Task scheduling method and device, and storage medium

Abstract

The application discloses a task scheduling method and device and a storage medium, and relates to the technical field of big data. The method comprises the following steps: analyzing a target task submitted by a target user to obtain target information for the target task, wherein the target information is used to represent the attributes and requirements of the target task; updating a target model according to the target information, wherein the target model at least comprises a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that the task time delay and the task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain the carbon emission amount of completing the target task; generating a target scheduling strategy meeting the target constraint condition and the scheduling target for the target task based on the target model; and scheduling and processing the target task through the target scheduling strategy. The application solves the technical problem that the traditional computing power network scheduling mode in the big model training scene is insufficiently optimized in terms of energy consumption, time delay and carbon emission.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data, in particular to a task scheduling method and device and a storage medium. BACKGROUND

[0002] With the rapid development of the field of artificial intelligence (AI), especially the large-scale application of deep learning models, the demand for computing power has increased exponentially. Large-scale model training not only requires high-performance computing resources, but also involves a large amount of data transmission and processing activities, which poses unprecedented challenges to traditional computing power network scheduling. Traditional scheduling methods often focus on the optimization of a single indicator, such as computing efficiency or network bandwidth, and are not sufficient when faced with the high energy consumption and high latency requirements of AI large model training. Especially in energy-intensive training tasks, simply pursuing speed while ignoring energy consumption and carbon emissions can lead to serious energy waste and environmental pollution, which goes against the concept of green computing.

[0003] In traditional computing power network scheduling, resource allocation often adopts a static or single-target optimization strategy, such as only considering minimizing task processing time. However, this mode has obvious drawbacks when dealing with AI large model training. First, model training requires long-term occupation of a large amount of computing resources, and the resulting high energy consumption can exceed expectations, leading to a sharp rise in energy costs. Second, the latency of data transmission is an important factor limiting task efficiency, especially in distributed training scenarios, where communication efficiency between nodes directly affects overall training performance. Moreover, carbon emissions are not taken into account.

[0004] To address the above problems, no effective solutions have been proposed so far. SUMMARY

[0005] The present application provides a task scheduling method, device and storage medium to at least solve the technical problem of insufficient optimization in terms of energy consumption, latency and carbon emissions in the traditional computing power network scheduling method in the large model training scenario.

[0006] According to one aspect of the present application, a task scheduling method is provided, comprising: analyzing a target task submitted by a target user to obtain target information for the target task, wherein the target information is used to represent the attributes and requirements of the target task; updating a target model according to the target information, wherein the target model at least includes a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that the task latency and task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain the carbon emissions of completing the target task; generating a target scheduling strategy that satisfies the target constraint condition and the scheduling target for the target task based on the target model; and scheduling and processing the target task through the target scheduling strategy.

[0007] Optionally, the target model is obtained by the following steps: constructing a network model according to network environment information, wherein the network environment information at least includes computing capability of network nodes, bandwidth capacity of network links, and network topology structure; constructing a task model according to task requirements, wherein the task requirements at least include task computing requirements, task transmission requirements, and data volume of the task; determining a transmission model according to the network model, the task model, and transmission information, wherein the transmission information at least includes transmission path of the task, transmission delay, and transmission rate; and determining the target model based on the network model, the task model, and the transmission model.

[0008] Optionally, the target model is determined based on the network model, the task model, and the transmission model, including: determining a delay model, an energy consumption model, and a carbon emission model based on the network model, the task model, and the transmission model, wherein the delay model is used to quantify delay of task completion, the energy consumption model is used to quantify energy consumption required for executing the task, and the carbon emission model is used to quantify influence of energy consumption on the environment when the task is executed; determining a scheduling target and N constraint conditions according to the delay model, the energy consumption model, and the carbon emission model, wherein N is an integer greater than or equal to 1, and the N constraint conditions at least include a target constraint condition; and determining the target model according to the scheduling target and the N constraint conditions.

[0009] Optionally, based on the target model, a target scheduling strategy satisfying the target constraint condition and the scheduling target is generated for the target task, including: setting a first variable, wherein the first variable is used to represent iteration number of a target algorithm, and an initial value of the first variable is 1, and the target algorithm is used to generate a scheduling strategy satisfying the target model; generating M first strategies based on target information and using the target algorithm, wherein M is an integer greater than 1; dividing the M first strategies into S levels according to a target manner, wherein S is an integer less than or equal to M, and the target manner is used to divide the M first strategies into levels based on the scheduling target and the constraint condition; performing a target operation on the M first strategies based on the S levels, to obtain T second strategies, wherein T is an integer greater than 1, and the target operation is used to generate variants of the M first strategies according to the M first strategies; increasing the first variable by 1, and merging the M first strategies and the T second strategies to obtain a first strategy set; and determining the target scheduling strategy for the target task according to the first variable and the first strategy set.

[0010] Optionally, determining the target scheduling strategy for the target task according to the first variable and the first strategy set comprises: step one, performing hierarchical division on the strategies in the first strategy set by a target method, and performing crowding degree calculation on the divided strategies in the first strategy set to obtain K third strategies, wherein K is an integer greater than 1, and the crowding degree calculation is used to quantify the distribution density of the strategies in the first strategy set; step two, performing target operation on the K third strategies to obtain J fourth strategies, wherein J is an integer greater than 1; step three, judging whether the first variable is greater than a preset iteration number; step four, if the first variable is less than the preset iteration number, increasing the first variable by 1, combining the K third strategies and the J fourth strategies into a second strategy set, taking the second strategy set as the first strategy set, and repeating steps one to four until the first variable is greater than or equal to the preset iteration number, and determining a fourth strategy meeting the scheduling target and the target constraint condition from the J fourth strategies as the target scheduling strategy.

[0011] Optionally, based on the S levels, performing target operation on the M first strategies to obtain T second strategies: based on the S levels, screening the M first strategies to obtain L first strategies, wherein L is an integer greater than 1 and less than M; exchanging any two first strategies in the L first strategies to obtain R variants of the first strategies, wherein R is an integer greater than 1; and performing random mutation on the R variants of the first strategies to obtain the T second strategies.

[0012] Optionally, performing hierarchical division on the strategies in the first strategy set by a target method, and performing crowding degree calculation on the divided strategies in the first strategy set to obtain K third strategies comprises: taking each strategy in the first strategy set as a fifth strategy; dividing the fifth strategies in the first strategy set into U levels by a target method, wherein U is an integer greater than 1; determining the crowding degree of the fifth strategies in each level; based on the crowding degree of the fifth strategies in each level, extracting K fifth strategies, and taking the K fifth strategies as the K third strategies.

[0013] Optionally, performing scheduling processing on the target task by the target scheduling strategy comprises: determining a target coefficient of the target scheduling strategy, wherein the target coefficient is used to represent the load balancing performance of the target scheduling strategy; when the target coefficient is greater than or equal to a first preset threshold, updating the target scheduling strategy; and when the target coefficient is less than the first preset threshold, performing scheduling processing on the target task by the target scheduling strategy.

[0014] According to another aspect of the present application, a task scheduling apparatus is also provided, comprising: an analyzing unit configured to analyze a target task submitted by a target user to obtain target information of the target task, wherein the target information is used to represent attributes and requirements of the target task; an updating unit configured to update a target model according to the target information, wherein the target model at least comprises a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that a task delay and a task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain carbon emissions for completing the target task; a generating unit configured to generate a target scheduling strategy satisfying the target constraint condition and the scheduling target for the target task based on the target model; and a scheduling unit configured to perform scheduling processing on the target task through the target scheduling strategy.

[0015] According to another aspect of the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program, wherein when the computer program runs, the computer readable storage medium causes a device where the computer readable storage medium is located to perform the task scheduling method described above.

[0016] In the present application, first, a target task submitted by a target user is analyzed to obtain target information of the target task, wherein the target information is used to represent attributes and requirements of the target task, then a target model is updated according to the target information, wherein the target model at least comprises a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that a task delay and a task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain carbon emissions for completing the target task, then a target scheduling strategy satisfying the target constraint condition and the scheduling target for the target task is generated based on the target model, and finally the target task is processed through the target scheduling strategy, that is, through intelligent scheduling under the scheduling target and the constraint condition, the purposes of simultaneously optimizing task completion delay, reducing task energy consumption and controlling carbon emissions are achieved, thereby realizing technical effects of efficiency, energy efficiency and environmental protection of resource scheduling, and further solving the technical problem of insufficient optimization in energy consumption, delay and carbon emissions of a traditional computing power network scheduling mode in a large model training scene. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which are included to provide a further understanding of the present application and are incorporated in and constitute a part of this application, illustrate embodiments of the present application and serve to explain the present application. In the drawings:

[0018] Figure 1 is a flowchart of an optional task scheduling method according to an embodiment of the present application;

[0019] Figure 2 is a system diagram of an optional computing network scheduling according to an embodiment of the present application;

[0020] Figure 3 is a structural diagram of an optional task scheduling method according to an embodiment of the application;

[0021] Figure 4 is a flowchart of an optional NSGA-II algorithm according to an embodiment of the application;

[0022] Figure 5 is a schematic diagram of an optional task scheduling device according to an embodiment of the application. DETAILED DESCRIPTION

[0023] In order to enable persons skilled in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] It should be noted that the collected information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present application are information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary security measures are taken, do not violate public order and good customs, and provide corresponding operation portal for user to choose authorization or refusal. For example, the system and related users or institutions are provided with an interface to provide the user with a corresponding operation portal for the user to choose to agree or refuse the automatic decision result; if the user chooses to refuse, the expert decision process is entered.

[0026] According to the embodiment of the present application, a method embodiment of a task scheduling method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0027] It should be noted that an intelligent scheduling system can be used as the execution subject of the task scheduling method of the embodiment of the present application. It can be understood that the task scheduling method provided by the embodiment of the present application can also be used as the execution subject by other systems or devices, which is not limited in the embodiment of the present application.

[0028] Figure 1 is a flowchart of an optional task scheduling method according to the embodiment of the present application, as shown in Figure 1 The method comprises the following steps:

[0029] Step S101, analyzing the target task submitted by the target user to obtain target information for the target task.

[0030] In step S101, the target information is used to represent the attributes and requirements of the target task.

[0031] Optionally, the target user refers to a customer or a device that needs a computing power network service. In the present application, the target user can be an individual or an enterprise that performs AI model training.

[0032] Optionally, the target task refers to a computing task submitted by the target user to the computing power network, such as a training task of an AI model. These tasks have specific data volume, computing complexity and QoS (Quality of Service, service quality) requirements.

[0033] Optionally, the target information is used to describe the key characteristics of the target task, including the total data volume of the task, the task load (the number of cycles required to process the task), and the user's requirements for the task delay and energy consumption, etc. These information is the basis for intelligent scheduling.

[0034] Optionally, taking the AI model training task as an example, the target task submitted by the target user is to train a visual recognition model. The target information may include the size of the model (total data volume), the computing period required for model training (task load), and the company's tolerance range for task completion time (delay) and energy consumption.

[0035] Step S102, updating the target model according to the target information.

[0036] In step S102, the target model at least includes a scheduling target and a target constraint condition.

[0037] In step S102, the scheduling target is used to represent that the task latency of the target task and the task energy consumption are less than the preset threshold, and the target constraint condition is used to constrain the carbon emission amount of completing the target task.

[0038] Optionally, the target model is used to systematically describe the optimization target and constraint condition of the algorithm network scheduling. Updating the target model is a process of adjusting the model parameters based on the specific needs of the target user and the current state of the algorithm network.

[0039] Optionally, the scheduling target at least includes minimizing task latency and minimizing task energy consumption. Specifically, the target is to ensure that the task reduces the energy consumption and carbon emission of the algorithm node as much as possible while meeting the user QoS requirements.

[0040] Optionally, the target constraint condition refers to the rules or restrictions that must be followed during the optimization process, for example, the total carbon emission of the system cannot exceed the preset threshold. The constraint condition ensures that the optimization result is feasible in the real world.

[0041] In step S103, based on the target model, a target scheduling strategy that meets the target constraint condition and the scheduling target is generated for the target task.

[0042] Optionally, the target scheduling strategy refers to a set of decision schemes generated by an optimization algorithm (such as NSGA-II (Nondominated Sorting Genetic Algorithm II)) according to the updated target model, which is used to guide the allocation of algorithm nodes and network resources. This strategy ensures that the task is reasonably allocated to the nodes in the algorithm network under the conditions of meeting the scheduling target (task latency and energy consumption) and constraint condition (such as carbon emission).

[0043] Optionally, through the NSGA-II algorithm, the intelligent scheduling system can generate the optimal scheduling strategy according to the QoS requirements of the user task and the energy supply situation. This includes selecting the most suitable algorithm node (for example, a node with a higher proportion of clean energy) and the optimal network path to minimize latency and energy consumption while ensuring that carbon emissions are within control.

[0044] In step S104, the target task is scheduled and processed through the target scheduling strategy.

[0045] Optionally, once the target scheduling strategy is determined, the intelligent scheduling system will allocate algorithm and network resources according to the strategy. For example, the user task may be allocated to the algorithm node n1 and the data is transmitted through the network link N1. At the same time, the system will ensure that clean energy is used during the task execution process, and real-time monitoring of carbon emissions will be carried out to ensure that it does not exceed the preset limit.

[0046] As can be known from the contents of steps S101 to S104, in the present application, first, the target task submitted by the target user is analyzed to obtain target information for the target task, wherein the target information is used to represent the attributes and requirements of the target task, then the target model is updated according to the target information, wherein the target model at least includes a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that the task delay and the task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain the carbon emission amount for completing the target task, then based on the target model, a target scheduling strategy meeting the target constraint condition and the scheduling target is generated for the target task, and finally the target task is scheduled by the target scheduling strategy, that is, through the intelligent scheduling under the scheduling target and the constraint condition, the purposes of simultaneously optimizing the task completion delay, reducing the task energy consumption and controlling the carbon emission are achieved, thereby realizing the technical effects of high efficiency, energy efficiency and environmental protection of resource scheduling, and further solving the technical problem of insufficient optimization in energy consumption, time delay and carbon emission of the traditional computing power network scheduling mode in the large model training scene.

[0047] In an optional embodiment, the target model is obtained by the following steps: first, the intelligent scheduling system constructs a network model according to network environment information, wherein the network environment information at least includes the computing power of network nodes, the bandwidth capacity of network links and the network topology structure, then constructs a task model according to task requirements, wherein the task requirements at least include task computing requirements, task transmission requirements and the data volume of the task, then determines a transmission model according to the network model, the task model and transmission information, wherein the transmission information at least includes the transmission path, the transmission delay and the transmission rate of the task, and finally determines the target model based on the network model, the task model and the transmission model.

[0048] Optionally, the network model is used to describe the physical architecture and logical connection of the computing power network, providing a basis for subsequent task matching and resource allocation; the task model is used to represent the characteristics of the task, including its data and computing requirements, which is crucial for scheduling appropriate resources to meet the task requirements; the transmission model optimizes the data transmission process, reduces transmission delay and energy consumption, and improves the efficiency and response speed of the entire system.

[0049] Optionally, the intelligent scheduling system collects and analyzes network environment information, mainly including the computing power of network nodes (such as CPU frequency and available memory), the bandwidth capacity of network links (i.e. data transmission speed) and the topology structure of the network (the connection mode between nodes), which are the basis for computing power network scheduling decisions, and then constructs a network model according to the network environment information.

[0050] Optionally, the intelligent scheduling system analyzes the task requirements, including the computational requirements (the amount of computing resources required to complete the task), transmission requirements (the transmission requirements of the task data), and the data volume of the task. These requirement information helps the system understand the complexity and scale of the task. Then a task model is constructed according to the task requirements.

[0051] Optionally, the intelligent scheduling system further determines a transmission model according to the network model and the task model, considering the transmission path of the task, the transmission delay (the time required for data to move from one node to another node), and the transmission rate. The transmission model ensures that data can be transmitted to the correct destination in the shortest time and with the lowest energy consumption.

[0052] Optionally, based on the models constructed in the foregoing, the intelligent scheduling system determines a target model that can guide the system to meet user requirements while achieving efficient use of resources and environmental friendliness.

[0053] From the above, it can be seen that through the above steps, the intelligent scheduling system can achieve fine and intelligent management of the computing power network. Specifically, the construction of the target model enables the system to: intelligently select the most suitable computing power nodes and network links according to the computing and transmission requirements of the task; minimize the total delay and energy consumption of task processing by reasonably allocating computing and transmission resources; by considering the network topology and link bandwidth, the system can adapt to network changes and ensure efficient resource scheduling even in complex network environments.

[0054] In an optional embodiment, the intelligent scheduling system determines a delay model, an energy consumption model, and a carbon emission model based on the network model, the task model, and the transmission model, wherein the delay model is used to quantify the delay of task completion, the energy consumption model is used to quantify the energy consumption required to execute the task, and the carbon emission model is used to quantify the impact of energy consumption on the environment when executing the task, then determines a scheduling target and N constraint conditions according to the delay model, the energy consumption model, and the carbon emission model, wherein N is an integer greater than or equal to 1, at least including a target constraint condition in the N constraint conditions, and then determines a target model according to the scheduling target and the N constraint conditions.

[0055] Optionally, the network model describes the topology of the computing power network and the connection of each computing power node, including link bandwidth, network delay, and communication energy consumption, etc. attributes; the task model defines the characteristics of the user-submitted task, such as data volume, computational requirements, and QoS requirements, including the delay sensitivity and energy consumption sensitivity of the task, for example, the user-submitted task picture content: a drawing symbol with an arrow. This symbol will be parsed to determine its specific requirements; the transmission model covers the characteristics of data transmission in the network, including transmission delay and transmission energy consumption.

[0056] Optionally, the intelligent scheduling system determines a time delay model, an energy consumption model, and a carbon emission model based on the network model, the task model, and the transmission model. The time delay model is based on the task amount, the network delay, and the processing time of the computing power node to comprehensively calculate the total time delay of task completion. The energy consumption model quantifies the total energy consumption of task execution according to the energy consumption of the computing power node and the energy consumption of data transmission. The carbon emission model calculates the total carbon emission by considering the energy type (clean or non-clean energy) used by the computing power node and the network link.

[0057] Optionally, the scheduling target aims to optimize the time delay and energy consumption of task processing.

[0058] Optionally, the N constraint conditions at least include a target constraint condition, i.e., controlling the carbon emission of completing the task to not exceed the maximum value that the system can tolerate, and in addition, may also include resource limitations of the computing power node, available bandwidth limitations of the network link, etc.

[0059] Optionally, the intelligent scheduling system provides a mathematical framework for intelligent scheduling in combination with the scheduling target and the constraint condition. The target model guides the system to perform optimal allocation of resources while meeting the user QoS requirements, ensuring efficient, low-energy, and environmentally friendly task processing.

[0060] Optionally, Figure 2 is an optional system diagram of algorithm and network scheduling according to an embodiment of the present application, as shown in Figure 2 , mainly including users, computing power nodes, network links, edge gateways, base stations, and in addition, energy suppliers, power grids, and algorithm and network scheduling platforms. Each component part operates independently, making it easy to maintain and upgrade. Users are represented by , computing power nodes are represented by , and network paths are represented by . By selecting computing power service nodes for AI large model training and transmitting data streams through network links to complete respective tasks, these tasks may have different requirements for time delay, network reliability, and other QoS requirements.

[0061] Optionally, Figure 3 is a structural diagram of an optional task scheduling method according to an embodiment of the present application, as shown in Figure 3As shown, the user initiates a task through the computing network scheduling platform, and the task requirements include at least the task type, task volume, and QoS requirements (such as latency, reliability, etc.). After receiving the task request, the computing network scheduling platform performs task allocation. First, it performs preliminary task analysis, including parsing user requirements and obtaining resource conditions. Parsing user requirements includes task data volume, computational complexity, etc. This process generates a scheduling decision, which takes into account relevant decision variables, optimization objectives (scheduling objectives), and relevant constraint conditions. Resource conditions are obtained from computing power suppliers (resources required for task execution) and network suppliers (resources required for task transmission). The network supplier purchases electricity from the electricity trading platform (including non-fossil energy and fossil energy), and the electricity trading platform relies on energy suppliers to provide energy.

[0062] Optionally, if the user sets the model training task as , a two-tuple is formed, where represents the total data volume of the task, measured in MB; specifies the task load, which refers to the number of cycles required to process the task. The computing network scheduling platform intelligently matches computing power nodes and network links based on the user's QoS requirements and the current system state. The platform considers the availability, performance, and energy consumption of computing power nodes, as well as the bandwidth and stability of network links. The allocation results are fed back to the user to confirm the execution details of the task. The computing network scheduling platform intelligently matches computing power nodes and network links based on the user's QoS requirements and the current system state. The platform considers the availability, performance, and energy consumption of computing power nodes, as well as the bandwidth and stability of network links. The allocation results are fed back to the user to confirm the execution details of the task. After receiving the task, the computing power node allocates computing resources based on the task characteristics. It executes the AI large model training task, monitors the task progress and resource usage, and returns the results to the user through the network link after the task is completed.

[0063] Optionally, the computing power node calculates the processing time based on the CPU frequency and the proportion of computing power allocated to the task. The intelligent scheduling system dynamically adjusts the allocation of computing resources in real time to optimize the processing time. If the CPU frequency of the computing power node is , then the time required for the computing power node to process the task can be represented as: where represents the proportion of computing power allocated to the task by the computing power node m. This proportion ranges from 0 to 1. If the value of is low, it means that the computing power supplier only allocates a small portion of its computing power to the task , which will cause the processing time to increase. Conversely, if the value of is high, it means more computing power is allocated to the task , and the processing time will decrease.

[0064] Optionally, the intelligent scheduling system calculates the transmission path delay from the user to the base station and from the base station to the computing power node. The transmission delay of the path from the user to the base station can be represented as: , where represents the transmission rate of the communication link between the user and the base station, and the transmission delay of the path from the base station to the internal computing power node can be represented as: , where represents the transmission rate of the path from the base station to the internal computing power node. Then the delay of the task transmitted to the designated computing power node is represented as: The system reduces transmission delay and improves data transmission efficiency by optimizing network link selection.

[0065] Optionally, the intelligent scheduling system combines the calculation delay and the transmission delay to obtain the total delay of the task. Assuming that the amount of task data after processing does not change significantly, based on the calculation delay and the communication delay, the delay of the task processed by the computing power node m selected by the user is: The system further optimizes the task scheduling and resource allocation strategy based on the total delay feedback.

[0066] Optionally, the intelligent scheduling system calculates the energy consumption of the computing power node based on the task amount and device parameters. Assuming that the user needs to complete a task , and has selected the computing power node m and the network path n after reasonable task scheduling. The computing power node is only responsible for providing its own computing power resources to execute the task submitted by the user, and the network path is only responsible for providing network bandwidth resources to transmit data. The electrical energy consumed by each computing power node for processing is related to the task amount and the parameters of the device itself. Let represent the energy coefficient of the computing power node for consuming electrical energy for processing computing tasks. The energy consumption of the node for processing tasks can be represented as: The system predicts the energy consumption of task execution through the energy consumption model to provide a basis for scheduling decisions.

[0067] Optionally, the intelligent scheduling system calculates the transmission energy consumption from the base station to the edge gateway and from the edge gateway to the computing power node. Compared with the transmission energy consumption of the wired link between the base station and the edge gateway and within the computing power network, the transmission energy consumption from the user to the base station is small and can be ignored. Let the transmission power of the link from the base station to the edge gateway and the transmission power between the edge gateway and the designated computing power node be equal, both being The transmission energy consumption can be represented as: The system reduces the transmission energy consumption by optimizing the transmission power and link selection.

[0068] Optionally, the intelligent scheduling system calculates the total carbon emissions according to the energy consumption carbon emission coefficient. The environmental impact of task execution is evaluated through the carbon emission model, supporting green scheduling.

[0069] Optionally, the intelligent scheduling system formulates the computing power node and network path decision variables of the task to optimize resource allocation. To calculate the total energy consumption of the intelligent scheduling system, the computing power node and network path decision variables of the task need to be formulated. First, the scheduling decision of the task is formulated As shown in equation (1):

[0070] (1)

[0071] wherein, represents the scheduling decision of task on computing power node m, which is 1 if the scheduling strategy is 1, and 0 if the task is not scheduled on computing power node m.

[0072] Similarly, the network path selection decision variable y is shown in equation (2):

[0073] (2)

[0074] wherein, represents the network path n through which task passes, which is 1, and 0 if the task does not pass through network path n.

[0075] Optionally, the intelligent scheduling system processes the total energy consumption of the task which can be represented as shown in equation (3):

[0076] (3)

[0077] wherein, represents the energy consumption of computing power node m processing task , represents the transmission energy consumption of task passing through network path n, represents the scheduling decision of task , represents the network path selection decision variable of task , and U represents the set of users, M represents the set of computing power nodes, and N represents the set of network paths.

[0078] Optionally, it is assumed that the energy consumption carbon emission coefficient (the amount of carbon dioxide emissions per unit energy consumption) This represents the carbon emissions from thermal power generation. For computing nodes... , respectively This indicates the proportion of its thermal energy and clean energy, and meets the following conditions. Similarly, Indicates network link The energy share, while meeting the conditions .

[0079] Optionally, the scheduling decision of computing nodes is based on the AI ​​model training task. With respect to the network path selection decision variable y, the total carbon emissions of the tasks processed by the intelligent scheduling system can be expressed as shown in formula (4):

[0080] (4)

[0081] Optionally, as shown in the delay model, the total delay of the intelligent scheduling system in processing tasks can be expressed as shown in formula (5):

[0082] (5)

[0083] in, Indicates the task processed by computing node m. The time required This represents the transmission delay along the path from the user to the base station. This represents the transmission delay of the path from the base station to the computing node m.

[0084] Optionally, as shown in the energy consumption model, the total energy consumption of the intelligent scheduling system in processing tasks can be expressed as shown in formula (6):

[0085] (6)

[0086] in, Indicates the task processed by computing node m. energy consumption Indicates task Transmission energy consumption when traversing network path n Indicates task Scheduling decisions, Indicates task Network path selection decision variables.

[0087] Optionally, based on the above, the scheduling target can be set as shown in formula (7):

[0088] (7)

[0089] Formula (7) indicates that the scheduling objective is to minimize the energy consumption and latency of task processing.

[0090] Alternatively, the constraints can be as shown in equations (8)-(10):

[0091] (8)

[0092] (9)

[0093] (10)

[0094] in, User Maximum tolerable latency It is the maximum electrical energy that the system can provide. This represents the maximum carbon emissions that the system can tolerate. When there are multiple constraints, the generated target scheduling strategy must satisfy multiple constraints. Formula (8) represents the computing power node selected by the user. Processing tasks The total latency must be less than or equal to the user's latency. The maximum tolerable latency, formula (9) means that the total energy consumption of processing tasks must be less than or equal to the maximum electrical energy that the system can provide, and formula (10) means that the total carbon emissions of processing tasks must be less than or equal to the maximum tolerable carbon emissions of the system.

[0095] As described above, the intelligent scheduling system can optimize task scheduling within a computing network by comprehensively considering various factors. It not only significantly reduces task processing latency and energy consumption but also controls carbon emissions, achieving the goal of green computing. By dynamically adjusting resource allocation, the system ensures load balancing across computing nodes, avoiding resource waste and overloading, thereby improving the overall operating efficiency and response speed of the computing network. Furthermore, by incorporating an environmental impact assessment of energy consumption, the intelligent scheduling system promotes sustainable development and provides technical support for building a low-carbon, environmentally friendly computing network. Overall, this implementation method achieves intelligent, efficient, and green scheduling of the computing network, providing users with high-quality computing services while promoting the rational use of resources and environmental protection.

[0096] In an optional embodiment, the intelligent scheduling system sets a first variable, wherein the first variable is used to represent the number of iterations of a target algorithm, the initial value of the first variable is 1, the target algorithm is used to generate a scheduling strategy satisfying a target model, then M first strategies are generated based on target information by using the target algorithm, wherein M is an integer greater than 1, then the M first strategies are divided into S levels according to a target mode, wherein S is an integer less than or equal to M, the target mode is used to classify the M first strategies according to scheduling targets and constraint conditions, then target operations are performed on the M first strategies based on the S levels, to obtain T second strategies, wherein T is an integer greater than 1, the target operation is used to generate variants of the M first strategies according to the M first strategies, then the first variable is increased by 1, and the M first strategies and the T second strategies are combined to obtain a first strategy set, finally a target scheduling strategy for a target task is determined according to the first variable and the first strategy set.

[0097] Optionally, the intelligent scheduling system first sets a first variable for recording the number of iterations of a target algorithm, and the initial value of the first variable is set to 1. The target algorithm refers to an algorithm for generating a scheduling strategy satisfying a target model. In this scheme, the NSGA-II algorithm, a multi-objective optimization algorithm based on non-dominated sorting, is used to find the optimal scheduling strategy under the constraints of multiple targets such as time delay, energy consumption and carbon emission.

[0098] Optionally, based on the target information, the intelligent scheduling system generates M first strategies using a target algorithm (NSGA-II algorithm), where M is an integer greater than 1. These strategies are the results of preliminary optimization according to the optimization objectives and constraints in the target model, such as task latency, energy consumption, and carbon emission requirements. The system then divides the M first strategies into S levels according to the target method, where S is less than or equal to M and is an integer. The target method refers to the process of ranking strategies based on scheduling objectives and constraints, usually implemented using the fast non-dominated sorting method to reflect the performance level of strategies in each target dimension. Based on the S levels after level division, the system performs target operations on the M first strategies to obtain T second strategies, where T is greater than 1. Target operations refer to the generation of variants of the first strategies through evolutionary algorithm operations (such as selection, crossover, and mutation), which helps to explore a wider strategy space and find better solutions. Each time a target operation is completed, the first variable increases by 1, indicating an increase in the number of algorithm iterations. Then, the system combines the M first strategies and T second strategies to obtain a first strategy set, which ensures that all optimized strategies are considered while maintaining the diversity of the population. Finally, the system determines the optimal target scheduling strategy for the target task based on the current value of the first variable and the content of the first strategy set. This process may involve multiple iterations until the algorithm converges or the first variable reaches the maximum number of iterations set in advance, ensuring that the resulting strategy achieves the best balance point in terms of latency, energy consumption, and carbon emissions.

[0099] From the above, through the above steps, the intelligent scheduling system can: efficient optimization: through multiple iterations, the system can effectively explore the strategy space and find the optimal scheduling strategy that meets the requirements of latency, energy consumption, and carbon emissions by using the characteristics of the NSGA-II algorithm; strategy diversity: the combination of level division and target operation ensures the diversity of the strategy set, which helps to avoid local optimal solutions and find the optimal solution that truly covers all objectives; dynamic adaptability: the strategy updating and selection mechanism in the iteration process enables the system to dynamically adapt to changes in the target model, so that it can quickly generate new optimal scheduling strategies even when the target constraints or scheduling objectives change; resource optimal allocation: the final determination of the target scheduling strategy ensures the optimal allocation of computing nodes and network resources, which can meet the QoS requirements of user tasks and minimize energy consumption and environmental impact, achieving efficient use of resources and green computing.

[0100] In an alternative embodiment, the intelligent scheduling system determines the target scheduling strategy for the target task according to the first variable and the first strategy set, comprising: step one, classifying the strategies in the first strategy set by target method, and calculating the crowding degree of the classified strategies in the first strategy set to obtain K third strategies, wherein K is an integer greater than 1, and the crowding degree calculation is used to quantify the distribution density of the strategies in the first strategy set; step two, performing target operation on the K third strategies to obtain J fourth strategies, wherein J is an integer greater than 1; step three, judging whether the first variable is greater than a preset iteration number; step four, if the first variable is less than the preset iteration number, increasing the first variable by 1, combining the K third strategies and the J fourth strategies into a second strategy set, taking the second strategy set as the first strategy set, repeating steps one to four until the first variable is greater than or equal to the preset iteration number, and determining the fourth strategy that meets the scheduling target and the target constraint condition from the J fourth strategies as the target scheduling strategy.

[0101] Optionally, the intelligent scheduling system uses the fast non-dominated sorting mechanism in the NSGA-II algorithm to sort the strategies in the first strategy set, divides the strategies into different levels according to their performance in multi-objective optimization, i.e. non-dominated levels, then calculates the number of adjacent strategies in the target space around each strategy in each level to quantify the distribution density of the strategies, wherein the greater the crowding degree of a strategy, the more similar performance strategies there are around the strategy in the target space, and vice versa, which means that the strategy is relatively unique in the target space. Based on the classification levels and the calculation of the crowding degree, K strategies are selected from the first strategy set, which are considered to have higher diversity and better performance and can be further evaluated and optimized.

[0102] Optionally, the intelligent scheduling system performs genetic algorithm operations such as selection, crossover and mutation on the selected K strategies, and through these operations, the system generates a new strategy set, i.e. J fourth strategies, based on the K third strategies.

[0103] Optionally, after obtaining the J fourth strategies, it is determined whether the first variable is greater than a preset iteration number, wherein the preset iteration number is one of the termination conditions set in the algorithm, and when the iteration number reaches or exceeds the preset value, the algorithm stops. If the first variable is less than the preset iteration number, the first variable is increased by 1, the K third strategies and the J fourth strategies are combined into a second strategy set (which contains the strategies with better performance in the current iteration and the strategies newly generated by the target operation), and then the second strategy set is taken as the first strategy set of the next iteration, and the above steps are repeated. The algorithm iterates continuously, selects better strategies for combination each time, generates new strategies through the target operation, and determines the fourth strategy that meets the scheduling target and the target constraint condition from the J fourth strategies as the target scheduling strategy until the first variable is greater than or equal to the preset iteration number.

[0104] From the above, it can be seen that through the implementation of the above steps, the intelligent scheduling system can effectively explore and optimize the scheduling strategies to meet the demand for multi-objective optimization in the computing power network. Specifically, the system gradually filters out the best-performing scheduling strategy set through the iterative process of the genetic algorithm. These strategies not only minimize the time delay and energy consumption of task processing, but also do so without exceeding the carbon emission limit, embodying the concept of green scheduling. Through the level division and congestion degree calculation, the system ensures the diversity and performance of the strategy set, avoiding premature convergence to a local optimal solution. The final target scheduling strategy can provide good scheduling effect under multi-objective optimization conditions, improve the operation efficiency of the overall computing power network and the quality of service for users, and contribute to environmental protection. The overall effect is the intelligentization, high efficiency and environmental protection of resource scheduling. Through continuous iteration and optimization, the system can gradually approach the Pareto optimal solution, achieving the best balance between energy consumption, time delay and carbon emission in the computing power network.

[0105] In an optional embodiment, the intelligent scheduling system filters M first strategies based on S levels to obtain L first strategies, wherein L is an integer greater than 1 and less than M, then exchanges any two first strategies in the L first strategies to obtain R variants of the first strategies, wherein R is an integer greater than 1, and finally randomly mutates the R variants of the first strategies to obtain T second strategies.

[0106] Optionally, the intelligent scheduling system filters M first strategies based on S levels obtained previously. The S levels are obtained by level division according to the performance of each strategy in the optimization targets such as time delay, energy consumption and carbon emission through the non-dominated sorting method. From the M first strategies, the system selects L first strategies with better performance, and L is an integer greater than 1 and less than M. This process usually involves evaluating the fitness of strategies in various optimization targets and selecting the best-performing strategy set as the basis for the next generation population.

[0107] Optionally, the intelligent scheduling system performs a crossover operation on any two strategies in the L first strategies to generate R variants of the first strategies. The crossover operation here refers to the crossover mechanism in genetic algorithms, which creates the "offspring" of two randomly selected strategies as "parents", with the purpose of exploring new strategy combinations and possibly finding better solutions. These variants are new strategies generated through the crossover operation, which combine the characteristics from different "parent" strategies, increasing the diversity and exploration ability of the population.

[0108] Optionally, a random mutation operation is performed on the R variants of the first strategies to obtain T second strategies, where T is greater than 1. Mutation operation is similar to mutation in biological evolution, which aims to make minor adjustments to certain random characteristics of the strategy to further expand the search range. The new strategy set generated after mutation may be different from the "parent" strategy in some aspects, which helps the system to jump out of the local optimal solution and explore a wider strategy space.

[0109] From the above, through the above steps, the intelligent scheduling system can: optimize strategy selection: by selecting the best L first strategies from multiple levels, the system ensures that the quality of the offspring strategy set is high, which is beneficial for the subsequent optimization process; enhance strategy diversity: through the crossover operation, the system generates R variants of the first strategies, which increases the diversity of the strategy set and helps the algorithm avoid falling into a local optimal solution, improving the optimization ability; further explore the optimization space: the mutation operation introduces randomness to the strategy set, which helps to explore a wider solution space and improves the probability of finding a global optimal solution.

[0110] In an optional embodiment, the intelligent scheduling system classifies the strategies in the first strategy set by target method, and calculates the crowding degree of the classified strategies in the first strategy set to obtain K third strategies, including: taking each strategy in the first strategy set as a fifth strategy, then classifying the fifth strategies in the first strategy set into U levels by target method, where U is an integer greater than 1, then determining the crowding degree of the fifth strategies in each level, and finally extracting K fifth strategies based on the crowding degree of the fifth strategies in each level, and taking the K fifth strategies as the K third strategies.

[0111] Optionally, the target method refers to a mechanism based on fast non-dominated sorting, which is used to evaluate and sort the strategies in the first strategy set to determine their relative advantages in multi-objective optimization.

[0112] Optionally, the strategies in the first strategy set are divided into U levels by the target method, and U is greater than 1. The level division here is based on the non-dominated sorting result of the strategies in the multi-objective optimization problem, and each level represents a group of strategies with similar performance in the multi-objective optimization dimension.

[0113] Optionally, in each level, the system further calculates the crowding degree of the fifth strategy. The crowding degree reflects the sparsity of the strategy in the target space relative to other strategies, and the higher the crowding degree of the strategy, the fewer similar performance strategies around it, and vice versa. The system extracts K best-performing and high-diversity fifth strategies based on the crowding degree of the fifth strategy in each level, which will be referred to as K third strategies. K is an integer greater than 1, but less than M, M is the size of the initial strategy set. By extracting strategies with high crowding degree, the system can retain strategies with good performance but high uniqueness, thereby maintaining the diversity of the population and avoiding premature convergence to local optimal solutions.

[0114] From the above, the above-mentioned embodiments effectively enhance the ability of the intelligent scheduling system based on the NSGA-II algorithm by introducing fast non-dominated sorting and crowding degree calculation. The specific effects are as follows: level division enhances strategy sorting: by dividing the strategies into multiple levels, the system can more finely identify and classify strategies with different performance levels, which helps to accurately select a set of strategies with excellent performance in multi-objective optimization; crowding degree calculation maintains diversity: the calculation of crowding degree not only considers the performance of the strategy, but also considers the uniqueness of the strategy. By extracting strategies with high crowding degree, the system ensures the diversity of the population, enhances the global search ability of the algorithm, and avoids premature convergence; selection of K third strategies: the extracted K third strategies are strategies that have been selected and evaluated layer by layer. They have high performance and uniqueness while meeting the scheduling objectives and constraint conditions, providing a high-quality strategy set for subsequent strategy optimization and combination.

[0115] In an optional embodiment, the intelligent scheduling system determines a target coefficient of the target scheduling strategy, wherein the target coefficient is used to represent the load balancing performance of the target scheduling strategy. When the target coefficient is greater than or equal to a first preset threshold, the target scheduling strategy is updated, and when the target coefficient is less than the first preset threshold, the target scheduling strategy is used to schedule the target task.

[0116] Optionally, the intelligent scheduling system first focuses on the load balancing performance of the target scheduling strategy, which is represented by the target coefficient. The target coefficient reflects the balance of the computing power network in processing tasks, that is, whether the utilization rates of each node are close, and whether there is excessive load.

[0117] Optionally, the system sets a first preset threshold as the evaluation criterion of load balancing performance. If the calculated target coefficient L is greater than or equal to this threshold, it indicates that the load distribution under the current target scheduling strategy is not ideal and needs to be further optimized. When the target coefficient does not meet the preset threshold, the system will update the target scheduling strategy and redistribute tasks to each computing power node. The updating process may involve re-running the NSGA-II algorithm to generate a new set of scheduling strategies, from which the optimal strategy is selected to improve the load balancing performance. If the target coefficient is less than the first preset threshold, the system considers that the current target scheduling strategy has achieved good load balancing, and at this time the strategy will be executed to schedule user tasks to the corresponding computing power nodes to ensure effective allocation and use of network resources.

[0118] Optionally, the intelligent scheduling system uses variance to reflect the load balancing performance of the system, and the load balancing performance is used to constrain the size of the load processed by different computing power nodes. In this embodiment, the variance of the total amount of computation of the tasks processed by each node is used to reflect the load balancing performance of the edge system. Wherein, the smaller the variance, the better the load balancing effect of the system. The calculation method of the target coefficient is shown in formula (11):

[0119] (11)

[0120] Wherein, represents the total amount of computation of the tasks to be processed by the node after this scheduling, represents the average value of the total amount of computation of the tasks processed by the node, that is, the average number of the total amount of computation of the tasks to be processed by all nodes after scheduling, which can be represented by the task model and the scheduling decision as shown in formula (12):

[0121] (12)

[0122] Wherein, designates the task load, represents the scheduling decision of the task.

[0123] From the above, by calculating and monitoring the target coefficient, the intelligent scheduling system can dynamically evaluate the performance of the scheduling strategy, especially the load balancing, and once imbalance is detected, the system immediately starts the update process to improve the uniformity of resource allocation; by optimizing the target scheduling strategy, the system ensures fair distribution of computing tasks among nodes, avoids resource waste and excessive load, and improves the overall resource utilization and processing efficiency of the computing power network; balanced load distribution helps to reduce system bottlenecks, improve the stability and reliability of the computing power network, and ensure that user tasks can be processed in a relatively stable environment, reducing the risk of system interruption due to single-point failure; the application of load balancing strategy makes the processing of user tasks more rapid and stable, shortens the task response time, and improves user satisfaction and trust.

[0124] In an alternative embodiment, Figure 4 is a flowchart of an alternative NSGA-II algorithm according to an embodiment of the present application, as Figure 4 shown, showing the process from initialization of the population to generation of the final scheduling strategy. The flowchart details the key steps in the algorithm, such as non-dominated sorting, crowding calculation, selection of individuals to form a new parent population, until the iteration number is met and the Pareto optimal solution set is output, i.e. a series of scheduling strategies that achieve the best balance between latency and energy consumption. The detailed process is as follows:

[0125] Before the algorithm starts, node information, energy information, and task information need to be input first, which form the basis for system decision. Node information includes computing resources and status of computing power nodes; energy information describes the status of energy supply; task information includes user-submitted task requirements and characteristics. Based on the target information of the current target task and the target model, a set of scheduling strategies is randomly generated as the initial population (initial population) of the algorithm, and is used as the parent population, each individual representing a possible scheduling strategy, including decision variables for computing power node selection and network path selection.

[0126] Then it is determined whether there is a first generation of sub-population, if not, it is sorted using the fast non-dominated sorting method, and based on the sorting result, a first generation of sub-population is generated through genetic operations such as selection, crossover and mutation (the algorithm uses a binary tournament selection method to decide which individuals to reproduce, then generates new individuals through crossover operation, and finally increases the diversity of the population through mutation operation), when the first generation of sub-population is generated, the algorithm starts iterative evolution, where gen=2 represents that the algorithm enters the second generation of evolution, at this time the population is combined from the first generation of parents and children.

[0127] Then the parent individuals of the last generation and the newly generated offspring individuals are combined into a larger population, which helps to retain the individuals of the last generation with good performance, and determines whether to generate a new parent population, if not, the combined population is quickly non-dominantly sorted, and the solutions are classified according to their non-dominant levels, and the non-dominant level refers to the superiority of the solution in the multi-objective optimization problem relative to other solutions, wherein the first non-dominant front contains the best solutions in the population, and the subsequent fronts are sorted according to the relative advantage of the solutions.

[0128] Then, for the solutions in the same non-dominant level, the crowding degree is calculated, which reflects the density of the solution in the multi-dimensional target space, avoids the algorithm from converging to a local optimum too early, and maintains the diversity of the population, and from the results of sorting and crowding degree calculation, the individuals with good performance are selected to form a new parent population for the next round of genetic operation, and the selection strategy here is based on the non-dominant level and the crowding degree of the solution, to ensure that there is enough diversity in the population.

[0129] After generating the new parent population, the operations such as selection, crossover and mutation are also performed, and then it is determined whether the current iteration number (gen) is greater than the maximum iteration number, if less than the maximum iteration number, the iteration number is increased by 1, and the loop starts from the combination of the parent individuals and the offspring individuals, until the predetermined maximum iteration number is reached or other stop conditions are met, and if greater than or equal to the maximum iteration number, the optimal individual (i.e. the optimal strategy) is selected according to the constraint conditions.

[0130] The embodiments of the present application also provide a task scheduling device, and it should be noted that the task scheduling device of the embodiments of the present application can be used to execute the task scheduling method provided by the embodiments of the present application. The task scheduling device provided by the embodiments of the present application is introduced as follows.

[0131] According to the embodiments of the present application, a device for implementing the above task scheduling method is also provided, Figure 5 is a schematic diagram of an optional task scheduling device according to the embodiments of the present application, as shown in Figure 5 including a parsing unit 501, an updating unit 502, a generating unit 503 and a scheduling unit 504.

[0132] Optionally, the parsing unit 501 is configured to parse the target task submitted by the target user to obtain target information of the target task, wherein the target information is used to represent attributes and requirements of the target task; the updating unit 502 is configured to update a target model according to the target information, wherein the target model at least includes a scheduling target and a target constraint condition, the scheduling target is used to represent that the task time delay and the task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain the carbon emission amount of completing the target task; the generating unit 503 is configured to generate a target scheduling strategy satisfying the target constraint condition and the scheduling target for the target task based on the target model; and the scheduling unit 504 is configured to schedule and process the target task through the target scheduling strategy.

[0133] Optionally, the task scheduling apparatus further includes a first construction unit, a second construction unit, a third construction unit, and a first determination unit. The first construction unit is configured to construct a network model according to network environment information, wherein the network environment information at least includes a computing capability of a network node, a bandwidth capacity of a network link, and a network topology structure. The second construction unit is configured to construct a task model according to task demand, wherein the task demand at least includes a task computing demand, a task transmission demand, and a data volume of the task. The third construction unit is configured to determine a transmission model according to the network model, the task model, and transmission information, wherein the transmission information at least includes a transmission path of the task, a transmission time delay, and a transmission rate. The first determination unit is configured to determine a target model based on the network model, the task model, and the transmission model.

[0134] Optionally, the first determination unit includes a first determination subunit, a second determination subunit, and a third determination subunit. The first determination subunit is configured to determine a time delay model, an energy consumption model, and a carbon emission model based on the network model, the task model, and the transmission model, wherein the time delay model is used to quantify a time delay of completing the task, the energy consumption model is used to quantify an energy consumption required for executing the task, and the carbon emission model is used to quantify an influence of energy consumption on the environment when the task is executed. The second determination subunit is configured to determine a scheduling target and N constraint conditions according to the time delay model, the energy consumption model, and the carbon emission model, wherein N is an integer greater than or equal to 1, and the N constraint conditions at least include the target constraint condition. The third determination subunit is configured to determine the target model according to the scheduling target and the N constraint conditions.

[0135] Optionally, the generating unit 503 comprises a first setting subunit, a first generating subunit, a first dividing subunit, a first processing subunit, a second processing subunit, and a fourth determining subunit. The first setting subunit is configured to set a first variable, wherein the first variable is used to represent the number of iterations of a target algorithm, and the initial value of the first variable is 1, and the target algorithm is used to generate a scheduling strategy satisfying a target model; the first generating subunit is configured to generate M first strategies based on target information by using the target algorithm, wherein M is an integer greater than 1; the first dividing subunit is configured to divide the M first strategies into S levels according to a target manner, wherein S is an integer less than or equal to M, and the target manner is used to perform hierarchical division on the M first strategies based on a scheduling target and a constraint condition; the first processing subunit is configured to perform a target operation on the M first strategies based on the S levels to obtain T second strategies, wherein T is an integer greater than 1, and the target operation is used to generate variants of the M first strategies according to the M first strategies; the second processing subunit is configured to increase the first variable by 1 and combine the M first strategies and the T second strategies to obtain a first strategy set; and the fourth determining subunit is configured to determine a target scheduling strategy for a target task according to the first variable and the first strategy set.

[0136] Optionally, the fourth determining subunit comprises a first dividing module, a first processing module, a first judging module, and a first determining module. The first dividing module is configured to perform hierarchical division on the strategies in the first strategy set by using a target manner, and perform crowding degree calculation on the divided strategies in the first strategy set to obtain K third strategies, wherein K is an integer greater than 1, and the crowding degree calculation is used to quantify the distribution density of the strategies in the first strategy set; the first processing module is configured to perform a target operation on the K third strategies to obtain J fourth strategies, wherein J is an integer greater than 1; the first judging module is configured to judge whether the first variable is greater than a preset iteration number; and the first determining module is configured to, if the first variable is less than the preset iteration number, increase the first variable by 1, combine the K third strategies and the J fourth strategies into a second strategy set, take the second strategy set as the first strategy set, repeat steps one to four until the first variable is greater than or equal to the preset iteration number, and determine a fourth strategy satisfying the scheduling target and the target constraint condition from the J fourth strategies as the target scheduling strategy.

[0137] Optionally, the first processing subunit comprises a first screening module, a second processing module and a third processing module. The first screening module is configured to screen the M first strategies based on the S levels to obtain L first strategies, where L is an integer greater than 1 and less than M. The second processing module is configured to exchange any two first strategies in the L first strategies to obtain R variants of the first strategies, where R is an integer greater than 1. The third processing module is configured to randomly mutate the R variants of the first strategies to obtain T second strategies.

[0138] Optionally, the first dividing module comprises a first determining submodule, a first dividing submodule, a first determining submodule and a second determining submodule. The first determining submodule is configured to take each strategy in the first strategy set as a fifth strategy. The first dividing submodule is configured to divide the fifth strategy in the first strategy set into U levels by a target manner, where U is an integer greater than 1. The first determining submodule is configured to determine the crowding degree of the fifth strategy in each level. The second determining submodule is configured to extract K fifth strategies based on the crowding degree of the fifth strategy in each level, and take the K fifth strategies as K third strategies.

[0139] Optionally, the scheduling unit 504 comprises a fifth determining submodule, a first updating submodule and a first executing submodule. The fifth determining submodule is configured to determine a target coefficient of the target scheduling strategy, where the target coefficient is used to represent the load balancing performance of the target scheduling strategy. The first updating submodule is configured to update the target scheduling strategy when the target coefficient is greater than or equal to a first preset threshold. The first executing submodule is configured to schedule the target task by the target scheduling strategy when the target coefficient is less than the first preset threshold.

[0140] According to another aspect of the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program. When the computer program runs, the computer readable storage medium makes the device where the computer readable storage medium is located execute the task scheduling method described above. The above sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0141] In the above embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0142] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented by other ways. Among them, the above-described device embodiments are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0143] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place or can be distributed to multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0144] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0145] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various program code storage media.

[0146] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A task scheduling method, characterized by, The method comprises the following steps: analyzing a target task submitted by a target user to obtain target information of the target task, wherein the target information is used to represent attributes and requirements of the target task; updating a target model according to the target information, wherein the target model comprises at least a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that the task time delay and the task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain the carbon emission amount of completing the target task; generating a target scheduling strategy for the target task based on the target model, wherein the target scheduling strategy satisfies the target constraint condition and the scheduling target; scheduling the target task through the target scheduling strategy; wherein generating the target scheduling strategy for the target task based on the target model, wherein the target scheduling strategy satisfies the target constraint condition and the scheduling target, comprises: setting a first variable, wherein the first variable is used to represent the number of iterations of a target algorithm, the initial value of the first variable is 1, and the target algorithm is used to generate a scheduling strategy that satisfies the target model; generating M first strategies based on the target information using the target algorithm, wherein M is an integer greater than 1; dividing M first strategies into S levels according to a target method, wherein S is an integer less than or equal to M, and the target method is used to grade M first strategies based on the scheduling target and the target constraint condition; performing a target operation on M first strategies based on S levels to obtain T second strategies, wherein T is an integer greater than 1, and the target operation is used to generate variants of the first strategies according to M first strategies; increasing the first variable by 1, and merging M first strategies and T second strategies to obtain a first strategy set; determining a target scheduling strategy for the target task according to the first variable and the first strategy set; wherein determining a target scheduling strategy for the target task according to the first variable and the first strategy set comprises: Step one: grading strategies in the first strategy set through the target method, and calculating the crowding degree of the strategies in the first strategy set after grading to obtain K third strategies, wherein K is an integer greater than 1, and the crowding degree calculation is used to quantify the distribution density of strategies in the first strategy set; Step two: performing the target operation on K third strategies to obtain J fourth strategies, wherein J is an integer greater than 1; Step three: determining whether the first variable is greater than a preset iteration number; Step four: if the first variable is less than the preset iteration number, increasing the first variable by 1, combining K third strategies and J fourth strategies into a second strategy set, taking the second strategy set as the first strategy set, repeating steps one to four until the first variable is greater than or equal to the preset iteration number, and determining a fourth strategy that satisfies the scheduling target and the target constraint condition from J fourth strategies as the target scheduling strategy.

2. The task scheduling method of claim 1, wherein, The target model is obtained by the following steps: constructing a network model according to network environment information, wherein the network environment information at least includes computing capability of network nodes, bandwidth capacity of network links and network topology structure; constructing a task model according to task requirements, wherein the task requirements at least include task computing requirements, task transmission requirements and data volume of tasks; determining a transmission model according to the network model, the task model and transmission information, wherein the transmission information at least includes transmission path of tasks, transmission delay and transmission rate; determining the target model based on the network model, the task model and the transmission model.

3. The task scheduling method of claim 2, wherein, Determining the target model based on the network model, the task model and the transmission model includes: determining a delay model, an energy consumption model and a carbon emission model based on the network model, the task model and the transmission model, wherein the delay model is used to quantify delay of task completion, the energy consumption model is used to quantify energy consumption required for executing tasks, and the carbon emission model is used to quantify influence of energy consumption on environment when executing tasks; determining the scheduling target and N constraint conditions according to the delay model, the energy consumption model and the carbon emission model, wherein N is an integer greater than or equal to 1, and at least the target constraint condition is included in the N constraint conditions; determining the target model according to the scheduling target and the N constraint conditions.

4. The task scheduling method of claim 1, wherein, Based on S levels, target operation is performed on M first strategies to obtain T second strategies: Based on S levels, screening is performed on M first strategies to obtain L first strategies, wherein L is an integer greater than 1 and less than M; any two first strategies in L first strategies are exchanged to obtain R variants of the first strategies, wherein R is an integer greater than 1; random mutation is performed on the R variants of the first strategies to obtain T second strategies.

5. The task scheduling method of claim 1, wherein, The strategies in the first strategy set are divided into levels by the target method, and the strategies in the divided first strategy set are calculated for congestion degree to obtain K third strategies, including: each strategy in the first strategy set is taken as a fifth strategy; the fifth strategy in the first strategy set is divided into U levels by the target method, wherein U is an integer greater than 1; the congestion degree of the fifth strategy in each level is determined; based on the congestion degree of the fifth strategy in each level, K fifth strategies are extracted, and the K fifth strategies are taken as K third strategies.

6. The task scheduling method of claim 1, wherein, Scheduling processing is performed on the target task by the target scheduling strategy, including: determining a target coefficient of the target scheduling strategy, wherein the target coefficient is used to represent load balancing performance of the target scheduling strategy; when the target coefficient is greater than or equal to a first preset threshold, updating the target scheduling strategy; when the target coefficient is less than the first preset threshold, scheduling processing is performed on the target task by the target scheduling strategy.

7. A task scheduling apparatus characterized by comprising: The task scheduling apparatus is used for implementing the task scheduling method in any one of claims 1 to 6, and comprises: an analysis unit configured to analyze a target task submitted by a target user to obtain target information of the target task, wherein the target information is used to represent attributes and requirements of the target task; an updating unit configured to update a target model according to the target information, wherein the target model at least comprises a scheduling target and a target constraint condition, wherein the scheduling target is used to represent that a task time delay and a task energy consumption of the target task are less than a preset threshold, and the target constraint condition is used to constrain carbon emission amount for completing the target task; a generating unit configured to generate a target scheduling strategy satisfying the target constraint condition and the scheduling target for the target task based on the target model; a scheduling unit configured to perform scheduling processing on the target task through the target scheduling strategy.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, wherein when the computer program runs, the device where the computer readable storage medium is located executes the task scheduling method in any one of claims 1 to 6.

Images