Time delay and energy consumption optimization method for task hypergraph construction oriented to computing power network

Abstract

The invention relates to the technical field of computing power networks, in particular to a computing power network-oriented time delay and energy consumption optimization method for constructing a task hypergraph, which comprises the following steps of: S1, establishing a directed acyclic graph according to a computing power task, and establishing a hypergraph model according to a dependency relationship of the directed acyclic graph; s2, calculating task time and task energy consumption according to the hypergraph model, and minimizing total task time and task energy consumption; and S3, obtaining an unloading decision model according to the hypergraph model and the MDP model, training the unloading decision model according to a near-end strategy optimization PPO algorithm, and finally obtaining a task unloading prediction network. According to the method, the hypergraph model is established through the dependency relationship of the DAG, the topological relationship among the tasks representing the computing power and the different attribute dependency relationship and the spatial dependency relationship among the different tasks, under the condition that multiple hyperedges are established, the multi-association attributes of different partitions are balanced by using hypergraph partitions, the reward function is designed through the hypergraph model, and the reward effect is improved. The efficient unloading optimization effect is achieved, and the final unloading energy consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the field of computing power network technology, and in particular to a method, device and storage medium for optimizing latency and energy consumption of building a task hypergraph for a computing power network. Background Art

[0002] With the rapid development of Internet technologies such as big data and artificial intelligence, the generation of massive amounts of data and increasingly powerful computing power have provided strong support for the training and reasoning of large models. Computing Power Network (CPN) is a new architecture that deeply integrates computing and networking. It is a new information infrastructure that takes computing as the center and the network as the foundation and provides computing resource services. For large computing tasks, its training can be accelerated by task offloading. Task offloading determines how to offload computing tasks to the resource pool for further intensive computing and ensure the quality of computing resource services. However, the energy consumption of CPN has become a core issue. Recent reports show that data centers alone contribute 1.54% of total carbon dioxide emissions. The task offloading problem is closely related to energy consumption, and important factors that lead to increased energy consumption include the lack of exploration of multiple dependencies of heterogeneous tasks and the low efficiency of offloading due to information loss.

[0003] Computing tasks lack information exploration of multiple dependencies and the limitations of graph representation. The computing tasks in CPN are heterogeneous and complex. The CPN framework generally treats tasks as similar entities, but many computing tasks can actually be decomposed or reused. Some work divides tasks into directed acyclic graphs in a fine-grained manner. However, for heterogeneous tasks and complex multi-task relationships, simple graph structure representations are already somewhat weak and cumbersome. In addition to simple temporal dependencies, there are other relationships between tasks, which are often not considered by existing technologies.

[0004] Current task offloading methods usually use heuristic algorithms or neural networks for task allocation. However, due to the lack of information on the representation of the task side and the resource side, the lack of information will lead to incomplete consideration of the problem, which will reduce the final offloading efficiency. In addition, the low adaptability of the reward function design of many reinforcement learning algorithms to the final goal will also lead to unsatisfactory final training results. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and provide a method for optimizing the latency and energy consumption of building a task hypergraph for a computing network, comprising the following steps: S1: Establish a directed acyclic graph according to the computing power task, and establish a hypergraph model according to the dependency relationship of the directed acyclic graph; S2: Calculate the task time and task energy consumption according to the hypergraph model, and minimize the overall task time and task energy consumption according to the task time and the task energy consumption; S3: Obtain an offloading decision model according to the hypergraph model and the MDP model, train the offloading decision model according to the proximal strategy optimization PPO algorithm, and finally obtain a task offloading prediction network.

[0006] Preferably, in step S1, establishing a hypergraph model according to the dependency relationship of the directed acyclic graph further comprises: S11: The directed acyclic graph is , the nodes of the directed acyclic graph represent tasks and the dependency pairs between tasks, the dependency pairs between tasks are connected through the edge set It means that each directed edge Defined as , where parent task j should be executed before child task i, and each task cannot be interrupted during execution; S12: Establishing the hyperedges of the hypergraph model according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections, and obtaining the hypergraph model according to the hyperedges , where V is a non-empty finite vertex set, i.e., a task set, is a family of subsets of V called hyperedges, such that and ; S13: Divide the hypergraph model to obtain new hyperedges for representing global information.

[0007] Preferably, in step S12, the hyperedges of the hypergraph model are established according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections: The hyperedges of the hypergraph model are established according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections. , as shown below: in, is the dependency relationship of the directed acyclic graph DAG, is the dependency of network attributes, Dependencies for local connections; Wherein, the DAG hyperedge is obtained according to the dependency relationship of the directed acyclic graph DAG , as shown below: Among them, the network attribute hyperedge is obtained according to the dependency relationship of network attributes. , as shown below: Among them, network attributes The vertex subset of can be expressed as , and the network attributes Corresponding hyperedge network properties The vertex subset of can be expressed as ; Among them, the k-hop hyperedge is obtained according to the dependency of the local connection, as shown below: Among them, the k-hop field is established according to the dependency relationship of local connections. The k-hop field is , which is defined as the set of nodes that are separated by a k-hop distance from one task node to another.

[0008] Preferably, in step S13, dividing the hypergraph model to obtain new hyperedges for representing global information further comprises: According to the cut metric cuts(·), the vertex set of the hypergraph model H is divided into k non-overlapping blocks. , re-evaluating the partitioning of various hyperedge attributes can form new hyperedges for representing global information , task dependencies are as follows: in, is a factor that balances the size of each block, represents the jth block, where a block is a set of vertices; The cutting metric cuts(·) is as follows: .

[0009] Preferably, in step S3, obtaining an offloading decision model according to the hypergraph model and the MDP model further includes: Based on reinforcement learning, the task offloading process is modeled as MDP. For the hypergraph H composed of computing tasks, the state space S, action space A and reward function R are defined in the MDP. The MDP state of offloading the i-th task can be expressed as the current hypergraph H and the offloading decision of the previous i tasks: The action space can be simply represented as the offloading decision of the i-th task . The path with the longest duration is identified as the critical path and is marked as , others are , during the scheduling process, the reward is expressed as the reward function can be expressed as the increment of the optimization target P: in and They are the delay weights of tasks on the critical path and non-critical path respectively. The reward function will encourage learning in the direction of reducing energy consumption and delay on the critical path. The complete time in the target often depends on the completion time of tasks on the critical path in many cases. Increasing the weights of tasks on the critical path can improve the training effect. Uninstalled policy functions is in state Next select action The probability of generating an overall unloading decision can be expressed as: Preferably, in step S3, the offloading decision model is trained according to the proximal policy optimization PPO algorithm, further comprising: The proximal policy optimization (PPO) algorithm is based on the policy parameters Gradient ascent continuously optimizes the unloading strategy: in Represents the learning rate, which uses the cosine learning rate adjustment scheme. Represents the gradient of the loss function calculation. After Z steps, it will replace the policy function. The gradient calculation uses the Adam first-order approximation.

[0010] Preferably, the loss function is calculated based on the cropping similarity loss and the MSE loss, further comprising: The calculation formula of the loss function is as follows: in, is the MSE loss of the value function, is the cropping similarity loss; Among them, the calculation formula for cropping similarity loss is as follows: in, Restrict the probability distribution ratio to is the generalized advantage estimation function; The calculation formula of the new and old strategy ratio of the jth task is as follows: Among them, the i-th task strategy parameter of the old strategy function is For a single learning task Sampling to obtain its trajectory , used to update the i-th task strategy parameter of the new strategy function for multiple rounds , and are the old policy parameters and the policy parameters that need to be updated; Among them, the generalized advantage estimation function The calculation formula is as follows: in, is the discount factor used to calculate the cumulative error, are parameters that control the bias and variance.

[0011] Preferably, the MSE loss is calculated according to the network value function, including: in, represents the network value function, represents the MDP state of the jth task, represents the actual estimated value function; in, represents the MDP reward of the j+kth task.

[0012] Based on the same concept, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the delay and energy consumption optimization method for constructing a task hypergraph for a computing power network as described in any one of the embodiments.

[0013] Based on the same concept, the present invention also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the latency and energy consumption optimization method for constructing a task hypergraph for a computing power network as described in any of the embodiments.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention establishes a hypergraph model through the dependency relationship of a directed acyclic graph. In addition to representing the topological relationship between computing tasks, it can also represent the different attribute dependencies and spatial dependencies between different tasks. When establishing multiple hyperedges, hypergraph partitions are used to balance the multi-association attributes of different partitions. (2) The present invention involves a reward function that is more suitable for the unloading target through a hypergraph model, thereby achieving a highly efficient unloading optimization effect and reducing the final unloading energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0016] Figure 1 A flowchart of a method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to the present invention; Figure 2 The weight training effect diagram of improving the tasks on the critical path of the present invention; Figure 3 Schematic diagram for comparing the training effects of different types of hyperedges in the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0018] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0019] See also Figure 1 As shown, the latency and energy consumption optimization method for constructing a task hypergraph for a computing power network provided in this embodiment includes the following steps: S1: Establish a directed acyclic graph according to the computing power task, and establish a hypergraph model according to the dependency relationship of the directed acyclic graph. Specifically, in this embodiment, in the autonomous driving scenario, formal data and target recognition are collected, and each task depends on the successful completion of all previous tasks.

[0020] In order to effectively represent the multiple dependencies of tasks, this embodiment develops a hypergraph construction model to provide a richer and more comprehensive task view. The hypergraph is chosen as the representation method mainly because using bidirectional edges in DAG will lead to explosive growth.

[0021] Preferably, in step S1, establishing a hypergraph model according to the dependency relationship of the directed acyclic graph further comprises: S11: The directed acyclic graph is , the nodes of the directed acyclic graph represent tasks and the dependency pairs between tasks, the dependency pairs between tasks are connected through the edge set It means that each directed edge Defined as , where the parent task j should be executed before the child task i, and each task cannot be interrupted during execution. Specifically, in this embodiment, the application is modeled as a directed acyclic graph DAG.

[0022] S12: According to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections, the hyperedges of the hypergraph model are established, and the hypergraph model is obtained according to the hyperedges, where V is a non-empty finite vertex set, i.e., a task set. is a family of subsets of V called hyperedges, such that and .

[0023] Preferably, in step S12, the hyperedges of the hypergraph model are established according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections: The hyperedges of the hypergraph model are established according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections. , as shown below: in, is the dependency relationship of the directed acyclic graph DAG, is the dependency of network attributes, is the dependency of the local connection. Specifically, in this embodiment, according to the considered scenario, a task corresponds to a vertex of H, and the task set is equal to the vertex set of H, that is, , retaining the dependency of the original directed acyclic graph DAG in the hypergraph representation, so the edge set E can be directly transformed into a set of 2-uniform hyperedges, maintaining the dependency with DAG to obtain DAG hyperedges , DAG hyperedge It can fully cover the sequential dependencies, which is the basic information required for task scheduling; Wherein, the DAG hyperedge is obtained according to the dependency relationship of the directed acyclic graph DAG , as shown below: Among them, the network attribute hyperedge is obtained according to the dependency relationship of network attributes. , as shown below: Among them, network attributes The vertex subset of can be expressed as , and the network attributes Corresponding hyperedge network properties The vertex subset of can be expressed as Specifically, in this embodiment, in order to accurately represent the dependency relationship between network attributes, the corresponding hyperedges containing network features are established with similar bandwidth attributes. The vertex subset of can be expressed as , and the attributes The corresponding hyperedge can be expressed as: in is a collection of bandwidth attributes and models dependencies in the attribute space at the group level. Similarly, other network attributes are defined, following the same hyperedge principle. Specifically, and Represents different network attribute hyperedges, including resolution, bit rate, and mobility, which together constitute the network hyperedge set ; Among them, the k-hop hyperedge is obtained according to the dependency of the local connection, as shown below: Among them, the k-hop field is established according to the dependency relationship of local connections. The k-hop field is , which is defined as the set of nodes that are separated by a k-hop distance from one task node to another.

[0024] S13: Divide the hypergraph model to obtain new hyperedges for representing global information.

[0025] Preferably, in step S13, dividing the hypergraph model to obtain new hyperedges for representing global information further comprises: According to the cut metric cuts(·), the vertex set of the hypergraph model H is divided into k non-overlapping blocks. , re-evaluating the partitioning of various hyperedge attributes can form new hyperedges for representing global information , task dependencies are as follows: in, is a factor that balances the size of each block, represents the jth block, where a block is a set of vertices; The cutting metric cuts(·) is as follows: .

[0026] S2: Calculate the task time and task energy consumption according to the hypergraph model, and minimize the overall task time and task energy consumption according to the task time and the task energy consumption.

[0027] Assume that the uninstall decision given by the system is ,in , indicating that the i-th task will be offloaded to In the resource pool. Once the offloading decision is made, the tasks will start to execute in the specified order. , assuming , then the time required for uninstallation is: in and They represent the transmission time and execution time respectively. and Respectively represent the transmission rates of the data upload channel and download channel.

[0028] For decomposable tasks, the overall time calculation needs to follow the topological structure of the tasks. The earliest time each task can start executing can be expressed as: Representation Task All predecessor tasks: The task nodes in the hypergraph can be arranged into a sequence according to the earliest execution time of each task, and they will be gradually unloaded to the corresponding resource pool after the completion of the predecessor task.

[0029] For VMs in the resource pool, there are two states during task offloading: one is the waiting state when there is no task running, and the other is the running state when there is a task. These two states can be converted to each other. We use Indicates the status of the vth task type running on the VM of the uth CPN resource pool. 0 indicates waiting and 1 indicates running. Indicates the expected availability time of the VM. When the VM is not running tasks, is equal to the current time, otherwise it is equal to the time of the previous waiting state Plus the tasks currently running on it Uninstall time: In order to calculate the estimated completion time of each task, a waiting queue needs to be established for each VM. Abbreviated as , the offloaded tasks are queued according to their position in the task sequence, assuming that the task The position assigned to the waiting queue is , qlen represents the queue length. Note that the tasks in the task queue indicate that the tasks are still waiting for their execution, so the expected completion time It can be expressed as: According to the topological structure of the hypergraph task, the earliest start time of each task can be calculated: in Assume that It indicates that there is no practical exit task, which indicates that all tasks are completed. Then the total time of the overall task unloading can be expressed as: Compared with latency calculation, energy consumption calculation is simpler. Similar to latency calculation, tasks will generate corresponding energy consumption during transmission and execution. : Among them represents the transmission power, represents the effective switch capacitance of the u-th resource pool, so the total energy consumption of the task execution can be expressed as: Since the calculation of energy consumption does not need to consider the topological relationship between the tasks represented by the hypergraph, the total energy consumption can be expressed as: By establishing a computational model, we can obtain the optimization goal of the task offloading model. The optimization goal is to minimize the overall latency and energy consumption under the condition that all tasks can be offloaded: S3: Obtain an offloading decision model according to the hypergraph model and the MDP model, train the offloading decision model according to the proximal strategy optimization PPO algorithm, and finally obtain a task offloading prediction network.

[0030] In step S3, obtaining an offloading decision model according to the hypergraph model and the MDP model further includes: Based on reinforcement learning, the task offloading process is modeled as MDP. For the hypergraph H composed of computing tasks, the state space S, action space A and reward function R are defined in the MDP. The MDP state of offloading the i-th task can be expressed as the current hypergraph H and the offloading decision of the previous i tasks: The action space can be simply represented as the offloading decision of the i-th task , The path with the longest duration is identified as the critical path and is marked as , others are , during the scheduling process, the reward is expressed as the reward function can be expressed as the increment of the optimization target P: in and are the task delay weights on the critical path and the non-critical path, respectively. The reward function encourages learning in the direction of reducing energy consumption and delay on the critical path. The complete time in the target often depends on the task completion time on the critical path in many cases. Increasing the weight of the task on the critical path can improve the training effect. Specifically, in this embodiment, please refer to Figure 3 As shown in the figure, if the delay on the critical path is reduced and the delay of the nodes on the non-critical path is increased, even if the total delay is the same, the modified maximum delay is 25, while it was 30 before, which has an optimization effect on our problem P.

[0031] Uninstalled policy functions is in state Next select action The probability of generating an overall unloading decision can be expressed as: Specifically, in this embodiment, as the unloading process continues, the corresponding states, rewards and actions will continue to change. The strategy function and the overall reward function can be learned through some deep neural networks to achieve the effect of optimizing the target.

[0032] Preferably, in step S3, the offloading decision model is trained according to the proximal policy optimization PPO algorithm, further comprising: The proximal policy optimization (PPO) algorithm is based on the policy parameters Gradient ascent continuously optimizes the unloading strategy: in Represents the learning rate, which uses the cosine learning rate adjustment scheme. Represents the gradient of the loss function calculation. After Z steps, it will replace the policy function. The gradient calculation uses the Adam first-order approximation.

[0033] Preferably, the loss function is calculated based on the cropping similarity loss and the MSE loss, further comprising: The calculation formula of the loss function is as follows: in, is the MSE loss of the value function, is the cropping similarity loss; Among them, the calculation formula for cropping similarity loss is as follows: in, Restrict the probability distribution ratio to is the generalized advantage estimation function; The calculation formula of the new and old strategy ratio of the jth task is as follows: Among them, the i-th task strategy parameter of the old strategy function is For a single learning task Sampling to obtain its trajectory , used to update the i-th task strategy parameter of the new strategy function for multiple rounds , and are the old policy parameters and the policy parameters that need to be updated; Among them, the generalized advantage estimation function The calculation formula is as follows: in, is the discount factor used to calculate the cumulative error, are parameters that control the bias and variance.

[0034] Preferably, the MSE loss is calculated according to the network value function, including: in, represents the network value function, represents the MDP state of the jth task, represents the actual estimated value function; in, represents the MDP reward of the j+kth task.

[0035] Through the above training method, a highly generalized task offloading prediction network can be obtained, which can capture the relationship between tasks in many complex situations and allocate the correct computing requirements to each task.

[0036] See also Figure 3 As shown, according to the DAG (the dependency of the original DAG of the task ), DAG and k-hop (the original DAG dependency of the task and local connection structure )DAG and network (dependencies of the original DAG of the task and network properties ), DAG and k-hop and network (the original DAG dependency of the task , local connection structure and network properties ) to compare the training effects of different types of hyperedges.

[0037] Figure 3 As shown in Figure (a) in Figure 2, compared with the task hypergraph containing only the original DAG dependencies, the two types of hyperedges ( and ), which proves the effectiveness of the attributed hyperedges proposed in this embodiment. However, the performance of the three hyperedge types (0.493) in Figure (b) is almost equivalent to the performance of the two hyperedge types (0.496) in the above figure. This shows that simply adding attributed hyperedges does not necessarily improve scheduling efficiency. Figure (b) illustrates the implementation of the hypergraph partitioning method, with an overall reward increase of 5%. Hypergraph partitioning can integrate global information based on complex task topology and provide more critical representation information, including detailed insights into the internal structure of hyperedges and the correlations between different hyperedges.

[0038] Example 2 In some embodiments of the present application, a computer device is also provided, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the method for optimizing latency and energy consumption for constructing a task hypergraph for a computing power network as described in any one of Example 1.

[0039] The present invention also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the latency and energy consumption optimization method for constructing a task hypergraph for a computing power network as described in any one of Example 1.

[0040] It is understandable that for the aforementioned methods for optimizing the delay and energy consumption of constructing a task hypergraph for a computing power network, if they are all implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0041] Computer readable storage media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program codes are carried. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program codes contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination thereof.

[0042] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for optimizing latency and energy consumption of constructing a task hypergraph for a computing network, characterized in that: The following steps are involved: S1: Establish a directed acyclic graph according to the computing power task, and establish a hypergraph model according to the dependency relationship of the directed acyclic graph; S2: Calculate the task time and task energy consumption according to the hypergraph model, and minimize the overall task time and task energy consumption according to the task time and the task energy consumption; S3: Obtain an offloading decision model according to the hypergraph model and the MDP model, train the offloading decision model according to the proximal strategy optimization PPO algorithm, and finally obtain a task offloading prediction network.

2. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 1, characterized in that: In step S1, a hypergraph model is established according to the dependency relationship of the directed acyclic graph, further comprising: S11: The directed acyclic graph is The nodes of the directed acyclic graph represent tasks and the dependency pairs between tasks, the dependency pairs between tasks are connected through the edge set It means that each directed edge Defined as The parent task j should be executed before the child task i, and each task cannot be interrupted during execution; S12: Establishing the hyperedges of the hypergraph model according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections, and obtaining the hypergraph model according to the hyperedges in is a non-empty finite set of vertices, i.e., a set of tasks, is called a hyperedge The subset family, such that and S13: Divide the hypergraph model to obtain new hyperedges for representing global information.

3. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 2, characterized in that: In step S12, the hyperedges of the hypergraph model are established according to the dependencies of the directed acyclic graph DAG, the dependencies of the network attributes and the dependencies of the local connections: The hyperedge ε of the hypergraph model is established according to the dependency relationship of the directed acyclic graph DAG, the dependency relationship of the network attributes and the dependency relationship of the local connection hy , as shown below: Among them, ε DAG is the dependency relationship of the directed acyclic graph DAG, is the dependency of network attributes, Dependencies for local connections; Among them, the DAG hyperedge ε is obtained according to the dependency relationship of the directed acyclic graph DAG DAG , as shown below: ε DAG ={{v j ,v i }|(v j ,v i )∈ε} Among them, the network attribute hyperedge is obtained according to the dependency relationship of network attributes. As shown below: Among them, the network attribute β b The vertex subset of can be represented as N a (β b ), and the network property β b The corresponding hyperedge network attribute β b The vertex subset of can be represented as N a (β b ); Among them, the k-hop hyperedge is obtained according to the dependency of the local connection, as shown below: Among them, the k-hop field is established according to the dependency relationship of local connections. The k-hop field is That is, it is defined as a set of nodes that are separated by a k-hop distance from one task node to another.

4. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 3, characterized in that: In step S13, the hypergraph model is divided to obtain new hyperedges for representing global information, further comprising: According to the cutting metric cuts(·), the hypergraph model The vertex set of is divided into k disjoint blocks Re-evaluating the partitioning of various hyperedge attributes can form new hyperedges for representing global information Task dependencies are as follows: Where,∈ is a factor that balances the size of each block, st|V j | represents the jth block, where a block is a set of vertices; The cutting metric cuts(·) is as follows:

5. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 4, characterized in that: In step S3, obtaining an offloading decision model according to the hypergraph model and the MDP model further includes: Based on reinforcement learning, the task offloading process is modeled as an MDP. Defining the state space in an MDP Action Space And the reward function The MDP state of offloading the i-th task can be represented as the current hypergraph And the offloading decision of the first i tasks: The action space can be simply represented as the offloading decision a of the i-th task i . The path with the longest duration is identified as the critical path and is labeled σ m =1, others are σ m = 0, during the scheduling process, the reward is expressed as the reward function can be expressed as the increment of the optimization target P: where w h and w l They are the delay weights of tasks on the critical path and non-critical path respectively. The reward function will encourage learning in the direction of reducing energy consumption and delay on the critical path. The complete time in the target often depends on the completion time of tasks on the critical path in many cases. Increasing the weights of tasks on the critical path can improve the training effect. The unloaded policy function π is in state s i Next select action a i The probability of generating an overall unloading decision can be expressed as:

6. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 5, characterized in that: In step S3, the offloading decision model is trained according to the proximal policy optimization PPO algorithm, further comprising: The proximal policy optimization PPO algorithm is based on the policy parameter θ i Gradient ascent continuously optimizes the unloading strategy: Where ω represents the learning rate, and the learning rate uses the cosine learning rate adjustment scheme. Represents the gradient of the loss function calculation. After z steps, the policy function will be replaced. The gradient calculation uses the Adam first-order approximation.

7. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 6, characterized in that: The loss function is calculated based on the cropping similarity loss and the MSE loss, further including: The calculation formula of the loss function is as follows: in, is the MSE loss of the value function, is the cropping similarity loss; Among them, the calculation formula for cropping similarity loss is as follows: in, It means limiting the probability distribution ratio to [1-∈, 1+∈], is the generalized advantage estimation function; The calculation formula of the new and old strategy ratio of the jth task is as follows: Among them, the i-th task strategy parameter of the old strategy function is For a single learning task t i The trajectory τ is obtained by sampling and used to update the i-th task strategy parameter π(a i |s i θ i ), and θ i are the old policy parameters and the policy parameters that need to be updated; Among them, the generalized advantage estimation function The calculation formula is as follows: Among them, γ∈[0,1] is the discount factor used to calculate the cumulative error, and λ is the parameter that controls the deviation and variance.

8. The method for optimizing latency and energy consumption of constructing a task hypergraph for a computing power network according to claim 7, characterized in that: The MSE loss is calculated based on the network value function, including: Among them, v π represents the network value function, s j represents the MDP state of the jth task, represents the actual estimated value function; Among them, r j+k represents the MDP reward of the j+kth task.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the latency and energy consumption optimization method for constructing a task hypergraph for a computing power network as described in any one of claims 1 to 8.

10. A storage medium storing computer-readable instructions, characterized in that: When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the latency and energy consumption optimization method for constructing a task hypergraph for a computing power network as described in any one of claims 1 to 8.

Images