A dual feature-aware resource allocation method and system based on learnable tree navigation
By constructing a learnable resource catalog tree and reinforcement learning agents, and combining them with a game theory framework, the problems of low resource retrieval efficiency and multi-objective scheduling imbalance in computing power networks are solved, achieving efficient resource allocation and balance of interests among multiple parties, and improving the system's adaptability and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG COMP SCI CENTNAT SUPERCOMP CENT IN JINAN
- Filing Date
- 2026-03-19
- Publication Date
- 2026-05-29
AI Technical Summary
In computing networks, there are problems such as low resource retrieval efficiency and imbalance in multi-objective scheduling. Existing methods are difficult to adapt to the dynamic task requirements and conflicting objectives of multiple stakeholders in heterogeneous resource environments.
A dual-feature perception resource allocation method based on learnable tree navigation is adopted. By constructing a learnable resource catalog tree, a reinforcement learning agent is used for resource retrieval. The optimal resource allocation scheme is generated through game theory between resource providers and users, thereby achieving a balance of interests among multiple parties.
It significantly improves resource retrieval efficiency, reduces latency and memory overhead in large-scale environments, achieves a balance of interests among multiple parties and enhances overall system performance, and supports dynamic expansion and multi-scenario scheduling needs.
Smart Images
Figure CN121880029B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a dual-feature perception resource allocation method and system based on learnable tree navigation, which is particularly suitable for efficient retrieval and resource allocation in a multi-vendor heterogeneous resource integration environment, and belongs to the field of computing network technology. Background Technology
[0002] Computing networks integrate geographically distributed heterogeneous resources such as computing, storage, and networking to build a multimodal infrastructure supporting advanced applications such as machine learning, artificial intelligence, and the Internet of Things. Their core value lies in achieving cross-domain resource aggregation and collaborative scheduling to alleviate problems such as uneven resource distribution and unbalanced task loads in time and space. Existing research mostly employs non-machine learning, neural network, and deep reinforcement learning methods to optimize task offloading and resource allocation, but still faces the following technical bottlenecks:
[0003] 1. Existing resource retrieval methods suffer from efficiency bottlenecks in large-scale data environments, while matching techniques based on fixed rules are difficult to adapt to the heterogeneous and dynamic task requirements of computing power networks.
[0004] 2. Imbalance in multi-objective scheduling: The optimization objectives of multiple stakeholders, such as users and resource providers, are inconsistent (e.g., minimizing task completion time versus maximizing resource utilization, distance, affinity, etc.). Traditional methods struggle to coordinate multiple objectives and balance the needs of multiple parties. Summary of the Invention
[0005] This invention aims to solve problems such as heterogeneous entity descriptions, low retrieval efficiency, and scheduling target conflicts in computing power networks, and provides a resource allocation method and system based on resource representation for efficient retrieval and balance of interests among multiple parties.
[0006] Terminology Explanation:
[0007] 1. Resource Representation: Standardize the description of heterogeneous and multi-source entities such as tasks, system nodes, data, and resources in the computing power network, and abstract the attributes of entities in different dimensions such as location, category, characteristics, spatiotemporal, and extension into standardized feature units that can be read by machines.
[0008] 2. Semantic atom: A unit that can express feature information and cannot be further subdivided.
[0009] 3. Learnable Resource Directory Tree: A data model with a hierarchical index structure built based on resource representations. This tree structure can be used by machine learning (such as reinforcement learning agents) to learn features, thereby achieving efficient retrieval of massive resources.
[0010] The present invention adopts the following technical solution:
[0011] A dual-feature-aware resource allocation method based on learnable tree navigation includes the following steps:
[0012] S1 constructs a learnable resource directory tree based on resource representation. The resource directory tree includes top-level service nodes, regional service nodes, computing center nodes, cluster nodes, and resource nodes from top to bottom.
[0013] S2, based on the user's task description, generates a task requirement representation. The scheduler parses the requirements of each subtask for the top-level service node, regional service node, computing center node, and cluster node according to the task requirement representation, thereby constructing a retrieval requirement vector.
[0014] S3, based on the retrieval demand vector, the scheduler retrieves the resource subgraph that meets the requirements of each subtask from the resource directory tree through the reinforcement learning agent (Root-Navigator);
[0015] S4, based on the resource subgraph, generates an optimal resource allocation scheme that balances the interests of all parties through a two-way game between resource providers and users.
[0016] In this invention, all entity representations in the system are stored in the database of the representation management server. The reinforcement learning agent (Root-Navigator) loads the resource directory tree and provides a directory tree update interface to update the directory tree in real time (the representation management server calls the interface provided by the agent when performing operations such as creation or modification); it can also learn the search path corresponding to the retrieval requirements, thereby quickly retrieving the corresponding resources when applying them.
[0017] After a user submits their task description to the computing network system, the representation management server generates a task requirement representation, which is then submitted to the scheduler. Workflow tasks, on the other hand, represent the resource requirements and network requirements of each subtask's dependencies.
[0018] The scheduler includes a task requirement parsing method, a reinforcement learning agent (Root-Navigator), and a method for calculating the optimal solution based on a resource subgraph and the game theory between resource providers and users. After receiving the task requirement representation, the scheduler uses the task requirement parsing method to obtain the top-level position, regional position, computing center position, and category feature requirement of each subtask's resource requirement and each dependent edge's network resource requirement, thereby constructing a corresponding retrieval requirement vector. Using the retrieval requirement vector of each subtask and dependent edge, the scheduler, using the Root-Navigator, can obtain a resource subgraph that matches the requirements. Therefore, the optimal solution is calculated based on the resource subgraph using the game theory between resource providers and users.
[0019] Preferably, in step S1, resource characterization It can be expressed as follows:
[0020] (1)
[0021] in, This represents the top-level position of the entity. To represent the regional location of an entity. These represent the computing power center location of an entity, together constituting the logical location features of the expression; , , These represent the first, second, and third categories of the entity, which together express the category dimension features; A unique identifier for the entity, created based on time, and unique throughout the entire system; It is a list representing the attributes of an entity, such as the name of the GPU cloud server, processor type, storage size, etc. It is also a list, representing the quantity of entities, such as total memory size, total video memory size, number of processor cores, etc. , , The three semantic atoms together express the feature dimension; It is a set of high and low load time periods for an entity, where each element represents the load situation for a time period. Latitude and longitude represent the physical location information of an entity. This can be used in formula (9) to calculate the distance between the user's required location and the allocated resource location; and The two semantic atoms together express the spatiotemporal dimension characteristics; and The supplier and access permission information of an entity are expressed separately, which together constitute the extended dimension. Semantic atoms can be added to the extended dimension as needed.
[0022] By utilizing the top-level location representation, regional location representation, computing center location representation, and category dimension representation in the resource representation, a resource directory tree is constructed. The top-level service node is a server set up to manage the various regional location representations within it. The regional service node is a server set up to manage the computing center location representations within it. The computing center node is a server set up to manage the representations of various resource clusters within it. The cluster node consists of multiple resource nodes.
[0023] The top-level service node, regional service node, computing center node, and cluster node in the resource directory tree store information about their subordinate nodes and provide information query interfaces.
[0024] Preferably, in step S2, the demand vector is retrieved. The formal representation of the task's requirements for top-level service nodes, regional service nodes, computing center nodes, and cluster nodes, among which... Indicates the first Matching requirements for layer nodes This indicates that no requirements have been specified. , To meet the needs of top-level service nodes, The corresponding regional service node requirements, To meet the needs of the computing power center nodes, The requirements for corresponding cluster nodes.
[0025] Preferably, in step S3, the resource directory tree is defined as a hierarchical graph structure. ,in Indicates that it has a depth attribute The set of nodes, Denotes a set of edges subject to hierarchical connectivity constraints: if and only if hour ,in Represents a node Depth within the resource directory tree, Represents a node Depth within the resource directory tree;
[0026] Reinforcement learning agents learn resource features from the resource catalog tree. The resource subgraph (a computing network contains all computing, storage, and network resources, forming a general graph; the resource subgraph contains all resources that can satisfy the requirements) is another subgraph. The retrieval process for all required resources is formalized as a finite-state decision process, and the model of the reinforcement learning agent is used. ;in Representing the state space, ; Represents the action space, , The nodes in the resource directory tree, excluding the nodes, represent the nodes in the resource directory tree. Other nodes; Indicates the state transition probability; Indicates a reward; Indicates the discount factor. ;
[0027] The formulas for state transition probability, reward, and value table are shown in formulas (2), (3), and (4), respectively:
[0028] (2)
[0029] (3)
[0030] (4)
[0031] In the state transition probability formula (2), the current state is The current node is (Node level is) The next level of requirement is... Action space Represents a node All child nodes, The best child node; For the next state, where Indicates the current node The set of requirements for subsequent level nodes; function Indicates from the value table Search status Next action Corresponding value; expression Indicates selection status The action with the highest value at that time ;
[0032] In the reward calculation formula (3), Ensure that the selected node attribute values meet the requirements. This represents the node attribute value, such as the location attribute of a top-level service node being CN, while the top-level service node requires... If it is also CN, then we have ;
[0033] The value table update formula is shown in formula (4), where For the updated value, For learning rate, State Below, the maximum value of all actions;
[0034] The reinforcement learning agent operates in two phases: offline and online. In the offline phase, the model value table is updated iteratively using training data. The value table represents the reward corresponding to the current state and the optional action pair. In the online phase, the pre-trained value table is loaded to perform the retrieval task.
[0035] Preferably, in step S3, during the offline phase, the model of the reinforcement learning agent... The input is a resource directory tree. and retrieval demand vector The output is a value table for each state-action pair. The steps for a single training round are as follows:
[0036] ① Initialize the value table to empty, and start from the root node (depth 0), with the initial state set to ,in This indicates the requirement for the first layer, i.e., the requirement for the top-level service node;
[0037] ② Regarding the current state Select the optimal action using formula (2) based on the value table. If the value table does not contain information for the corresponding current state-action pair, a random optional action is returned as the optimal action.
[0038] ③Transition to the next state ,in Indicates transfer to Then, the matching requirements for the next level of nodes;
[0039] ④ Calculate the reward using formula (3) and update it using formula (4). ;
[0040] ⑤ Repeat steps ② to ④ until the demand vector is obtained. All requirements were matched.
[0041] Preferably, in step S3, during the online phase, the model of the reinforcement learning agent... The input is a resource directory tree. , Search demand vector Value table of each action in each state The output is a resource node graph, i.e., a resource subgraph, that matches the requirements of all top-level service nodes, regional service nodes, computing center nodes, and cluster nodes. The specific steps are as follows:
[0042] a. The initialization process starts from the root node (depth 0), with the initial state set to... Initialize a stack, and... Push it onto the stack;
[0043] b. Pop the top element from the stack. To perform state transition; if If an element is the parent node of a leaf node (i.e., a cluster node), then all its child nodes (i.e., resource nodes) are retained. The next element is popped from the stack and then checked to see if it is the parent node. If it is, all its child nodes are retained as well. This process is repeated continuously. When an element is no longer the parent node, step c is executed.
[0044] c. Regarding the current state ,like Then, according to the value table, use formula (2) to select the optimal action. and will Push onto the stack; if Then all possible next states (correspond The actions in the sequence are pushed onto the stack in order;
[0045] d. Repeat steps b and c until the stack is empty, then output all matching leaf nodes, i.e. resource nodes.
[0046] In the resource directory tree, leaf nodes represent all resource nodes. Step b determines whether the current node is the parent node (i.e., a cluster node) of a leaf node (i.e., a resource node). If so, all leaf nodes under the current node are retained. If not, it indicates that the current node is a top-level service node, a regional service node, or a computing center node. Step c will perform a state transition as needed, such as transitioning the state to a top-level service node. =CN and regional service node requirements , When an SDSW node is available, the service will be transferred from the top-level service node CN to the regional service node SDSW, because after training, the state ( CN The value corresponding to the SDSW key-value pair for the action is the largest in the value table.
[0047] Preferably, in practice, a task can typically be decomposed into multiple interrelated subtasks. The requirements of each subtask are characterized from five dimensions: logical location, category, characteristics, spatiotemporal context, and extension. The task is modeled as... , yes Requirements description matrix for each subtask This is the correlation matrix between the various subtasks, and the formula is as follows:
[0048] (5)
[0049] in, , , , , The first one describes the second one respectively. The logical location requirements, categories, characteristics, spatiotemporal requirements, and extended attribute requirements of each subtask can be used to create a complete requirement representation for that subtask. . Subtasks For subtasks The successor subtask, and To meet the requirements for network link types, This refers to the amount of data transmitted. Feature dimensions include task completion time, location constraints, and resource attributes and quantity requirements.
[0050] In step S4, to coordinate the competing interests of all parties, the workflow task resource allocation problem is modeled as a Stackelberg game. In the Stackelberg game model, the resource provider is considered as the leader, pursuing system-level goals: minimizing idle resources, providing geographically proximate resources, and allocating high-affinity resources; the user is considered as the follower, focusing on minimizing the workflow completion time.
[0051] When the task When submitted to the scheduler, the representations of each subtask are parsed: the logical location dimension of the subtask representation is used to determine the top-level service node. Regional service nodes and computing power center nodes The requirements; the category dimension is used to identify sub-task types, and based on the mapping relationship between sub-task types and resource categories, the required resource types are further determined, thereby clarifying the resource cluster. For example, in an artificial intelligence model training task, the corresponding resource cluster is a GPU server cluster. There will be predefined mapping relationships here; specific constraints are parsed from the feature dimension. (e.g., delay, position, etc.) final definition Search demand vector set :
[0052] (6)
[0053] in Indicates the first Vectors of subtasks Represents the vector of network resource demand;
[0054] Subsequently, the scheduler uses the trained reinforcement learning agent model. and resource directory tree Obtain resource subgraphs that meet the requirements of each subtask, designate them as candidate resources, and then determine the hierarchical representation of all candidate resources. :
[0055] (7)
[0056] in Indicates the first The set of candidate resource representations corresponding to each subtask This represents the set of network resource representations;
[0057] Resource representation In For the attribute representation of the entity, such as the name, processor type, and storage size of the GPU cloud server, the hierarchical representation of all candidate resources is obtained in formula (7). Next, it is necessary to determine whether the required resources are compatible with... If the resource representations are the same, then from... Remove representations that do not match the required resource characteristics. For example, if the required resource characteristics are 4 cores and 8GB, then a single characteristic of a resource representation is incorrect. When the system has 2 cores and 4GB of RAM, this resource representation will be removed.
[0058] On the leader side, acquire the candidate resource set. Then, calculate and generate subtasks. Rating matrix The specific process is as follows:
[0059] (8)
[0060] in, and These represent the completion times of computation tasks and storage tasks, respectively. Different calculation methods are required for the completion times of computation and storage tasks. For task computation, For the computing power of cloud servers, This represents the current utilization rate of the processor. Additional overhead for computational tasks; For data read and write volume, To store the read and write speed of the storage server, Read / write bandwidth utilization rate Delaying metadata operations for storage access;
[0061] For location-sensitive tasks, the physical distance between the required location and the resource location is... The calculation is performed using the Haversine formula, as shown in formula (9), where For the Earth's radius, , ; Indicates the longitude of the resource. Indicates the longitude required for the task. Indicates the dimension of the resource. This indicates the dimension of task requirements; this information is carried in the resource and task representations.
[0062] (9)
[0063] Define subtasks With resources Affinity is The calculation formula is shown in (10):
[0064] (10)
[0065] in , and Time matching degree Time matching degree Matching degree with access permissions The weights; For time matching degree, For supply matching, For access permission matching degree; when the representation is a resource representation, Expressed as This represents the time requirement of a subtask for resources; when represented as a task representation... Expressed as This indicates a period of low resource load. Used for calculation , The length of the overlapping time periods and They represent , The length of the time period; and These represent the subtask's demand for suppliers and the supplier information for the resources, respectively. and These represent the subtask's resource permission requirements and the resource permission information, respectively.
[0066] It is worth noting that resource characterization In and It is used in formula (10) to calculate the supply matching degree. Matching degree with access permissions If the representation is a task representation, Represented in the formula If the representation is a resource representation, Represented in the formula If the representation is a task representation, Represented in the formula If the representation is a resource representation, Represented in the formula .
[0067] The completion time of the computation task and the storage task is obtained based on formula (8); the physical distance between the required location of the subtask and the resource is obtained based on formula (9); and the affinity between the subtask and the resource is obtained based on formula (10).
[0068] use Subtasks In resources Completion time; Subtasks Demand location and resources The distance between them; Subtasks With resources Affinity was used to obtain a rating matrix. :
[0069] (11)
[0070] To eliminate the influence of dimensions between indicators and to address different subtasks Special constraints Given the different requirements for completion time, location, and affinity, for Each element in After normalization and weighted calculation, the formula is updated to (12):
[0071] (12)
[0072] in, Indicates weight, if task For time-delay and location-sensitive tasks, the completion time weight is... The position weight is 0.4. The affinity weight is 0.4. It is 0.2, and ;
[0073] based on Determine the ideal solution With negative ideal solution And calculate each element Proximity to positive and negative ideal solutions and The expression is as shown in formula (13):
[0074] (13)
[0075] in, , , ; Indicates list index, such as ,but ;
[0076] The leader's objective function is defined as follows: :
[0077] (14)
[0078] in, and The allocated resources Scores for each target With the The ideal solution for each sub-task With positive ideal solution The distance between them.
[0079] Preferably, in step S4, on the follower side, a candidate resource set is obtained. Then, the priority of each subtask is calculated; for subtasks without predecessors... Its priority Set to 1; for subtasks with a predecessor. Priority is defined as ,in Subtasks Given all direct predecessor tasks, calculate the completion time of all subtasks in each priority subtask set under different resource allocation schemes. The maximum completion time is defined as the completion time of that priority set. .
[0080] On the follower side, the main goal is to minimize workflow tasks. The overall completion time, which requires considering the candidate resource set Select a specific resource instance and for each dependent edge Match compatible network links Subtasks priority Based on the task association matrix Calculated. For subtasks without predecessors. Its priority Set to 1. For subtasks with a predecessor. Priority is defined as ,in Subtasks The set of all direct predecessor tasks. Priority is... The set of subtasks is defined as Define decision variables. When subtask Assigned to its resource set The first in Individual resources hour, .definition When dependent edges Select candidate network resource sets The first in Link hour, . Indicates all priorities are The maximum execution time in a subtask.
[0081] The objective function for followers is:
[0082] (15)
[0083] in, Indicates the target symbol. To minimize; Indicates the completion time;
[0084] constraint With constraints Ensure that each subtask is assigned to a unique resource instance and that each dependency edge is assigned to a unique network link of the matching type; constraints With constraints Ensure that the selected link corresponds to its associated resources, where This represents a set of links that conform to the required type. (Function) The value is 1 if the condition is true, and 0 otherwise. (Constraint) The rules specify the priority level for each priority level. Its completion time It must be no less than the completion time of any subtask under this priority. Subtasks In resources The completion time is calculated using formula (8). If the task is a computational task, then... If the task is a storage task .
[0085] Preferably, based on the leader's objective function Completion time of followers Determine the optimal resource allocation plan:
[0086] First, the leader assigns each task From its candidate resource representation set Select the best set of resource representations:
[0087] Leaders based on each resource Scores for each target With the The ideal solution for each sub-task With positive ideal solution Calculate the optionality value based on the distance between them: The larger the selectivity value, the better, because the numerator is the distance to the negative ideal solution; select the top m resources with the largest selectivity values to form the optimal resource representation set; m is preferably 10.
[0088] Then, the followers, from the best set of resource representations selected by the leader, traverse all possible combinations based on minimizing the overall workflow task completion time, thereby selecting the best resource allocation scheme.
[0089] A dual-feature-aware resource allocation system based on learnable tree navigation, used to implement the above-mentioned dual-feature-aware resource allocation method based on learnable tree navigation, includes: a representation system and a scheduler;
[0090] The representation system is used to: construct a learnable resource catalog tree based on resource representations;
[0091] The scheduler is used to: parse the requirements of each subtask for top-level service nodes, regional service nodes, and computing center nodes based on the task requirement representation, thereby constructing a retrieval requirement vector; based on the retrieval requirement vector, retrieve resource subgraphs that meet the requirements of each subtask in the resource directory tree through a reinforcement learning agent (Root-Navigator); based on the resource subgraphs, generate an optimal resource allocation scheme that balances the interests of all parties through a two-way game between resource providers and users.
[0092] For any details not covered in this invention, please refer to the prior art.
[0093] The beneficial effects of this invention are as follows:
[0094] 1. Optimize retrieval performance: The learnable resource directory tree combined with reinforcement learning agents significantly reduces retrieval latency and memory overhead in large-scale environments.
[0095] 2. Achieve a balance of interests among multiple parties: Introduce a game theory framework to coordinate the conflicting goals of users and providers in resource allocation, thereby improving the overall efficiency and fairness of the system.
[0096] 3. Enhanced system adaptability: The modular design supports dynamic expansion and is suitable for multi-scenario and multi-objective computing network resource scheduling needs. Attached Figure Description
[0097] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0098] Figure 1 This is a flowchart of the dual-feature perception resource allocation method based on learnable tree navigation of the present invention;
[0099] Figure 2 This is an architecture diagram of the dual-feature perception resource allocation system based on learnable tree navigation according to the present invention. Detailed Implementation
[0100] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. However, this is not the only description; all aspects not described in detail herein are based on conventional techniques in the art.
[0101] Example 1
[0102] A dual-feature-aware resource allocation method based on learnable tree navigation, such as Figure 1 As shown, it includes the following steps:
[0103] S1 constructs a learnable resource directory tree based on resource representation. The resource directory tree includes top-level service nodes, regional service nodes, computing center nodes, cluster nodes, and resource nodes from top to bottom.
[0104] S2 generates a task requirement representation based on the user's task description, as shown below:
[0105] / CN.SDSW.JN531-R1 / task.Universal.experiment / task0001.[[delaysensitive, location sensitive],[8x32G,base]].[CPU30%,Mem35%] / [(22,22,L_s)].(22.1,114.3) / ChinaMobile.rank1
[0106] CN.SDSW.JN531-R1 represents the logical location requirement, and task.Universal.experiment represents the category dimension value. The status dimension is divided into the identifier task0001; task attribute requirements [[delay sensitive, location sensitive], [8x32G, base]]; and quantity requirements [CPU30%, Mem35%]. Delay sensitive and location sensitive indicate that the task is latency-sensitive and location-sensitive, respectively; 8x32G represents the required resource specifications, and base represents the required resource type. CPU30% and Mem35% indicate that the CPU and memory utilization of the resources should be less than 30% and 35%, respectively. (22,22,L_s) indicates the time load status of the resources required by the task; (22.1,114.3) indicates the physical location requirement. ChinaMobile and rank1 indicate the preferred supplier and access permissions, respectively.
[0107] Based on the task requirement representation, the scheduler parses CN, SDSW, and JN531-R1 as tasks from the logical location dimension, and then parses their requirements for top-level service nodes, regional service nodes, computing center nodes, and cluster nodes. Based on the mapping relationship between task category `task.Universal.experiment` and resources, the scheduler determines the task's cluster requirements, thereby constructing a retrieval requirement vector. ;
[0108] S3, based on the retrieval requirement vector, the scheduler uses a reinforcement learning agent (Root-Navigator) to retrieve resource subgraphs from the resource directory tree that meet the requirements of each subtask; the scheduler uses the trained Root-Navigator model. and resource directory tree Obtain resource subgraphs that meet the requirements of each subtask, designate them as candidate resources, and then determine the hierarchical representation of all candidate resources. .
[0109] S4, based on the resource subgraph, generates an optimal resource allocation scheme that balances the interests of all parties through a two-way game between resource providers and users.
[0110] On the leader side, acquire the candidate resource set. Then, subtasks are generated based on formulas (8), (9), and (10). rating matrix To eliminate the influence of dimensions between indicators and to address different subtasks Special constraints Given the different requirements for completion time, location, and affinity, for Each element in After normalization and weighted calculation, the updated formula is (12), resulting in the normalized formula. .based on Determine the ideal solution With negative ideal solution And calculate each element using formula (13). Proximity to positive and negative ideal solutions and This allows us to determine the scores of different dimensions of objectives under different resource allocation schemes.
[0111] On the follower side, obtain the candidate resource set. Next, the priority of each subtask needs to be calculated. For subtasks without predecessors... Its priority Set to 1. For subtasks with a predecessor. Priority is defined as ,in Subtasks The set of all direct predecessor tasks. From this, the completion time of all subtasks in each priority subtask set under different resource allocation schemes can be calculated, and the maximum completion time is defined as the completion time of that priority set. .
[0112] Based on the leader's objective function Based on the objective functions of the leader and followers, the optimal resource allocation scheme is determined. Specifically, in the solution process, the leader assigns resources to each task... From its candidate resource representation set Select the best set of resource representations, based on each resource Scores for each target With the The ideal solution for each sub-task With positive ideal solution Calculate the optionality value based on the distance between them: The top 10 resources with the highest selectivity values are selected to form the optimal resource representation set;
[0113] Then, the followers, from the best set of resource representations selected by the leader, traverse all possible combinations based on minimizing the overall workflow task completion time, and thus select the best solution.
[0114] Example 2
[0115] A dual-feature-aware resource allocation system based on learnable tree navigation, such as Figure 2 As shown, the dual-feature-aware resource allocation method based on learnable tree navigation for implementing Embodiment 1 includes: a representation system and a scheduler;
[0116] The representation system is used to: construct a learnable resource catalog tree based on resource representations;
[0117] The scheduler is used to: parse the requirements of each subtask for top-level service nodes, regional service nodes, and computing center nodes based on the task requirement representation, thereby constructing a retrieval requirement vector; based on the retrieval requirement vector, retrieve resource subgraphs that meet the requirements of each subtask in the resource directory tree through a reinforcement learning agent (Root-Navigator); based on the resource subgraphs, generate an optimal resource allocation scheme that balances the interests of all parties through a two-way game between resource providers and users.
[0118] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A dual-feature-aware resource allocation method based on learnable tree navigation, characterized in that, Includes the following steps: S1 constructs a learnable resource directory tree based on resource representation. The resource directory tree includes top-level service nodes, regional service nodes, computing center nodes, cluster nodes, and resource nodes from top to bottom. S2, based on the user's task description, generates a task requirement representation. The scheduler parses the requirements of each subtask for the top-level service node, regional service node, computing center node, and cluster node according to the task requirement representation, thereby constructing a retrieval requirement vector. S3, Based on the retrieval demand vector, the scheduler uses a reinforcement learning agent to retrieve resource subgraphs that meet the requirements of each subtask from the resource catalog tree; S4, based on the resource subgraph, generates an optimal resource allocation scheme that balances the interests of all parties through a two-way game between resource providers and users; In step S1, resource characterization Expressed as follows: （1） in, This represents the top-level position of the entity. To represent the regional location of an entity. It represents the computing power center location of the entity and jointly expresses the logical location characteristics; , , These represent the first, second, and third categories of the entity, which together express the category dimension features; A unique identifier for the entity; It is a list representing the attributes of the entity; It is also a list, representing the number of entities; , , The three semantic atoms together express the feature dimension; It is a set of high and low load time periods for an entity, where each element represents the load situation for a time period; Latitude and longitude represent the physical location information of an entity; and The two semantic atoms together express the spatiotemporal dimension characteristics; and The supplier and access permission information of the entity are expressed separately, which together constitute the extended dimension; By utilizing the top-level location representation, regional location representation, computing center location representation, and category dimension representation in the resource representation, a resource directory tree is constructed. The top-level service node is a server set up to manage the various regional location representations within it. The regional service node is a server set up to manage the computing center location representations within it. The computing center node is a server set up to manage the representations of various resource clusters within it. The cluster node consists of multiple resource nodes. In the resource directory tree, the top-level service node, regional service node, computing center node, and cluster node store information about their subordinate nodes and provide information query interfaces. In step S2, the demand vector is retrieved. The formal representation of the task's requirements for top-level service nodes, regional service nodes, computing center nodes, and cluster nodes, among which... Indicates the first Matching requirements for layer nodes This indicates that no requirements have been specified. , To meet the needs of top-level service nodes, The corresponding regional service node requirements, To meet the needs of the computing power center nodes, The requirements for corresponding cluster nodes; In step S3, the resource directory tree is defined as a hierarchical graph structure. ,in Indicates that it has a depth attribute The set of nodes, Denotes a set of edges subject to hierarchical connectivity constraints: if and only if hour ,in Represents a node Depth within the resource directory tree, Represents a node Depth within the resource directory tree; The reinforcement learning agent learns resource features from a resource catalog tree, and the retrieval process is formalized as a finite-state decision process. The model of the reinforcement learning agent... ;in Representing the state space, ; Represents the action space. , The nodes in the resource directory tree, excluding the nodes, represent the nodes in the resource directory tree. Other nodes; Indicates the state transition probability; Indicates a reward; Indicates the discount factor. .
2. The dual-feature perception resource allocation method based on learnable tree navigation according to claim 1, characterized in that, In step S3, the formulas for the state transition probability, reward, and value table are shown in formulas (2), (3), and (4), respectively: （2） （3） （4） In the state transition probability formula (2), the current state is The current node is The next level of requirements is ; Action space Represents a node All child nodes, The best child node; For the next state, where Indicates the current node The set of requirements for subsequent level nodes; function Indicates from the value table Search status Next action Corresponding value; expression Indicates selection status The action with the highest value at that time ; In the reward calculation formula (3), Ensure that the selected node attribute values meet the requirements. Indicates the node attribute value; The value table update formula is shown in formula (4), where For the updated value, For learning rate, State Below, the maximum value of all actions; The reinforcement learning agent operates in two phases: offline and online. In the offline phase, the model value table is updated iteratively using training data. The value table represents the reward corresponding to the current state and the optional action pair. In the online phase, the pre-trained value table is loaded to perform the retrieval task.
3. The dual-feature perception resource allocation method based on learnable tree navigation according to claim 2, characterized in that, In step S3, the offline phase, the model of the reinforcement learning agent... The input is a resource directory tree. and retrieval demand vector ; The output is a value table for each state-action pair. The steps for a single training round are as follows: ① The initial value table is empty, and starting from the root node, the initial state is set to... ,in This indicates the requirement for the first layer, i.e., the requirement for the top-level service node; ② Regarding the current state Select the optimal action using formula (2) based on the value table. If the value table does not contain information for the corresponding current state-action pair, a random optional action is returned as the optimal action. ③Transition to the next state ,in Indicates transfer to Then, the matching requirements for the next level of nodes; ④ Calculate the reward using formula (3) and update it using formula (4). ; ⑤ Repeat steps ② to ④ until the demand vector is obtained. All requirements were matched.
4. The dual-feature perception resource allocation method based on learnable tree navigation according to claim 3, characterized in that, In step S3, during the online phase, the model of the reinforcement learning agent... The input is a resource directory tree. , Search demand vector Value table of each action in each state The output is a resource node graph, i.e., a resource subgraph, that matches the requirements of all top-level service nodes, regional service nodes, computing center nodes, and cluster nodes. The specific steps are as follows: a. Start the initialization process from the root node, setting the initial state to... Initialize a stack, and... Push it onto the stack; b. Pop the top element from the stack. To perform state transition; if If an element is the parent node of a leaf node (i.e., a cluster node), then all its child nodes (i.e., resource nodes) are retained. The next element is popped from the stack and then checked to see if it is the parent node. If it is, all its child nodes are retained as well. This process is repeated continuously. When an element is no longer the parent node, step c is executed. c. Regarding the current state ,like Then, according to the value table, use formula (2) to select the optimal action. and will Push onto the stack; if Then all possible next states Push them onto the stack one by one; d. Repeat steps b and c until the stack is empty, then output all matching leaf nodes, i.e. resource nodes.
5. The dual-feature perception resource allocation method based on learnable tree navigation according to claim 4, characterized in that, In step S4, the resource allocation problem is modeled as a Stackelberg game. In the Stackelberg game model, the resource provider acts as the leader, pursuing system-level goals: minimizing idle resources, providing geographically proximate resources, and allocating high-affinity resources; the user acts as the follower, focusing on minimizing workflow completion time. When the task When submitted to the scheduler, define Search demand vector set : （6） in Indicates the first Vectors of subtasks Represents the vector of network resource demand; Subsequently, the scheduler uses the trained reinforcement learning agent model. and resource directory tree Obtain resource subgraphs that meet the requirements of each subtask, designate them as candidate resources, and then determine the hierarchical representation of all candidate resources. : （7） in Indicates the first The set of candidate resource representations corresponding to each subtask This represents the set of network resource representations; On the leader side, acquire the candidate resource set. Then, calculate and generate subtasks. rating matrix The specific process is as follows: （8） in, and These represent the completion times of the computation task and the storage task, respectively. For task computation, For the computing power of cloud servers, This represents the current utilization rate of the processor. Additional overhead for computational tasks; For data read and write volume, To store the read and write speed of the storage server, Read / write bandwidth utilization rate Delaying metadata operations for storage access; For location-sensitive tasks, the physical distance between the required location and the resource location is... The calculation is performed using the Haversine formula, as shown in formula (9), where For the Earth's radius, , ; Indicates the longitude of the resource. Indicates the longitude required for the task. Indicates the dimension of the resource. Indicates the dimension of task requirements; （9） Define subtasks With resources Affinity is The calculation formula is shown in (10): （10） in , and Time matching degree Time matching degree Matching degree with access permissions The weights; For time matching degree, For supply matching degree, For access permission matching degree; when the representation is a resource representation, Expressed as This represents the time requirement of a subtask for resources; when represented as a task representation... Expressed as This indicates a period of low resource load. Used for calculation , The length of the overlapping time periods and They represent , The length of the time period; and These represent the subtask's demand for suppliers and the supplier information for the resources, respectively. and These represent the subtask's resource permission requirements and the resource permission information, respectively. The completion time of the computation task and the storage task is obtained based on formula (8); the physical distance between the required location of the subtask and the resource is obtained based on formula (9); and the affinity between the subtask and the resource is obtained based on formula (10). use Subtasks In resources Completion time; Subtasks Demand location and resources The distance between them; Subtasks With resources Affinity was used to obtain a rating matrix. : （11） To eliminate the influence of dimensions between indicators and to address different subtasks Special constraints Given the different requirements for completion time, location, and affinity, for Each element in After normalization and weighted calculation, the formula is updated to (12): （12） in, Indicates weight, and ; based on Determine the ideal solution With negative ideal solution And calculate each element Proximity to positive and negative ideal solutions and The expression is as shown in formula (13): （13） in, , , ; Indicates a list index; The leader's objective function is defined as follows: : （14） in, and The allocated resources Scores for each target With the The ideal solution for each sub-task With positive ideal solution The distance between them.
6. The dual-feature perception resource allocation method based on learnable tree navigation according to claim 5, characterized in that, In step S4, on the follower side, a candidate resource set is obtained. Then, the priority of each subtask is calculated; for subtasks without predecessors... Its priority Set to 1; for subtasks with a predecessor. Priority is defined as ,in Subtasks Given all direct predecessor tasks, calculate the completion time of all subtasks in each priority subtask set under different resource allocation schemes. The maximum completion time is defined as the completion time of that priority set. .
7. The dual-feature perception resource allocation method based on learnable tree navigation according to claim 6, characterized in that, Based on the leader's objective function Completion time of followers Determine the optimal resource allocation plan: First, the leader assigns each task From its candidate resource representation set Select the best set of resource representations: Leaders based on each resource Scores for each target With the The ideal solution for each sub-task With positive ideal solution Calculate the optionality value based on the distance between them: Select the top m resources with the largest selectivity values to form the optimal resource representation set; Then, the followers, from the best set of resource representations selected by the leader, traverse all possible combinations based on minimizing the overall workflow task completion time, thereby selecting the best resource allocation scheme.
8. A dual-feature sensing resource allocation system based on learnable tree navigation, characterized in that, A dual-feature-aware resource allocation method based on learnable tree navigation as described in any one of claims 1-7 includes: a representation system and a scheduler; The representation system is used to: construct a learnable resource catalog tree based on resource representations; The scheduler is used to: parse the requirements of each subtask for top-level service nodes, regional service nodes, and computing center nodes based on the task requirement representation, thereby constructing a retrieval requirement vector; based on the retrieval requirement vector, use a reinforcement learning agent to retrieve resource subgraphs that meet the requirements of each subtask in the resource catalog tree; and based on the resource subgraphs, generate an optimal resource allocation scheme that balances the interests of all parties through a two-way game between resource providers and users.