing power-power joint optimization scheduling method employs a heterogeneous decomposition algorithm for distributed solution across upper and lower layers. This algorithm decomposes the original problem into independent master and slave problems by setting boundary variables. These boundary variables include a coordination cost vector and an edge-side aggregation unloading matrix. The coordination cost vector corresponds to the computing service price decided in the upper-layer problem, while the edge-side aggregation unloading matrix corresponds to the task unloading amount decided in the lower-layer problem. Thus, the master problem is: given the aggregation unloading matrix, solve a profit maximization problem covering the entire scheduling cycle and with cross-time-period coupling constraints to obtain a new coordination cost vector. The slave problem is: given the coordination cost vector, all edge base stations solve their respective hourly optimization problems in parallel. Within each hour, each edge base station progressively completes the second-order cone programming problem for each sub-time period, extracting the optimization results of the unloading variables from the solution set to collectively constitute the task unloading amount for that edge base station.

Images