An electric carbon perception-based computing power network task scheduling method and system

Abstract

The application discloses a kind of electric carbon perception-based computing power network task scheduling method, system and device, the method includes: introducing carbon tax mechanism, constructs electric carbon perception-based computing power network task scheduling model;Based on the computing power network task scheduling model, the problem of minimizing system electric carbon cost is generated;Based on Lyapunov optimization, the problem is transformed, and the optimal solution of scheduling scheme is obtained by using DPRA algorithm solution.The system includes: model construction module, problem definition module and solving module.Through using the application, green and stable computing power network task scheduling can be realized.The application can be widely applied to computer computing power network field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer computing power network, and particularly relates to a computing power network task scheduling method and system based on electric carbon perception. BACKGROUND

[0002] With the continuous emergence of new applications and new businesses, including virtual reality, Internet of Things, Internet of Vehicles, etc., while providing customers with better services, these applications also put forward higher requirements for network bandwidth resources, computing resources and other resources. Under this background, as a supplement to traditional cloud technology, edge computing is proposed to provide users with more targeted and demand-oriented network and computing services.

[0003] After several years of rapid development, edge computing has entered a stable development stage. The large-scale deployment of edge computing and intelligent terminal devices makes it more convenient and fast for user terminals to access and use these massive distributed computing resources. However, with such a large number of edge computing nodes, the computing resources of each individual node are limited and difficult to complete some computing-intensive tasks or provide high-speed services. Moreover, the coordination mechanism between edge computing nodes, the cloud center and edge nodes, and the allocation and scheduling mechanism of computing tasks are not yet perfect, resulting in low utilization of computing resources. Under this background, compute first networking (CFN) has emerged.

[0004] Compute first networking is a new solution to the unified supply of different levels of computing resources, including cloud computing, edge computing, and terminal computing. The purpose is to achieve "cloud-edge-end" three-end coordination through reasonable task allocation and scheduling mechanism, combine network information and user demand, and provide and allocate computing resources, network resources and other resources, so as to ultimately realize the optimal configuration and use of the entire computing power network, maximize the utilization of computing resources and network resources, etc. At the same time, compute first networking has high flexibility and scalability, and any data center can become a node in the compute first networking.

[0005] However, with the rapid development of digital economy, the demand for computing power, as the core productivity supporting the development of digital economy, is growing explosively. The energy consumption problem of compute first networking, as the main carrier of computing power, is increasingly prominent. The consumption of power resources and carbon emissions caused by this problem have attracted increasing attention. Therefore, exploring the development path of the integration of electricity, carbon sink and computing power, and realizing green and low-carbon development of computing power, has far-reaching significance for the sustainable development of digital economy and the realization of the "double carbon" goal.

[0006] Currently, the task scheduling method of traditional compute first networking often prioritizes resource availability and computing efficiency, while ignoring energy and environmental factors. SUMMARY

[0007] Therefore, in order to solve the technical problem that the existing computing power network task scheduling method does not consider energy and environmental protection factors, thereby causing the energy consumption of the computing power network to surge, the present application proposes a computing power network task scheduling method based on electricity-carbon perception, which comprises the following steps:

[0008] Introducing a carbon tax mechanism to build an electricity-carbon-aware computing power network task scheduling model;

[0009] Based on the computing power network task scheduling model, a problem of minimizing the system electricity-carbon cost is generated;

[0010] Based on Lyapunov optimization, the problem is transformed and solved using a dependent probability rounding algorithm (DPRA) algorithm to obtain the optimal solution of the scheduling scheme.

[0011] In some embodiments, the electricity-carbon-aware computing power network task scheduling model specifically includes:

[0012] The wireless terminal device communicates with the regional access point and uses the computing service;

[0013] The regional access point schedules the received tasks, communicates with the edge server and the cloud server, considers the overall electricity-carbon cost of the computing power network, and selects to schedule the tasks to the edge server or the cloud server for processing.

[0014] Among them, the computing power network architecture: covers the computing power network topology structure of three-level computing power resources of cloud-edge-terminal (cloud computing, edge computing, terminal computing);

[0015] Electricity price and carbon tax mechanism: considering the overall electricity-carbon cost of the computing power network, the electricity cost comes from the electricity price proportional to the electricity consumed for paying and processing tasks, and the carbon cost comes from the carbon tax mechanism, paying carbon tax proportional to carbon emissions.

[0016] In some embodiments, the step of generating a problem of minimizing the system electricity-carbon cost based on the computing power network task scheduling model specifically includes:

[0017] Defining constraints and objective functions;

[0018] The constraints include task non-disassembly constraints, task complete processing constraints, quality of service constraints, resource constraints and stability constraints;

[0019] The objective function is to minimize the long-term average electricity-carbon cost of the overall computing power network system.

[0020] wherein, the task non-disassembly constraint: the task scheduling decision variable is a 0 / 1 integer variable;

[0021] the task complete processing constraint: the sum of the decision variables corresponding to the task scheduling to each server is 1;

[0022] the service quality constraint: the corresponding time of the task cannot exceed the predetermined maximum value;

[0023] the resource constraint: the size of the task waiting for processing cannot exceed the storage size of the corresponding server;

[0024] the stability constraint: the system queue is stable, that is, each task can be processed in a limited time.

[0025] In some embodiments, the expression of the problem of minimizing the system electric carbon cost is as follows:

[0026]

[0027]

[0028] wherein, X represents the set of system task scheduling decision variables, the first constraint condition represents that each decision variable x ij (τ) is a 0 / 1 variable, the second constraint condition ensures that any task in each time slot must be processed and can only be scheduled to a single server for processing, the third constraint condition is the resource constraint, which ensures that the task storage space size of each server will not exceed the storage space size of the server, the fourth constraint condition is the service quality constraint, which ensures that the task processing delay of each server will not exceed the upper bound d max , and the fifth constraint condition is the system stability constraint, which ensures the overall stability of the system and that all tasks can be processed eventually.

[0029] In some embodiments, the problem is transformed based on Lyapunov optimization and solved by using the DPRA algorithm to obtain the optimal solution of the scheduling scheme, and the step specifically comprises:

[0030] transforming the problem based on Lyapunov optimization and omitting the terms in the objective function that are irrelevant to the decision variables, to obtain a problem that is solved independently for each time slot;

[0031] relaxing the constraints in the problem that is solved independently for each time slot to linear constraints to obtain a relaxation problem;

[0032] solving the optimal solution of the relaxation problem based on the linear programming algorithm, rounding the variables, and marking the tasks that are destroyed after scheduling;

[0033] scheduling the marked tasks based on the greedy algorithm to obtain the optimal solution of the scheduling scheme.

[0034] wherein, the relaxation problem of the transformation problem is solved: the relaxation problem after the simplex method is applied to solve the transformation problem, that is, the corresponding problem of the transformation problem with the task non-disassembly constraint relaxed;

[0035] integerization: the optimal solution of the relaxation problem is set as the probability of each decision variable being set to 1, while ensuring that only one variable is set to 1 and the rest are set to 0 in all decision variables corresponding to each task, if the constraint is broken after the task is scheduled, the task is marked and the scheduling of the task is cancelled;

[0036] greedy algorithm: all marked tasks are scheduled by a greedy algorithm.

[0037] In some embodiments, the expression of the problem solved independently for each time slot is as follows:

[0038]

[0039]

[0040] wherein, X(τ) represents the task scheduling decision set corresponding to time slot τ, W j represents the computing speed of server j, V represents a control parameter introduced by Lyapunov optimization, represents the unit electricity price corresponding to server j, I j represents the carbon emission intensity of server j, represents the carbon tax corresponding to server j, and a represents the proportion factor of power resource consumption / server computing amount.

[0041] In some embodiments, it further comprises:

[0042] Through simulation and comparative experiments, the performance of the electricity-carbon-aware computing power network task scheduling model is verified.

[0043] wherein, the algorithm performance comparison: the performance of different strategy algorithms is compared, including the all-scheduling-to-cloud algorithm (ALL2CLOUD), the greedy algorithm (GREEDY), the algorithm (DPRA) and the integer programming optimal algorithm (OPT), the results prove that the all-scheduling-to-cloud algorithm has the highest electricity-carbon cost, the greedy algorithm and the DPRA algorithm have similar performance to the optimal algorithm, and the DPRA algorithm is better than the greedy algorithm, and the optimization interval of the DPRA algorithm is less than half of the optimization interval of the greedy algorithm;

[0044] Model and algorithm effectiveness verification: different control parameters V are set to draw the corresponding simulation results, and the simulation results conform to the theoretical derivation of the application;

[0045] Influence of part of parameters on model effect: different queue upper limits d maxSimulation experiments are carried out, corresponding simulation results are drawn, and corresponding analysis is given.

[0046] The application further provides an electric-carbon-aware computing power network task scheduling system, which comprises:

[0047] A model construction module is configured to introduce a carbon tax mechanism and construct an electric-carbon-aware computing power network task scheduling model.

[0048] A problem definition module is configured to generate a problem of minimizing system electric-carbon cost based on the computing power network task scheduling model.

[0049] A solution module is configured to transform the problem based on Lyapunov optimization and solve the problem by using a DPRA algorithm to obtain an optimal solution of the scheduling scheme.

[0050] The application further provides an electric-carbon-aware computing power network task scheduling device, which comprises:

[0051] At least one processor;

[0052] At least one memory configured to store at least one program;

[0053] When the at least one program is executed by the at least one processor, the at least one processor is caused to implement the electric-carbon-aware computing power network task scheduling method.

[0054] Based on the above scheme, the application provides an electric-carbon-aware computing power network task scheduling method and system, introduces a carbon tax mechanism in the carbon price signal into the computing power network, comprehensively considers the stability, service quality and resource consumption of the computing power network task scheduling system, constructs an electric-carbon-cost-aware computing power network task scheduling model, deduces an optimization problem of balancing the electric-carbon cost and the system stability and service quality, and proposes a DPRA algorithm to solve an approximate optimal solution. The method can effectively balance the electric-carbon cost and the system stability and service quality of the computing power network system according to actual demand by setting control parameters, and efficiently processes task scheduling. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a step flowchart of the electric-carbon-aware computing power network task scheduling method of the application;

[0056] Figure 2 is a scene schematic diagram of the electric-carbon-aware computing power network task scheduling in the application;

[0057] Figure 3 is a queue service model of a server;

[0058] Figure 4Table of simulation experiment parameter setting;

[0059] Figure 5 Performance diagram of each algorithm under large V value (including all scheduling to cloud strategy);

[0060] Figure 6 Optimization interval comparison diagram of better algorithm under large V value;

[0061] Figure 7 Long-term average electricity carbon cost comparison diagram of better algorithm under large V value;

[0062] Figure 8 Influence diagram of control parameter V on model performance;

[0063] Figure 9 Influence diagram of queue length upper limit on queue length of each server (real-time);

[0064] Figure 10 Influence diagram of queue length upper limit on queue length of each server (average);

[0065] Figure 11 Influence diagram of queue length upper limit on cloud edge distribution ratio of task scheduling. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0067] It should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings. The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0068] It should be understood that the "system", "device", "unit" and / or "module" used in the present application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

[0069] Unless the context explicitly indicates an exception, words such as "a," "an," "a kind," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list; a method or apparatus may also include other steps or elements. An element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0070] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0071] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0072] Reference Figure 1 This is a flowchart illustrating an optional example of the computing power network task scheduling method based on carbon dioxide sensing proposed in this invention. This method can be applied to computer devices, and the imaging method proposed in this embodiment may include, but is not limited to, the following steps:

[0073] Step S1: Introduce a carbon tax mechanism and construct a computing power network task scheduling model for electricity carbon sensing;

[0074] Step S2: Based on the computing power network task scheduling model, generate a problem that minimizes the system's carbon dioxide cost;

[0075] Step S3: Transform the problem based on Lyapunov optimization and solve it using the DPRA algorithm to obtain the optimal solution of the scheduling scheme.

[0076] In some feasible embodiments, step S1 specifically includes:

[0077] The scenario of the task scheduling model of the computing power network is as follows: Figure 2 As shown, in this model, several wireless devices send their required tasks to a regional access point, which then schedules the tasks based on the task information and the state of the computing network system. The server adopts a queue service model (SIQ, Serve in Queue), such as... Figure 3As shown, the server receives a task package containing a plurality of tasks of the same batch assigned by the access point between each time slot, each task package is saved as a queue element in the memory waiting for the server to process, and the server encapsulates the task response and current state information of the processed task package and sends it to the receiving point for task scheduling.

[0078] Since the edge server can have more energy supply, including green energy and power grid, it is assumed that the carbon intensity of the edge server can be lower than that of the cloud server, and due to the instability and volatility of green energy, the carbon intensity of the edge server also fluctuates. Assuming that the task calculation consumes power resources proportional to the task size, therefore, the power resources consumed by task calculation are the same whether in cloud server processing or in edge server processing, but the power resource consumption generated by the task transmission to the cloud server is larger than that to the edge server, so the task scheduling to the cloud server processing will consume more power resources than the scheduling to the edge server. In order to achieve the "double carbon" goal and promote carbon emission reduction, the present application introduces the carbon tax mechanism in the carbon price signal, and the user needs to pay tax on the carbon emissions according to a certain proportion, although the average price of green power (such as photovoltaic power, wind power, etc.) is higher than that of traditional thermal power, but the carbon emission reduction target will automatically optimize the carbon tax mechanism, so as to achieve the condition that the average electric carbon price of using green power is lower than that of traditional thermal power, so as to promote the replacement of green and clean energy and carbon emission reduction. Therefore, it is concluded that scheduling tasks to the cloud server processing will cause higher electric carbon cost, and scheduling to the edge server with more diverse power resources can reduce the electric carbon cost. However, due to the limited task processing capacity of the edge server, when the number of tasks is large, it is not possible to schedule all tasks to the edge server for the purpose of saving electric carbon cost, and scheduling a large number of tasks to the edge server will cause the growth of the task backlog queue, which may on the one hand cause the instability of the overall system of the computing power network, and on the other hand, increase the task processing delay, reduce the service quality, and affect the user experience. In general, scheduling tasks to the edge server can effectively reduce the electric carbon cost caused by processing tasks, but excessive scheduling will reduce the stability and service quality of the system.

[0079] It is assumed that the task scheduler runs the computing power network system in the discrete time range {0, 1, …, t}, where each element represents a time slot, and t represents the length of the system running time. At the beginning of each time slot, a plurality of demand tasks arrive at the access point, and the access point schedules the arriving tasks according to the task information and the current system state. It is assumed that each demand task (where is the task set arriving at the access point at time slot τ, which satisfies ) has the attribute F i(T), to express the task computation amount. Let M = {0, 1, 2, …, m-1} be the total server set, where 0 represents a cloud server and the others represent edge servers, and m represents the total number of servers in the computing power network system. For each server j e M, Cap j represents the storage space size of server j, and the amount of task data received by the server cannot exceed the storage space size of the server. After the server completes processing of a single task package, the storage space occupied by the task package is released; W j represents the computing speed of server j (computing time = task computation amount / computing speed). It is assumed that the distribution of the computing power network system is reasonable (i.e., the computing power resources of the computing power network can always meet the task demand). Therefore, it is assumed in this embodiment that the computing speed of the cloud server W0→∞, and the task storage space of the cloud server can always meet the demand. Since the electricity price and the carbon emission intensity fluctuate slightly, in order to make the problem more intuitive and simple, it is assumed in this embodiment that the electricity price and the carbon emission intensity are constants, represents the unit electricity price corresponding to server j, I j represents the carbon emission intensity of server j, P C represents the carbon tax corresponding to the region where the computing power network is located. Therefore, the electricity-carbon cost per unit computation amount corresponding to server j is As shown in the previous paragraph, as an economic means of carbon emission reduction, the carbon tax can automatically optimize the electricity-carbon cost, so that the unit computation amount electricity-carbon cost corresponding to the cloud server with higher electricity consumption and carbon emission is higher than that corresponding to the edge server.

[0080] In some possible embodiments, the step S2 specifically comprises:

[0081] For each time slot τ = 0, 1, 2, …, t-1, the task scheduler obtains the task set N(τ) of the time slot τ and schedules the tasks according to the current system state, electricity-carbon cost information, and task information, so that represents the task scheduling decision set corresponding to the time slot τ, where x ij is a 0 / 1 variable, and x ij =1 indicates that task i is transmitted to server j for processing. Let A j (τ) represent the electricity consumption generated by server j in processing task N(τ). It is assumed in this embodiment that the computing electricity consumption is proportional to the processing task computation amount, A pj (τ) = a∑ i∈N(τ) x ij F i(τ), where a is the proportional factor between the electric energy consumption and the processing task computation amount. Since the cloud server transmission loss is large, the cloud server transmission loss electric energy proportional factor is denoted by b (b > 0), the edge server transmission loss is negligible compared with the cloud server loss, so the edge server loss electric energy proportional factor is assumed to be 0, then the cloud server electric energy consumption A0(τ) = a(1 + b)∑ i∈N(τ) x i0 F i (τ), the electric energy consumption of other edge servers A j (τ) = a∑ i∈N(τ) x ij F i (τ), j≠0. The carbon emission of server j Car j (τ) = A j (τ) * I j , so the electric carbon cost C j Further, the overall electric carbon cost of the system is

[0082]

[0083] Let Add j (τ) represent the system task time increase set corresponding to time slot τ, where Add (τ) represents the system task time increase of server j corresponding to time slot τ, then

[0084]

[0085] Let G (τ) represent the system task time backlog queue set corresponding to time slot τ, where G j (τ) represents the system task time backlog queue of server j corresponding to time slot τ. Then It is easy to know that the system task time backlog queue G(τ) changes in real time with the advancement of time slot τ.

[0086] Let o j (τ) represent the used storage space of server j in time slot τ. As shown in the system model, each server only releases the storage space occupied by the task package after the processing of the task package is completed. The processing of a task package may require multiple time slots, therefore, the change of the remaining storage space of the server does not change in real time with the advancement of time slot τ, o j (τ + 1) is not only related to o j (τ), but also related to the remaining processing time of the current task of the server.

[0087] The optimization objective of the present application is to minimize the long-term average electricity and carbon cost of the system as a whole under the premise of satisfying the overall stability of the system, the resource constraints of the edge server and the quality of service constraints, thus obtaining the mathematical expression of the optimization problem The mathematical expression of the optimization problem

[0088]

[0089]

[0090] The first constraint condition indicates that each decision variable x ij (τ) is a 0 / 1 variable, the second constraint condition ensures that any task in each time slot must be processed and can only be scheduled to a single server for processing, the third constraint condition is the resource constraint, which ensures that the size of the task storage space of each server will not exceed the storage space size of the server, the fourth constraint condition is the quality of service constraint, which ensures that the task processing delay of each server will not exceed the upper limit d max , and the fifth constraint condition is the system stability constraint, which ensures the overall stability of the system and that all tasks can be finally processed.

[0091] In some feasible embodiments, the step S3 specifically comprises:

[0092] The original problem is transformed by applying the Lyapunov optimization technique, and the terms irrelevant to the decision variables in the objective function are omitted, so that the original problem requiring future information is transformed into a problem that can be solved independently for each time slot The expression is as follows:

[0093]

[0094]

[0095] In addition, it is noted that the objective function and the constraint conditions of the problem except the first one are linear, so the first constraint of the problem is relaxed to a linear constraint, and the relaxed problem of the problem is obtained. The expression is as follows:

[0096]

[0097]

[0098] First, the optimal solution of the problem is obtained by using a commonly used linear programming algorithm independent rounding of the variables in the problem, in this step, the task schedules that will cause the third and fourth constraints to be violated are marked, and in the last step, the marked tasks are scheduled based on a greedy algorithm to obtain an approximate optimal solution of the problem

[0099] In some possible embodiments, further comprising:

[0100] The optimization gap and performance of the Lyapunov optimization technique and the DPRA algorithm are analyzed.

[0101] Specifically, the optimization gap between the problem and satisfies:

[0102]

[0103] where the left side of the inequality represents the long-term average electric-carbon cost corresponding to the optimal solution of the problem minus the optimal objective value of the original problem , that is, the optimization gap, and in the right side of the inequality, B represents a constant upper bound, is the maximum number of access point arrival tasks in a single time slot, f max is the maximum computation amount of a single task, W min is the computation speed of the slowest edge server, and V is a control parameter introduced by the Lyapunov optimization technique to balance the proportion of electric-carbon cost and system stability and service quality in the optimization objective. It is known that the larger the control parameter V is set, the smaller the optimization gap brought by the Lyapunov optimization technique is.

[0104] The upper bound of the queue length of the system as a whole satisfies

[0105]

[0106] where the left side of the inequality represents the sum of the long-term average queue lengths of all servers in the system, and its upper bound is determined by the right side of the inequality, where η≥0 is a constant related to the task amount, and the larger the task amount processed by the system is, the smaller η is, and C max is the worst objective function value of the problem , and it is easy to know that it corresponds to the case where all tasks are scheduled to the cloud server.

[0107] The time complexity of a single time slot of the DPRA algorithm includes the time complexity of a simplex method and the rounding time complexity. The complexity of the simplex method can reach an exponential level, but it is generally considered that the complexity in actual application is acceptable. The complexity of a single rounding is ​Therefore, the total complexity of the DPRA algorithm is where t is the length of the system running time.

[0108] The optimized interval of the DPRA algorithm, i.e., the obtained approximate optimal solution The corresponding problem The target function value of the problem The optimal target function value between the interval, satisfies

[0109]

[0110] where represents The target function value of the corresponding problem R opt represents the optimal target function value of the problem , φ is used to measure the optimal solution of the problem , Next, the ratio between the total amount of task calculation and the amount of task calculation allocated to the busy server represents the scheduling distribution ratio of the task between the idle and busy servers,

[0111] Based on the above scheme, simulation examples and comparative experiments are also given.

[0112] According to the established system model, parameters such as Figure 4 are set.

[0113] It is assumed that the computing power network has |M| = 10 servers participating in task processing, including one cloud server and nine edge servers, and the number of tasks received by the unit time slot access point is subject to uniform distribution, For the setting of server attributes, it is assumed that the computing speed and queue storage capacity of the cloud server tend to infinity, so that even if the number of tasks and the amount of calculation are large, the entire computing power network can be in a stable state to process these tasks, ensuring the feasibility of the original problem. The computing speed and queue storage capacity of the edge server are subject to uniform distribution of [1, 2]*10 8 bit / time slot and [2, 6]-10 8 bit. The task data volume F i (τ) is subject to uniform distribution of [4, 10]*10 8 bit. It is assumed that the energy consumption / calculation amount ratio a is 1.4*10 -11 kWh / bit, the cloud transmission loss ratio b is 1.1. The electricity price is set with reference to the average price of the living electricity price, and the carbon emission intensity is referenced to the average carbon emission factor of the national power grid in 2022. It is assumed that the electricity price of the cloud server is is 0.7 Yuan / kWh, and the carbon intensity I0 of the cloud server is 2*10 -3 tCO2 / kWh, and the electricity price of the edge server obeys a uniform distribution of [0.75, 0.85] Yuan / kWh, and the carbon intensity I j of the edge server obeys a uniform distribution of [1, 3]*10 -4 tCO2 / kWh. The carbon tax floor is set to $50, and the carbon tax P C is approximately set to 350 Yuan / tCO2 according to the floor and the exchange rate. The upper limit d max of the queue length is set to 150, and the control parameter V is set in [1, 10]*{10 4 , 10 8} to verify the effectiveness of the model and the algorithm. Finally, the system running time length t is set to 960 time slots, each of which represents 1.5 minutes, that is, the system running time length is 24 hours.

[0114] At the beginning of each time slot, an optimization problem is established according to the environmental information such as the electricity price and the carbon tax and the system information such as the queue length, then the task scheduling strategy is obtained by using the DPRA algorithm, and finally, the queue length and the server storage space size of the system are updated according to the task scheduling, and the task scheduling of the next time slot is entered.

[0115] The control parameter V affects the weight of the electricity-carbon cost and the length of the accumulated queue in the objective function, and the greater V is, the greater the influence of the electricity-carbon cost and the smaller the influence of the length of the accumulated queue. At the same time, it is noted that the electricity-carbon cost and the length of the accumulated queue have different characteristics, and the electricity-carbon cost in different time slots is independent of each other, while the length of the accumulated queue is dependent on each other. Therefore, the greater the value of V, the more the value of the objective function of the problem represents the performance of the algorithm for solving the single-time-slot optimization problem.

[0116] In order to verify the effectiveness of the DPRA algorithm of the present application and compare the performance of different algorithms, in this embodiment, V is set in [1, 10]*10 8 , which is a large numerical range, and the simulation results are shown in Figure 5 , Figure 6 , Figure 7 .

[0117] As can be seen from Figure 5 , the objective function values of the DPRA algorithm, the greedy algorithm and the optimal solution are much better than the objective function value of the algorithm of scheduling all tasks to the cloud, and with the increase of V, all the objective function values are approximately linearly increased, which conforms to the assumption of the problem model and the problem model is effective. In order to facilitate the comparison of the other three algorithms, the algorithm of scheduling all tasks to the cloud is ignored hereinafter.

[0118] To see the performance gap between DPRA and greedy algorithm more clearly, the vertical coordinate is changed to the ratio of the objective function value to the optimal objective value, as shown in Figure 6 Figure 6 It can be seen that the optimization gap of DPRA is 0.021%, and the optimization gap of greedy algorithm is 0.045%, the optimization gap of DPRA is less than half of that of greedy algorithm, which shows that the performance of DPRA is better than that of greedy algorithm.

[0119] The performance of different algorithms in the optimization of electricity and carbon cost is plotted in Figure 7 Figure 7 It can be seen that the electricity and carbon cost of DPRA is also lower than that of greedy algorithm. In addition, Figure 6 The vertical coordinate represents the ratio of the objective function value of the solution obtained by the algorithm to the optimal objective value of the problem , and the horizontal coordinate represents the value of V. Figure 7 The vertical coordinate represents the objective function value of the solution obtained by the same algorithm for the problem , and the horizontal coordinate represents the value of V. The trend of the two curves is consistent, which shows that in the case of large V value, the optimization gap of the problem and is very small, which proves the correctness of the theoretical analysis in this embodiment.

[0120] (1) Set V to an appropriate value to balance the electricity and carbon cost and the queue length.

[0121] In practical applications, it is necessary to adjust the size of the control parameter V according to the actual situation in a more suitable V value range to achieve the optimal performance. In this embodiment, V is set in the range of [1, 10]*10 4 This value range makes the trade-off between electricity and carbon cost and task time accumulated queue length more obvious, and the simulation results are shown in Figure 8

[0122] It can be seen from Figure 8 (a) that as the value of V increases, the objective function value also increases, which is obvious because the value of V is the coefficient of the electricity and carbon cost term in the objective function. Combined with Figure 8 (b) and Figure 8 , it can be seen that as the value of V increases, the queue length also increases, which is also obvious because the value of V is the coefficient of the queue length term in the objective function.​​(c) As can be seen, with the increase of V value, the system's carbon dioxide cost also increases, and the rate of increase gradually decreases, while the task-time accumulation queue length increases, and the rate of increase also gradually decreases, tending to stabilize. In practical applications, the V value can be adjusted according to requirements to balance the carbon dioxide cost with the task-time queue length, as well as the carbon dioxide cost with system stability and service quality, in order to achieve overall system optimization. Furthermore, as can be seen from the above three figures, the performance of different algorithms fluctuates. This is because a smaller V value means the task-time queue length has a certain influence on the objective function, leading to increased interdependence between optimization problems corresponding to different time slots, resulting in performance fluctuations. This does not contradict the conclusion in the previous section and further verifies the relationship between the optimization spacing and V value brought about by the Lyapunov optimization technique.

[0123] Depend on Figure 8 (d) As can be seen, as the value of V increases, the number of tasks scheduled to the cloud server decreases while the number of tasks scheduled to the edge server increases. This is because as the value of V increases, the impact of carbon electricity costs becomes greater, and the system tends to choose edge servers with richer green electricity resources and lower carbon electricity costs to process tasks in order to reduce carbon electricity costs. This result further verifies the effectiveness of the problem model of the present invention.

[0124] (2) Investigating the upper limit d of the queue length max Impact on the system.

[0125] Set V = 4 * 10 4 Using the DPRA algorithm of this invention, the upper limit d of the queue length is studied. max Impact on system performance. (d) max The values ​​are set to {50, 60, 70, 80}, and the simulation results are as follows. Figure 9 , Figure 10 and Figure 11 As shown.

[0126] exist Figure 9 In the diagram, each curve represents the change in queue length over time for a single edge server's task. First, when the computing network system starts from state zero, the queue lengths of all servers initially increase approximately proportionally until they reach a steady state when time slot τ is between 50 and 100, fluctuating around this steady-state value. Second, the steady-state values ​​differ between different edge servers. For edge servers with lower queue lengths and more distinct separation from other servers, their steady-state values ​​vary depending on the specific edge server. The objective function dictates that when the queue length is at a steady state, the increase in the objective function value caused by scheduling tasks to this edge server and scheduling them to the cloud server is the same; however, for edge servers with a large overlap at the top (e.g., ...), ...Figure 9 (a) the upper coincident curve), whose steady-state value is determined by the quality-of-service constraint, i.e. max d, the growth of the objective function value caused by scheduling tasks to the server is always less than that caused by scheduling tasks to the cloud server.

[0127] Figure 10 The average queue length distribution of each edge server is shown, and it can be found that, as d max increases, the overall queue length of the system also presents a slowing increasing trend, and the overall distribution becomes more uneven, and the steady-state value of some edge servers limited by d max increases, as analyzed in the previous paragraph. Figure 11 The cloud-edge task scheduling distribution is shown, and as d max increases, the proportion of tasks scheduled to the cloud server for processing also presents a certain rise.

[0128] Based on the above scheme and the above verification, the present application establishes an electricity-carbon-aware computing power network task scheduling model, and on this basis, deduces an optimization problem of minimizing the long-term average electricity-carbon cost while meeting the overall system stability and resource constraints of the computing power network. At the same time, the present application proposes a dynamic approximate algorithm to solve the problem, and theoretically analyzes the optimization interval and performance thereof. Finally, the effectiveness of the method of the present application is verified through simulation experiments, and the influence of different parameters on the performance of the algorithm is analyzed. The experimental results show that the method of the present application has good performance and applicability in terms of long-term average electricity-carbon cost and system stability.

[0129] An electricity-carbon-aware computing power network task scheduling system based on electricity-carbon awareness, comprising:

[0130] A model construction module for introducing a carbon tax mechanism to construct an electricity-carbon-aware computing power network task scheduling model;

[0131] A problem definition module for generating a problem of minimizing system electricity-carbon cost based on the computing power network task scheduling model;

[0132] A solution module for transforming the problem based on Lyapunov optimization and solving it by using a DPRA algorithm to obtain an optimal solution of the scheduling scheme.

[0133] The contents in the above method embodiments are all applicable to the present system embodiments, the present system embodiments specifically implement the same functions as the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0134] An electricity-carbon-aware computing power network task scheduling device based on electricity-carbon awareness, comprising:

[0135] at least one processor;

[0136] at least one memory for storing at least one program;

[0137] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned method for scheduling tasks in an algorithm power network based on electric carbon perception.

[0138] The content in the above method embodiments is applicable to the device embodiments, the device embodiments specifically implement the same functions as the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0139] A storage medium, wherein the storage medium stores processor-executable instructions, and the processor-executable instructions, when executed by a processor, are used to implement the above-mentioned method for scheduling tasks in an algorithm power network based on electric carbon perception.

[0140] The content in the above method embodiments is applicable to the storage medium embodiments, the storage medium embodiments specifically implement the same functions as the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0141] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the above-mentioned embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. A method for task scheduling of a computing power network based on electric carbon perception, characterized in that, The method comprises the following steps: Introducing a carbon tax mechanism to build an electricity-carbon-aware computing power network task scheduling model; Based on the computing power network task scheduling model, a problem of minimizing the system electricity-carbon cost is generated; Based on Lyapunov optimization, the problem is transformed and solved by using a coupling probabilistic rounding algorithm to obtain an optimal solution of the scheduling scheme; The electricity-carbon-aware computing power network task scheduling model specifically comprises: A wireless terminal device communicates with a regional access point and uses a computing service; The regional access point schedules the received tasks, communicates with an edge server and a cloud server, considers the overall electricity-carbon cost of the computing power network, and selects to schedule the tasks to the edge server or the cloud server for processing; The step of generating a problem of minimizing the system electricity-carbon cost based on the computing power network task scheduling model specifically comprises: Defining constraints and an objective function; The constraints include task non-divisible constraints, task complete processing constraints, quality of service constraints, resource constraints, and stability constraints; The objective function is to minimize the long-term average electricity-carbon cost of the overall computing power network system; The expression of the problem of minimizing the system electricity-carbon cost is as follows: Where X represents the set of system task scheduling decision variables, Represents decision variables, This represents the overall carbon cost of the system. Represents the system runtime. Indicates time slot The set of tasks arriving at the access point Represents a collection of servers. i Indicates the required task. j Indicates server, Indicates time slot server Used storage space Indicates the computational load of the task. Indicates server Storage space size, Indicates server In the time slot The corresponding system task is a backlog queue. Indicates server In the time slot The corresponding system task is added. Indicates the upper bound of task processing delay; The step of transforming the problem based on Lyapunov optimization and solving it by using a coupling probabilistic rounding algorithm to obtain an optimal solution of the scheduling scheme specifically comprises: Based on Lyapunov optimization, the problem is transformed and the terms irrelevant to the decision variables in the objective function are omitted, and the problem is transformed into a problem of solving each time slot independently; The constraints in the problem of solving each time slot independently are relaxed to linear constraints to obtain a relaxation problem; The optimal solution of the relaxation problem is solved based on a linear programming algorithm, the variables are rounded, and the tasks that break the constraints after scheduling are marked; The marked tasks are scheduled based on a greedy algorithm to obtain an optimal solution of the scheduling scheme.

2. The computing power network task scheduling method based on electric carbon perception according to claim 1, characterized in that, The expression of the problem of solving each time slot independently is as follows: wherein, denotes a time slot a corresponding set of task scheduling decisions, denotes a server a computing speed, V denotes a control parameter introduced by Lyapunov optimization, denotes a server a corresponding unit electricity price, denotes a server a carbon emission intensity, denotes a server a corresponding carbon tax, denotes a proportional factor of electricity resource consumption / server computing amount, denotes a time slot a binary decision variable indicating whether a task i is scheduled to a server j in the time slot.

3. The computing power network task scheduling method based on electric carbon perception according to claim 2, characterized in that, Further comprising: Through simulation and comparative experiments, the performance of the electricity-carbon-aware computing power network task scheduling model is verified.

4. A computing power network task scheduling system based on electric carbon perception, characterized in that, The method for performing the electricity-carbon-aware computing power network task scheduling method according to claim 1 comprises: A model building module for introducing a carbon tax mechanism to build an electricity-carbon-aware computing power network task scheduling model; A problem definition module for generating a problem of minimizing the system electricity-carbon cost based on the computing power network task scheduling model; A solving module for transforming the problem based on Lyapunov optimization and solving it by using a coupling probabilistic rounding algorithm to obtain an optimal solution of the scheduling scheme.

5. An electric carbon perception-based computing power network task scheduling apparatus, characterized in that, Comprise: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1-3.

Images