A cloud-edge collaborative elastic scheduling method and system for heterogeneous computing resources

Abstract

The application discloses a cloud-edge collaborative elastic scheduling method and system for heterogeneous computing resources, relates to the technical field of cloud computing and edge computing collaborative scheduling, and unifies the heterogeneous computing resources into a multi-dimensional computing resource package vector, introduces a task completion benefit value and an expected cooperation benefit base of an edge node, constructs a winning collaborative combination optimization mechanism with the maximum collaborative scheduling net benefit as the target, and ensures that the scheduled edge node combination realizes the optimal system net benefit on the premise of meeting the multi-dimensional resource demand. Meanwhile, the application takes the substantial marginal contribution of the edge node in the collaborative combination as the only basis to calculate the actual cooperation benefit value of each participating edge node, ensures that the benefit distribution strictly reflects the incremental value of individual contribution, gives the edge node the ability to dynamically adjust the upper limit of the schedulable computing power in the multi-homing environment through the periodic sealed collaborative participation response framework, and realizes the adaptive scheduling of the dynamic change of the heterogeneous computing power supply.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and edge computing collaborative scheduling technology, specifically to a cloud-edge collaborative elastic scheduling method and system for heterogeneous computing resources. Background Technology

[0002] With the application of the Internet of Things and artificial intelligence, the cloud-edge collaborative architecture, which combines cloud computing and edge computing, has become a key paradigm to support low-latency, high-bandwidth computing tasks. On the edge side, a large number of heterogeneous computing resources are distributed across edge nodes in different geographical locations, and these edge nodes usually belong to different owners. In order to improve the overall utilization of computing resources, it is necessary to design an effective collaborative scheduling mechanism to encourage edge nodes of different owners to actively share idle heterogeneous computing power and work with the cloud scheduling center to complete elastic computing tasks.

[0003] Existing collaborative scheduling technologies for multi-owner edge nodes typically employ centralized resource orchestration or simple allocation methods based on bidding mechanisms. However, existing bidding allocation mechanisms rely solely on bid prices as the sole decision-making criterion, making it difficult to adapt to the needs of multi-dimensional optimal combinations of heterogeneous computing power, and the scheduling results cannot guarantee optimal global collaborative efficiency. Furthermore, existing mechanisms use coarse-grained methods such as allocation based on bid ratios or equal allocation in the revenue distribution stage, failing to consider the differences in substantial contributions of different edge nodes in the collaborative combination. This results in high-contribution nodes not receiving rewards commensurate with their contributions, undermining the nodes' motivation to continue participating in collaboration. Moreover, when dealing with multi-ownership scenarios where edge nodes simultaneously belong to multiple scheduling domains, existing technologies lack constraints to ensure that nodes honestly declare their available computing power limits across multiple scheduling domains. The scheduling center cannot obtain the true available computing power information of nodes, making it difficult to guarantee the feasibility and stability of scheduling decisions.

[0004] In summary, existing cloud-edge collaborative scheduling methods struggle to simultaneously maximize global collaborative efficiency and ensure fair returns for participating nodes in scenarios with multiple owners and multiple affiliations of heterogeneous computing power. Therefore, a cloud-edge collaborative elastic scheduling method and system is needed that can comprehensively optimize multi-dimensional heterogeneous computing power combinations, distribute benefits fairly based on the substantial marginal contributions of nodes, and adapt to dynamic scheduling environments with multiple affiliations. Summary of the Invention

[0005] To address the technical problem of unstable edge node collaborative alliances and inability to achieve globally optimal scheduling results in multi-ownership, multi-homogeneous computing power scenarios due to the lack of a collaborative mechanism that balances global efficiency and individual fairness in existing cloud-edge collaborative scheduling technologies, this invention designs a cloud-edge collaborative elastic scheduling method and system for heterogeneous computing power resources. This invention models heterogeneous computing power resources as a multi-dimensional computing power resource package vector, introduces task completion benefit values ​​and the expected collaborative benefit base of edge nodes, and constructs a winning collaborative combination optimization mechanism aimed at maximizing the net benefit of collaborative scheduling. This ensures that the scheduled edge node combinations achieve optimal system net benefit while meeting multi-dimensional resource requirements. Simultaneously, this invention uses the substantial marginal contribution of edge nodes in the collaborative combination as the sole criterion to calculate the actual collaborative benefit value of each participating edge node, ensuring that the benefit distribution strictly reflects the incremental value of individual contributions. Through a periodically sealed collaborative participation response framework, it empowers edge nodes to dynamically adjust the upper limit of schedulable computing power in multi-homogeneous environments, achieving adaptive scheduling to dynamic changes in heterogeneous computing power supply.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a cloud-edge collaborative elastic scheduling method for heterogeneous computing resources, the specific steps of which are as follows: S100, the cloud scheduling center parses the arriving task set into a multi-dimensional resource demand vector, and broadcasts the multi-dimensional resource demand vector and the corresponding task completion benefit value to multiple registered edge nodes. Each edge node belongs to a different owner, and a single edge node is allowed to register to multiple scheduling domains at the same time. S200. Each edge node generates multiple computing power resource packages based on the real-time heterogeneous computing power availability of its own node, and sets a corresponding expected collaborative benefit base for each computing power resource package. The computing power resource package and the expected collaborative benefit base are encapsulated into a sealed collaborative participation response information and sent to the cloud scheduling center. S300. The cloud scheduling center collects the sealed collaborative participation response information submitted by all edge nodes within the response deadline. With the goal of maximizing the net benefit of collaborative scheduling, it determines a winning collaborative combination among all edge node combinations that can collaboratively meet the multi-dimensional resource demand vector. The net benefit of collaborative scheduling is the difference between the task completion benefit value and the sum of the expected collaborative benefit bases of all edge nodes in the winning collaborative combination. S400. The cloud scheduling center uses the Shapley value allocation algorithm to calculate the actual collaborative benefit value that each participating edge node in the winning collaborative combination should receive, and allocates the benefit to each participating edge node according to the actual collaborative benefit value. S500, the cloud scheduling center schedules the corresponding subtasks in the task set to each participating edge node in the winning collaborative combination for execution, and monitors the task execution status until completion.

[0007] Furthermore, in S200, the sealing and collaborative encapsulation process for the response information is as follows: Edge nodes collect the real-time availability of various heterogeneous computing resources on the local surface. These heterogeneous computing resources include at least one of the following: number of CPU cores, graphics processor memory, neural network processor computing power, and number of field-programmable gate array (FPGA) units. The edge nodes combine and generate multiple computing power resource packages based on the real-time availability. Each computing power resource package is described in the form of a multi-dimensional vector, which describes the type and quantity of various heterogeneous computing power resources contained in the package. Based on the local computing power resource occupation cost and expected retention utility, an expected collaboration benefit base is preset for each computing power resource package. The expected collaboration benefit base reflects the minimum actual collaboration benefit value required for the edge node to participate in this collaborative scheduling by transferring the corresponding computing power resource package. The multidimensional vector of the computing power resource package and the expected collaborative benefit base corresponding to the computing power resource package are encapsulated together to form the sealed collaborative participation response information, which is then sent to the cloud scheduling center through a secure channel.

[0008] Furthermore, in the multi-dimensional vector of the computing power resource package, each dimension corresponds to a type of heterogeneous computing power resource. The value of the dimension represents the maximum schedulable amount of the corresponding type of heterogeneous computing power resource that the edge node can transfer within the current scheduling cycle. The edge node generates multiple mutually exclusive computing power resource package options for the same scheduling cycle, and each computing power resource package option corresponds to a different combination configuration of heterogeneous computing power resources.

[0009] Furthermore, in S300, the process of determining the winning cooperative combination is as follows: The cloud scheduling center traverses all computing power resource packages submitted by edge nodes. For each edge node combination, it determines whether all computing power resource packages within the edge node combination meet the aggregated resource coverage condition. The aggregated resource coverage condition is: whether the aggregated resource vector of all computing power resource packages is not less than the corresponding dimension value of the multidimensional resource demand vector in each dimension. For each combination of edge nodes that satisfies the aggregated resource coverage condition, according to Calculate the net collaborative scheduling benefit value corresponding to the feasible edge node combination, where, Represents feasible combinations of edge nodes The corresponding net benefit value of coordinated scheduling, This represents the task completion benefit value broadcast by the cloud-based dispatch center. Represents the combination of edge nodes The Middle The expected collaboration revenue base submitted by each edge node for its respective computing power resource package. Represents a feasible combination of edge nodes. The expected collaboration revenue base of all edge nodes is summed. Compare the net benefits of collaborative scheduling for all feasible combinations of edge nodes. Select to make The feasible combination of edge nodes that achieves the maximum value is taken as the winning collaborative combination. .

[0010] Furthermore, when multiple different feasible combinations of edge nodes exist, all of which make If the same maximum value is achieved, priority is given to those that make the maximum value. Among multiple feasible edge node combinations that achieve the maximum value, the feasible edge node combination with the fewest edge nodes is selected as the winning collaborative combination. This reduces the scheduling overhead of cross-node task coordination and data transmission.

[0011] Furthermore, the specific steps of S400 are as follows: S401, Winning Synergy The total number of edge nodes included is The cloud-based dispatch center targets the winning collaborative combinations. Any target participating in the edge node Enumerate the winning combinations The target participating edge node is not included. All subcombinations, each subcombination is denoted as , yes subset and Not belonging to ; S402, For each sub-combination Calculate the sub-combination The resulting sub-combination collaborative scheduling net benefit value And calculate the participation of the target in the edge node. Add the sub-combination The net benefit of the new combination's coordinated scheduling that can be generated by the newly formed combination Subtracting the two yields the target participating edge node. Relative to subcombination The marginal synergistic benefit contribution value, expressed as... ; S403, according to Weighted summation calculation target participates in edge nodes The actual collaborative benefits that should be received ,in, Indicates the target participating edge node The actual collaborative benefits that should be received. Indicates the winning combination The target participating edge nodes are not included. Any sub-combination, Subcombinations The number of edge nodes included Indicates the winning combination The total number of edge nodes included, where ! represents factorial operation. This indicates that the target will participate in the edge node. Add sub-combinations The net benefit of the new combination's coordinated scheduling that can be generated by the newly formed combination Subcombinations The net revenue generated by the sub-combination collaborative scheduling, summation sign It means that for all satisfying and sub-combinations Perform summation; S404. Repeat steps S401-S403 until a winning combination is found. The actual collaborative benefit value of each participating edge node has been calculated.

[0012] Furthermore, S403 calculates the actual collaborative benefit value. Subsequently, it was verified that each target participated in the edge nodes. Actual collaborative benefits Are all of them no smaller than the target participating edge node? Submitted expected collaboration benefit base When any target participates in the edge node Appear If the current scheduling does not meet the conditions for rational individual participation, the cloud scheduling center will abandon the winning collaborative combination and select a backup edge node combination with the second-best collaborative scheduling net benefit value and satisfying that the actual collaborative benefit value of all participating edge nodes is not less than their respective expected collaborative benefit base as the scheduling object.

[0013] Furthermore, when the sub-combination is... Add target participating edge nodes Afterwards, the new combination If the aggregated resource vector still does not meet the corresponding dimensional requirements of the multidimensional resource demand vector in any dimension, then the new combination... Ineffective collaborative combinations deemed incapable of completing the task; new combinations The corresponding new combined collaborative scheduling net benefit value A value of 0 corresponds to the target participating edge node. Relative to the sub-combination The marginal synergistic benefit contribution value is calculated as follows: .

[0014] Secondly, a cloud-edge collaborative elastic scheduling system for heterogeneous computing resources, comprising: a cloud scheduling center and multiple edge nodes; The cloud-based dispatch center includes: The demand broadcasting module is used to parse the arriving task set into a multi-dimensional resource demand vector and broadcast the multi-dimensional resource demand vector and the corresponding task completion benefit value to multiple registered edge nodes. The response collection module is used to receive the sealed collaborative participation response information submitted by each edge node. The sealed collaborative participation response information includes the computing power resource package generated by each edge node and the corresponding expected collaborative benefit base. The combinatorial optimization module is used to gather all sealed collaborative response information, with the goal of maximizing the net benefit of collaborative scheduling, and to determine a winning collaborative combination among all edge node combinations that can collaboratively satisfy the multidimensional resource demand vector. The Shapley value allocation module is used to calculate the actual collaborative revenue value that each participating edge node in the winning collaborative combination should receive using the Shapley value allocation algorithm, and to allocate revenue according to the actual collaborative revenue value. The scheduling and execution module is used to schedule the corresponding subtasks in the task set to each participating edge node in the winning collaborative combination for execution, and to monitor the task execution status. The edge nodes include: The resource monitoring module is used to periodically collect the real-time availability of various heterogeneous computing resources on the local machine and generate a description of available computing resource packages. The collaborative participation generation module is used to generate at least one computing power resource package according to the description of the available computing power resource package, set a corresponding expected collaborative benefit base for each computing power resource package, and encapsulate the computing power resource package and the expected collaborative benefit base into sealed collaborative participation response information. The task execution module is used to receive sub-tasks issued by the cloud scheduling center and call the corresponding heterogeneous computing resources to complete the execution.

[0015] Compared with existing technologies, this cloud-edge collaborative elastic scheduling method for heterogeneous computing resources has the following advantages: I. This invention constructs a winning collaborative combination determination mechanism with the goal of maximizing the net benefit of collaborative scheduling by using the task completion benefit value and the expected collaborative benefit base. It uniformly describes the heterogeneous computing power capabilities of edge nodes in the form of multi-dimensional vectors. After collecting the sealed collaborative participation response information submitted by each edge node, the combination optimization module of the cloud scheduling center traverses all feasible edge node combinations, verifies whether the aggregated resource vector of each combination meets the multi-dimensional resource requirement vector of the task in each heterogeneous computing power dimension, and uses the collaborative scheduling net benefit value as the selection criterion. This ensures that in any scheduling cycle, the selected winning collaborative combination is the resource orchestration scheme with the largest net benefit at the system level, realizing the optimal allocation of heterogeneous computing power resources and avoiding suboptimal scheduling results that ignore the collaborative efficiency of combination due to the one-sided pursuit of low collaboration cost.

[0016] Second, this invention calculates the actual collaborative benefit value of each participating edge node based on the substantial marginal contribution made by the edge nodes in the collaborative scheduling. By enumerating the order of all sub-combinations that win the collaborative combination, the marginal collaborative benefit contribution value brought by each target participating edge node after joining different sub-combinations is calculated one by one, and the actual collaborative benefit value of the node is obtained by weighted summation according to the mathematical weight of the combination. This allocation method makes the actual collaborative benefit value obtained by each edge node equal to the expected incremental net benefit after joining the collaborative alliance, which satisfies the individual rational participation condition and solves the technical problem of unstable collaborative alliance caused by improper incentives in multi-homing scenarios.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart illustrating the steps of a cloud-edge collaborative elastic scheduling method for heterogeneous computing resources in an embodiment of the present invention. Figure 2 This is a flowchart of the revenue distribution process based on the Shapley value in embodiment S400 of the present invention; Figure 3 This is a schematic diagram of the composition of a cloud-edge collaborative elastic scheduling system for heterogeneous computing resources in an embodiment of the present invention. Detailed Implementation

[0020] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a cloud-edge collaborative elastic scheduling method for heterogeneous computing resources. This method is applicable to a cloud-edge collaborative scheduling system composed of a cloud scheduling center and multiple edge nodes. The cloud scheduling center is deployed on a central cloud platform or regional cloud platform, responsible for receiving sets of computing tasks from upper-layer applications and performing parsing, scheduling, and monitoring management of these tasks. Edge nodes are deployed at the network edge, each possessing one or more heterogeneous computing resources, including but not limited to CPU cores, GPU memory, neural network processor computing power, and field-programmable gate arrays (FPGAs). Each edge node belongs to a different owner, and a single edge node can register to multiple scheduling domains simultaneously. That is, the same edge node can participate in the resource pools of multiple cloud scheduling centers simultaneously, sharing its heterogeneous computing resources across different scheduling domains. The entire scheduling process is executed periodically according to a scheduling cycle, with the task set arriving in each scheduling cycle constituting the scheduling object for that cycle.

[0022] In this embodiment, as Figure 1 As shown, the specific steps of a cloud-edge collaborative elastic scheduling method for heterogeneous computing resources are as follows: S100, the cloud scheduling center parses the arriving task set into a multi-dimensional resource demand vector, and broadcasts the multi-dimensional resource demand vector and the corresponding task completion benefit value to multiple registered edge nodes; S200. Each edge node generates multiple computing power resource packages based on the real-time heterogeneous computing power availability of its own node, and sets a corresponding expected collaborative benefit base for each computing power resource package. The computing power resource package and the expected collaborative benefit base are encapsulated into a sealed collaborative participation response information and sent to the cloud scheduling center. S300. The cloud scheduling center collects the sealed collaborative participation response information submitted by all edge nodes within the response deadline. With the goal of maximizing the net benefit of collaborative scheduling, it determines a winning collaborative combination among all edge node combinations that can collaboratively satisfy the multidimensional resource demand vector. S400. The cloud scheduling center uses the Shapley value allocation algorithm to calculate the actual collaborative benefit value that each participating edge node in the winning collaborative combination should receive, and allocates the benefit to each participating edge node according to the actual collaborative benefit value. S500, the cloud scheduling center schedules the corresponding subtasks in the task set to each participating edge node in the winning collaborative combination for execution, and monitors the task execution status until completion.

[0023] In step S100 above, when the current scheduling cycle starts, the cloud scheduling center receives a task set submitted by the upper-layer application. The task set contains one or more computing tasks to be scheduled, each with its own resource requirement and priority attributes. The cloud scheduling center performs unified parsing on the task set, converting the overall resource requirement of the task set into a multi-dimensional resource requirement vector. In this embodiment, the number of dimensions of the multi-dimensional resource requirement vector corresponds to the total number of predefined heterogeneous computing resource types in the system. The system predefined heterogeneous computing resource types include four types: the number of CPU cores, the amount of GPU memory, the computing power of neural network processors, and the number of field-programmable gate array (FPGA) logic units. Therefore, the multi-dimensional resource requirement vector is a four-dimensional vector. ,in, This indicates the total number of CPU cores required for the task set. This represents the total amount of graphics processing unit (GPU) memory required for the task set. This represents the total computing power required by the neural network processors for the task set. This indicates the total number of field-programmable gate array (FPGA) logic cells required for the task set.

[0024] When a task in a task set does not require a certain type of heterogeneous computing resource, the requirement value of the task set in the corresponding dimension is set to 0. After parsing the multi-dimensional resource requirement vector, the cloud scheduling center obtains the task completion benefit value corresponding to the task set. The task completion benefit value is the total revenue obtained by the system after completing the task set. The unit of this total revenue is consistent with the unit of the expected collaborative revenue base of each edge node, and both are measured using a unified utility unit. The task completion benefit value is provided by the upper-layer application when submitting the task set, or it is generated by the cloud scheduling center after a comprehensive evaluation based on the computational load, urgency, and service quality requirements of the task set. In this embodiment, the task completion benefit value is denoted as... , For positive real numbers, the cloud scheduling center will use the multidimensional resource demand vector. and the corresponding task completion benefit value It is encapsulated as a broadcast message and broadcast to multiple registered edge nodes through a pre-established secure communication link.

[0025] The registered edge nodes are those that completed the registration process before the current scheduling cycle. The registration process includes each edge node reporting its node identifier, network address, owner information, and a list of supported heterogeneous computing resource types to the cloud scheduling center. The cloud scheduling center maintains an edge node registry, which records the registration information of all registered edge nodes. The cloud scheduling center broadcasts the message to all edge nodes in the registry. The secure communication link is protected using symmetric or asymmetric encryption mechanisms to ensure the confidentiality and integrity of the broadcast message during transmission.

[0026] In this embodiment, when there are multiple cloud scheduling centers running simultaneously in the system, the same edge node can complete an independent registration process with different cloud scheduling centers and become a registered edge node of multiple scheduling domains.

[0027] In step S200 above, after receiving the multi-dimensional resource demand vector D and task completion benefit value B broadcast by the cloud scheduling center, each edge node performs a sealing collaborative participation response information encapsulation process, which specifically includes: Edge nodes collect real-time availability data of various heterogeneous computing resources locally. A resource monitoring agent is deployed on each edge node, collecting the current usage status of these resources at a fixed sampling period. In this embodiment, the resource monitoring agent obtains the number of idle CPU cores and the total number of cores by reading the resource management interface provided by the operating system, and calculates the real-time availability of CPU cores; it obtains the free capacity of GPU memory by reading the memory usage information provided by the GPU driver, which serves as the real-time availability of GPU memory; it obtains the currently available computing power value of the neural network processor by reading the computing power utilization information provided by the neural network processor driver; and it obtains the number of currently available logic units by reading the logic unit usage status of the field-programmable gate array (FPGA).

[0028] Edge nodes aggregate the real-time availability of various heterogeneous computing resources collected, generating a real-time available computing power vector for that edge node within the current scheduling period. Let the real-time available computing power vector of edge node i within the current scheduling period be denoted as . , The dimensions and the multidimensional resource demand vector The same dimensions Represented as: ,in, This represents the number of CPU cores that edge node i can relinquish during the current scheduling cycle. This represents the amount of graphics processing unit (GPU) memory that edge node i can release during the current scheduling cycle. This represents the neural network processor computing power that edge node i can relinquish within the current scheduling cycle. This represents the number of field-programmable gate array (FPGA) units that edge node i can relinquish during the current scheduling period.

[0029] Edge nodes based on the real-time available computing power vector Multiple computing resource packages are generated by combining these packages. Each computing resource package is described in the form of a multi-dimensional vector, representing the types and quantities of various heterogeneous computing resources it contains. Edge nodes generate multiple mutually exclusive computing resource package options within the same scheduling cycle, with each option corresponding to a different combination of heterogeneous computing resources. Let the computing resource package vector declared by edge node i in the m-th computing resource package option be... , Represented as: ,in, This represents the number of central processing unit cores committed to be transferred in the computing power resource package m. This represents the amount of graphics processing unit (GPU) memory committed to be transferred in computing resource package m. This represents the neural network processor computing power committed to be transferred in computing power resource package m. This represents the number of field-programmable gate array (FPGA) units committed to be transferred in the computing resource package m. For the computing resource package vector... Each dimension of the vector has a value no greater than the real-time available computing power vector. The values ​​in the corresponding dimension satisfy: The multiple computing resource package options generated by the edge node satisfy mutual exclusion constraints. That is, the edge node can only select one computing resource package option to participate in the collaborative scheduling in the current scheduling cycle. The resource combination configurations corresponding to different computing resource package options are different from each other, and there is no case where one option is better than the other in all dimensions between any two computing resource package options, so as to ensure that each option has actual differentiated significance.

[0030] Edge nodes preset an expected collaboration benefit base for each computing resource package based on the local computing resource occupation cost and expected retention utility. When calculating the expected collaboration benefit base, edge nodes calculate the opportunity cost of relinquishing various heterogeneous computing resources within the computing resource package. This opportunity cost includes the loss incurred by the edge node due to being unable to execute local tasks because of relinquishing computing resources; this loss is estimated by the edge node based on the historical average resource occupation and average revenue of local tasks. Edge nodes calculate the energy consumption cost of relinquishing computing resources. The sum of the opportunity cost and the energy consumption cost is used as the base cost, and an expected retention utility increment is added to this base cost to obtain the expected collaboration benefit base. The expected retention utility increment is the additional benefit required for the edge node to participate in collaborative scheduling. Edge nodes independently set the expected collaboration benefit base for each computing resource package. Let the expected collaboration benefit base set by edge node i for computing resource package m be... , It is a positive real number.

[0031] The edge node encapsulates the multi-dimensional vector of the computing resource package together with the expected collaborative benefit base corresponding to the computing resource package to form the sealed collaborative participation response information. The data structure of the sealed collaborative participation response information includes an edge node identifier field, a computing resource package vector field, and an expected collaborative benefit base field.

[0032] When an edge node generates multiple computing resource package options, it encapsulates each option into an independent sealed collaborative participation response message. Each sealed collaborative participation response message contains one and only one computing resource package option and its corresponding expected collaborative benefit base. The edge node sends all sealed collaborative participation response messages to the cloud scheduling center through a pre-established secure channel.

[0033] In step S300 above, after sending a broadcast message, the cloud scheduling center starts a response timer. The duration of the response timer is equal to the system's preset response deadline. Within the response deadline, the cloud scheduling center collects the sealed collaborative participation response information submitted by all edge nodes. When the response timer expires, the cloud scheduling center no longer receives new sealed collaborative participation response information and performs validity verification on all received sealed collaborative participation response information. The validity verification includes verifying whether the format of the sealed collaborative participation response information is complete, verifying whether the values ​​of each dimension of the computing power resource package vector are non-negative real numbers, and verifying whether the expected collaborative benefit base is a positive real number.

[0034] The cloud scheduling center aims to maximize the net benefit of collaborative scheduling. Among all edge node combinations capable of collaboratively satisfying the multidimensional resource demand vector, it determines a winning collaborative combination. The cloud scheduling center iterates through all computing power resource packages submitted by all edge nodes. For each edge node combination, the cloud scheduling center determines whether all computing power resource packages within that combination meet the aggregated resource coverage condition. The aggregated resource coverage condition is: the aggregated resource vector of all computing power resource packages in the edge node combination is not less than the corresponding dimension value of the multidimensional resource demand vector D in each dimension. Let a certain edge node combination to be determined be denoted as S, containing K edge nodes, where K is a positive integer. Let the computing power resource package vector submitted by the i-th edge node in S be denoted as... , The dimension of is the same as the dimension of the multidimensional resource demand vector D. Represented as: Then the aggregated resource vector of S Each dimension is calculated as follows: , , , The aggregated resource coverage condition is expressed as follows: , , , The edge node combination is true if and only if all four inequalities are true. The feasible combination of edge nodes is determined to meet the aggregated resource coverage conditions.

[0035] For each feasible edge node combination that meets the aggregated resource coverage conditions, the cloud scheduling center calculates the net collaborative scheduling benefit value corresponding to that feasible edge node combination. The formula for calculating the net collaborative scheduling benefit value is as follows: ,in, Represents feasible combinations of edge nodes The corresponding net benefit value of coordinated scheduling, This represents the benefit value of completing the task. Represents feasible combinations of edge nodes The Middle The expected collaboration revenue base submitted by each edge node for its respective computing power resource package. Represents a feasible combination of edge nodes. The expected collaboration revenue base of all edge nodes is summed, and when feasible edge node combinations are found... When it is an empty set, , , The value of is a real number, which can be positive or negative. When the value is positive, it indicates that coordinated scheduling can generate positive net benefits; when... When the value is negative, it indicates that the benefits of coordinated scheduling are insufficient to cover the collaboration costs of participating edge nodes.

[0036] The cloud-based scheduling center compares the net benefits of collaborative scheduling for all feasible combinations of edge nodes. Select to make The feasible combination of edge nodes that achieves the maximum value is taken as the winning collaborative combination. The winning collaborative combination The method of determination is expressed as follows: ,in, This represents the set of all feasible edge node combinations that satisfy the aggregated resource coverage condition. This represents the value of the independent variable that maximizes the objective function. This is true when there are multiple different combinations of feasible edge nodes that maximize the objective function. When the same maximum value is obtained, the cloud-based scheduling center prioritizes using the one that... Among multiple feasible edge node combinations that achieve the maximum value, the feasible edge node combination with the fewest edge nodes is selected as the winning collaborative combination. .

[0037] In step S400 above, such as Figure 2 As shown, the cloud-based scheduling center determines the winning collaborative combination. Then, regarding the winning collaborative combination The winning collaborative combination is calculated using the Shapley value allocation algorithm. The actual collaborative benefit value that each participating edge node should receive is determined, and the benefits are allocated to each participating edge node according to the actual collaborative benefit value. The Shapley value allocation algorithm uses the substantial marginal contribution of each edge node in the collaborative combination as the allocation basis to ensure the fairness and incentive compatibility of the benefit distribution. Let the winning collaborative combination be... The total number of edge nodes included is The cloud-based scheduling center targets the winning collaborative combination. Any target participating in the edge node Enumerate the winning collaborative combinations The target participating edge node is not included. All subcombinations, each subcombination is denoted as , satisfy and The enumeration process iterates through... All possible values, including The case of an empty set and for Except The case of the set consisting of all other edge nodes.

[0038] For each sub-combination The cloud-based scheduling center calculates the sub-combinations. The resulting sub-combination collaborative scheduling net benefit value : ,when When it is an empty set, The cloud-based scheduling center calculates and participates the target in the edge nodes. Add the sub-combination The new combination formed later The resulting new combined collaborative scheduling net benefit value The cloud-based dispatch center will and Subtracting them yields the target participating edge nodes. Relative to subcombination The marginal synergistic benefit contribution value, expressed as: , For target participating edge nodes Add sub-combinations Subsequently, the net increase in coordinated scheduling revenue brought about by the coordinated combination, It can take positive, negative, or zero values. When When the value is positive, it indicates that the target participates in the edge node. The addition of this feature improved the net benefits of coordinated scheduling; when When the value is negative, it indicates that the target participates in the edge node. The addition of [something] reduced the net benefit of coordinated scheduling.

[0039] In this embodiment, when the sub-combination is... Add target participating edge nodes Afterwards, if a new combination The aggregated resource vector still does not satisfy the multidimensional resource demand vector in any dimension. If the corresponding dimension requirement is met, then the new combination This is considered an invalid collaborative combination that cannot complete the task. In this case, a new combination... The corresponding new combined collaborative scheduling net benefit value A value of 0 corresponds to the target participating edge node. Relative to the sub-combination The marginal synergistic benefit contribution value is calculated as follows: ,when It is an empty set and If the aggregated resource coverage condition is still not met, , This ensures that during the calculation of Shapley values, any combination that cannot independently or collaboratively meet the task's resource requirements is assigned zero revenue, thereby accurately reflecting the negative marginal contribution of the target participating edge nodes in ineffective combinations and avoiding assigning unreasonable positive revenue to ineffective combinations.

[0040] The cloud-based scheduling center calculates the target participating edge nodes using a weighted summation based on the Shapley value formula. The actual benefits of collaboration that you deserve ,in, Indicates the target participating edge node The actual collaborative benefits that should be received. The winning collaborative combination is indicated. The target participating edge nodes are not included. Any sub-combination, Subcombinations The number of edge nodes included The winning collaborative combination is indicated. The total number of edge nodes included. To represent factorial operation, This indicates that the target will participate in the edge node. Add sub-combinations The net benefit of the new combination's coordinated scheduling that can be generated by the newly formed combination Subcombinations The net revenue generated by the sub-combination collaborative scheduling, summation sign It means that for all satisfying and sub-combinations Summation is performed. The weighting factors... Shapley values ​​are weighting coefficients in combinatorics that satisfy the following condition for all subcombinations. The sum of their weights equals 1, that is: The actual collaborative benefit value The unit and the benefit value of the task completion and the base of expected collaborative benefits The units are consistent, all being unified units of utility. This continues until the winning cooperative combination is determined. The actual collaborative benefit value of each participating edge node has been calculated. In this embodiment, the cloud scheduling center calculates the actual collaborative benefit value of each participating edge node. Then, verify that each target participates in the edge node. Actual collaborative benefits Are all of them no less than the target participating edge nodes? Submitted expected collaboration benefit base When for all All When the scheduling is deemed to meet the conditions for rational individual participation, the cloud-based scheduling center determines the actual collaborative benefit value. Revenue is allocated to each participating edge node. This occurs when any target participating edge node exists. Appear If the current scheduling does not meet the conditions for rational individual participation, the cloud-based scheduling center will abandon the winning collaborative combination. The cloud scheduling center selects a backup edge node combination as the scheduling target, provided that the net collaborative scheduling benefit value of all participating edge nodes is not less than their respective expected collaborative benefit base value. The backup edge node combination is determined as follows: among all feasible edge node combinations, the cloud scheduling center sequentially verifies the individual rational participation condition in descending order of the net collaborative scheduling benefit value. The first feasible edge node combination to pass this verification is determined as the backup edge node combination, and the benefit allocation calculation in step S400 is re-executed for this backup edge node combination. If no feasible edge node combination satisfies the individual rational participation condition, the cloud scheduling center determines that the current scheduling cycle cannot complete the scheduling of the task set, returns a scheduling failure response to the upper-layer application, and places the task set in the waiting queue for the next scheduling cycle to await rescheduling.

[0041] In step S500 above, after completing the combination selection and revenue calculation, the cloud scheduling center issues and executes the tasks. Based on the resource capabilities of each edge node in the winning collaborative combination and the characteristics of the subtasks in the current task set, a task mapping and distribution strategy is formulated. This strategy schedules different subtasks or data slices in the task set to the corresponding participating edge nodes. Upon receiving the subtasks from the cloud scheduling center, the task execution modules of each participating edge node invoke their corresponding local heterogeneous computing resources to begin executing the computation tasks. During execution, the scheduling and execution module of the cloud scheduling center continuously monitors the task execution status of each edge node, including but not limited to task progress, resource utilization, and network connectivity.

[0042] This invention also provides a cloud-edge collaborative elastic scheduling system for heterogeneous computing resources, such as... Figure 3 As shown, the system architecture includes a cloud-based scheduling center and multiple edge nodes.

[0043] The cloud-based scheduling center is a server cluster or high-performance computing unit deployed in the cloud, comprising: Demand Broadcast Module: This module is responsible for receiving and parsing the arriving task set, transforming it into a multi-dimensional resource demand vector, and broadcasting this vector along with the corresponding task completion benefit value to all registered edge nodes. The input to this module is the raw task flow, and the output is a structured broadcast data packet.

[0044] Response Collection Module: This module is responsible for receiving and storing the sealed collaborative participation response information submitted by each edge node through a secure channel within a preset time window.

[0045] Combinatorial Optimization Module: This module is the core of the system's decision-making process, encapsulating the algorithm for determining the winning collaborative combination. This module obtains all response information from the response collection module, searches and compares all feasible edge node combinations with the goal of maximizing the net benefit of collaborative scheduling, and finally outputs a winning collaborative combination and its corresponding maximum net benefit value.

[0046] Shapley Value Allocation Module: This module is activated after the combinatorial optimization module determines the winning combination. Internally, it encapsulates the Shapley value algorithm, which enumerates all sub-combinations of the winning combination, calculates the marginal contribution of each participating node, and ultimately calculates the actual collaborative revenue value that each node should receive, generating a revenue allocation instruction.

[0047] Scheduling and Execution Module: This module is responsible for dividing the task set into subtasks and generating and issuing specific scheduling instructions based on the resource characteristics of each node in the winning collaborative combination. Simultaneously, this module continuously monitors the task execution status until the task is completed.

[0048] The edge nodes are devices or clusters of devices with heterogeneous computing capabilities deployed at the network edge, and each edge node belongs independently to a different owner. Each edge node includes the following functional modules: Resource monitoring module: This module runs in the background of the edge node. By calling the hardware monitoring interface at the underlying operating system, it periodically collects the real-time availability of various heterogeneous computing resources on the local machine and generates a vector describing the currently available resources.

[0049] Collaborative Participation Generation Module: This module obtains available resource information from the resource monitoring module and generates one or more mutually exclusive computing power resource packages according to a preset participation strategy. Simultaneously, based on local resource occupancy costs and expected retention utility, it sets an expected collaborative benefit base for each generated computing power resource package, encapsulates this information into sealed collaborative participation response information, and sends it to the cloud scheduling center.

[0050] Task execution module: Receives subtasks issued by the cloud scheduling center, calls corresponding heterogeneous computing resources according to the task type, creates an isolated execution environment, and starts the task execution process. After the task is completed, it returns the execution result and resource usage report to the cloud scheduling center.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A cloud-edge collaborative elastic scheduling method for heterogeneous computing resources, characterized in that, The specific steps of this method are as follows: S100, the cloud scheduling center parses the arriving task set into a multi-dimensional resource demand vector, and broadcasts the multi-dimensional resource demand vector and the corresponding task completion benefit value to multiple registered edge nodes; S200. Each edge node generates multiple computing power resource packages based on the real-time heterogeneous computing power availability of its own node, and sets a corresponding expected collaborative benefit base for each computing power resource package. The computing power resource package and the expected collaborative benefit base are encapsulated into a sealed collaborative participation response information and sent to the cloud scheduling center. S300. The cloud scheduling center collects the sealed collaborative participation response information submitted by all edge nodes within the response deadline. With the goal of maximizing the net benefit of collaborative scheduling, it determines a winning collaborative combination among all edge node combinations that can collaboratively satisfy the multidimensional resource demand vector. S400. The cloud scheduling center uses the Shapley value allocation algorithm to calculate the actual collaborative benefit value that each participating edge node in the winning collaborative combination should receive, and allocates the benefit to each participating edge node according to the actual collaborative benefit value. S500, the cloud scheduling center schedules the corresponding subtasks in the task set to each participating edge node in the winning collaborative combination for execution, and monitors the task execution status until completion.

2. The cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 1, characterized in that, In step S200, the sealing and collaborative encapsulation process for the response information is as follows: Edge nodes collect the real-time availability of various heterogeneous computing resources on the local surface. These heterogeneous computing resources include at least one of the following: number of CPU cores, graphics processor memory, neural network processor computing power, and number of field-programmable gate array (FPGA) units. The edge nodes combine and generate multiple computing power resource packages based on the real-time availability. Each computing power resource package is described in the form of a multi-dimensional vector, which describes the type and quantity of various heterogeneous computing power resources contained in the package. Based on the local computing power resource occupation cost and expected retention utility, a base of expected collaborative revenue is preset for each computing power resource package; The multidimensional vector of the computing power resource package and the expected collaborative benefit base corresponding to the computing power resource package are encapsulated together to form the sealed collaborative participation response information, which is then sent to the cloud scheduling center through a secure channel.

3. The cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 2, characterized in that, In the multi-dimensional vector of the computing power resource package, each dimension corresponds to a type of heterogeneous computing power resource. The value of the dimension represents the maximum schedulable amount of the corresponding type of heterogeneous computing power resource that the edge node can transfer in the current scheduling cycle. The edge node generates multiple mutually exclusive computing power resource package options for the same scheduling cycle, and each computing power resource package option corresponds to a different combination configuration of heterogeneous computing power resources.

4. The cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 1, characterized in that, In S300, the process of determining the winning cooperative combination is as follows: The cloud scheduling center traverses all computing power resource packages submitted by edge nodes. For each edge node combination, it determines whether all computing power resource packages within the edge node combination meet the aggregated resource coverage condition. The aggregated resource coverage condition is: whether the aggregated resource vector of all computing power resource packages is not less than the corresponding dimension value of the multidimensional resource demand vector in each dimension. For each combination of edge nodes that satisfies the aggregated resource coverage condition, according to Calculate the net collaborative scheduling benefit value corresponding to the feasible edge node combination, where, Represents feasible combinations of edge nodes The corresponding net benefit value of coordinated scheduling, This represents the task completion benefit value broadcast by the cloud-based dispatch center. Represents the combination of edge nodes The Middle The expected collaboration revenue base submitted by each edge node for its respective computing power resource package. Represents a feasible combination of edge nodes. Sum the expected collaboration revenue base of all edge nodes in the middle; Compare the net benefits of collaborative scheduling for all feasible combinations of edge nodes. Select to make The feasible combination of edge nodes that achieves the maximum value is taken as the winning collaborative combination. .

5. A cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 4, characterized in that, When there are multiple different feasible combinations of edge nodes, all of which make If the same maximum value is achieved, priority is given to those that make the maximum value. Among multiple feasible edge node combinations that achieve the maximum value, the feasible edge node combination with the fewest edge nodes is selected as the winning collaborative combination. .

6. The cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 1, characterized in that, The specific steps of S400 are as follows: S401, Winning Synergy Combination The total number of edge nodes included is The cloud-based dispatch center targets the winning collaborative combinations. Any target participating in the edge node Enumerate the winning combinations The target participating edge node is not included. All subcombinations, each subcombination is denoted as , yes subset and Not belonging to ; S402, For each sub-combination Calculate the sub-combination The net benefit value of sub-combination collaborative scheduling that can be generated And calculate the participation of the target in the edge node. Add the sub-combination The net benefit of the new combination's coordinated scheduling that can be generated by the newly formed combination Obtain the target participating edge nodes Relative to subcombination The marginal synergistic benefit contribution value, expressed as... ; S403, according to Weighted summation calculation target participates in edge nodes The actual collaborative benefits that should be received ,in, Indicates the target participating edge node The actual collaborative benefits that should be received. Indicates the winning combination The target participating edge nodes are not included. Any sub-combination, Subcombinations The number of edge nodes included Indicates the winning combination The total number of edge nodes included, where ! represents factorial operation. This indicates that the target will participate in the edge node. Add sub-combinations The net benefit of the new combination's coordinated scheduling that can be generated by the newly formed combination Subcombinations The net revenue generated by the sub-combination collaborative scheduling, summation sign It means that for all satisfying and sub-combinations Perform summation; S404. Repeat steps S401-S403 until a winning combination is found. The actual collaborative benefit value of each participating edge node has been calculated.

7. A cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 6, characterized in that, S403 calculates the actual collaborative benefit value. Subsequently, it was verified that each target participated in the edge nodes. Actual collaborative benefits Are all of them no smaller than the target participating edge node? Submitted expected collaboration benefit base When any target participates in the edge node Appear If the conditions for individual rational participation are not met, the cloud-based scheduling center will abandon the winning collaborative combination.

8. A cloud-edge collaborative elastic scheduling method for heterogeneous computing resources according to claim 6, characterized in that, When to the sub-combination Add target participating edge nodes Afterwards, the new combination If the aggregated resource vector still does not meet the corresponding dimensional requirements of the multidimensional resource demand vector in any dimension, then the new combination... Ineffective collaborative combinations deemed incapable of completing the task; new combinations The corresponding new combined collaborative scheduling net benefit value A value of 0 indicates that the target is an edge node. Relative to the sub-combination The marginal synergistic benefit contribution value is calculated as follows: .

9. A cloud-edge collaborative elastic scheduling system for heterogeneous computing resources, applicable to the cloud-edge collaborative elastic scheduling method for heterogeneous computing resources as described in any one of claims 1-9, characterized in that, The system includes: a cloud-based dispatch center and multiple edge nodes; The cloud-based dispatch center includes: The demand broadcasting module is used to parse the arriving task set into a multi-dimensional resource demand vector and broadcast the multi-dimensional resource demand vector and the corresponding task completion benefit value to multiple registered edge nodes. The response collection module is used to receive the sealed collaborative participation response information submitted by each edge node. The sealed collaborative participation response information includes the computing power resource package generated by each edge node and the corresponding expected collaborative benefit base. The combinatorial optimization module is used to gather all sealed collaborative response information, with the goal of maximizing the net benefit of collaborative scheduling, and to determine a winning collaborative combination among all edge node combinations that can collaboratively satisfy the multidimensional resource demand vector. The Shapley value allocation module is used to calculate the actual collaborative revenue value that each participating edge node in the winning collaborative combination should receive using the Shapley value allocation algorithm, and to allocate revenue according to the actual collaborative revenue value. The scheduling and execution module is used to schedule the corresponding subtasks in the task set to each participating edge node in the winning collaborative combination for execution, and to monitor the task execution status. The edge nodes include: The resource monitoring module is used to periodically collect the real-time availability of various heterogeneous computing resources on the local machine and generate a description of available computing resource packages. The collaborative participation generation module is used to generate at least one computing power resource package according to the description of the available computing power resource package, set a corresponding expected collaborative benefit base for each computing power resource package, and encapsulate the computing power resource package and the expected collaborative benefit base into sealed collaborative participation response information. The task execution module is used to receive sub-tasks issued by the cloud scheduling center and call the corresponding heterogeneous computing resources to complete the execution.

Images