Computing task scheduling method and system based on network resource status prediction of time sequence diagram

Abstract

The present invention discloses a computing task scheduling method and system based on time-series graph network resource state prediction, which relates to the field of computing resource allocation technology. The method includes: collecting computing resource state data based on computing power nodes and modeling it as time-series heterogeneous graph data; constructing a time-series heterogeneous graph attention network THGAT model, and using node feature and edge feature prediction loss as the loss function and minimizing the prediction error as the goal to perform offline training on the THGAT model; inputting the time-series heterogeneous graph data into the THGAT model to predict the resource state of the future time window; converting the computing task into a computing demand vector, adopting the multi-objective optimization resource allocation THGATRM algorithm, and allocating computing power resources to the computing task according to the resource state of the corresponding time window; and executing the computing task according to the computing power resource allocation plan. Through the technical solution of the present invention, the accuracy of computing power resource prediction is improved, and efficient allocation and optimization of resources are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of computing resource allocation, and in particular to a computing task scheduling method based on timing graph network resource state prediction and a computing task scheduling system based on timing graph network resource state prediction. Background Art

[0002] With the rapid development of artificial intelligence and high-performance computing, the demand for computing resources has exploded. Modern computing infrastructure is evolving towards a heterogeneous, distributed, and dynamic approach, comprised of diverse computing resources such as GPUs, CPUs, and TPUs. In such a complex computing environment, traditional static scheduling strategies are no longer sufficient to efficiently allocate resources.

[0003] Traditional temporal graph neural networks typically separate time and graph processing, such as applying a GNN followed by an RNN to process the time series, or simply overlaying the time dimension. Traditional graph attention mechanisms typically only consider node features, ignoring the important information of edge features. In traditional GNNs, message passing typically only considers the features of the source node and the attention weight. Summary of the Invention

[0004] In response to the above problems, the present invention provides a computing task scheduling method and system based on the prediction of network resource status of a time-series graph. By modeling the graph structure of the computing power network, a time-series heterogeneous graph attention network THGAT model is constructed. By deeply fusing time and space information and integrating edge features into the attention calculation and message passing process, the computing power resources of the computing power network are predicted in multiple steps. The computing demand vector of the computing task and the predicted resources are allocated based on multi-objective optimization to achieve efficient allocation and optimization of resources.

[0005] To achieve the above object, the present invention provides a computing task scheduling method based on network resource status prediction of a time sequence diagram, comprising:

[0006] The computing resource status data of the heterogeneous computing network is collected based on computing nodes and modeled as time-series heterogeneous graph data;

[0007] Constructing a temporal heterogeneous graph attention network (THGAT) model, and training the THGAT model offline with node feature prediction loss and edge feature prediction loss as loss functions and minimizing prediction error as the goal;

[0008] Inputting the time series heterogeneous graph data into the THGAT model to predict the resource status of the computing power network in the future time window;

[0009] The computing task requested by the user is converted into a computing demand vector, and the multi-objective optimization resource allocation THGATRM algorithm is used to allocate computing power resources to the computing task according to the resource status of the corresponding time window;

[0010] Execute computing tasks according to the computing power resource allocation plan.

[0011] In the above technical solution, preferably, the sequence of the time series heterogeneous graph data is , any time t The computing power network heterogeneous graph is ;

[0012] in, is a node set, is the node type mapping function, For the moment t The edge set of is the edge type mapping function, For the moment t The feature set of all nodes, For the moment t The feature set of all edges.

[0013] In the above technical solution, preferably, the THGAT model includes a heterogeneous graph encoder, a relationship-aware attention layer with edge feature enhancement, a temporal information fusion module, and a multi-step prediction decoder;

[0014] The heterogeneous graph encoder is used to process features of different types of nodes and edges;

[0015] The relationship-aware attention layer of edge feature enhancement performs relationship-aware transformation on node and edge features, adopts a K-head attention mechanism to calculate the attention score of edge feature enhancement, and normalizes the attention weights of neighbor nodes of the same type of edge. At the same time, edge features are introduced to directly modulate the message content, and multi-head attention messages of different variable types and different neighbor nodes are aggregated for each central node;

[0016] The temporal information fusion module captures temporal dependencies through the gated recurrent unit (GRU), and directly uses the spatial aggregation information of each time step to update the temporal state.

[0017] The multi-step prediction decoder uses a multi-task learning framework to predict the dynamic features and edge dynamic features of computing power nodes based on the final representation of each computing power node at a specific moment.

[0018] In the above technical solution, preferably, the THGAT model is trained offline using the node feature prediction loss and the edge feature prediction loss as loss functions and minimizing the prediction error as the goal. The specific process includes:

[0019] The training goal of the offline training of the THGAT model is to minimize the prediction error of the future preset time step, and the loss function includes the node feature prediction loss and the edge feature prediction loss;

[0020] The loss function is:

[0021]

[0022] Among them, the node feature prediction loss is:

[0023]

[0024] The edge feature prediction loss is:

[0025]

[0026] in, and is the error weight vector, and are the weights of node and edge losses;

[0027] Set decreasing weights for the predicted time step k ,in is the attenuation factor.

[0028] In the above technical solution, preferably, the node features include static features and dynamic features. , the static characteristics are

[0029]

[0030] in, is the embedding representation of the node type, is the peak computing power vector, is the total memory or video memory capacity, is the memory bandwidth, is the usage cost per unit time, is the location identifier;

[0031] Dynamic Features Changes over time:

[0032]

[0033] in, is a multi-dimensional load indicator, For real-time available computing power, is the available memory or video memory, is the instantaneous power consumption, is the temperature, is the task queue length or estimated waiting time, For health status;

[0034] The edge features include static features and dynamic features , the static characteristics are:

[0035]

[0036] in, is the embedding representation of the edge type, is the theoretical maximum bandwidth, is the basic delay, is the number of network hops, Data transmission costs;

[0037] The dynamic features are:

[0038]

[0039] in, For real-time available bandwidth, is the instantaneous end-to-end delay, is the packet loss rate, For delay jitter, is the link utilization rate, , Indicates the link status is {normal, degraded, disconnected}.

[0040] In the above technical solution, preferably, the time series heterogeneous graph data is input into the THGAT model to predict the resource status of the computing power network in the future time window. The specific process includes:

[0041] Performing linear transformation on the node features and edge features, and performing relationship-aware feature projection transformation on different types of node features and edge features;

[0042] Based on the transformed node features and edge features, the K-head attention mechanism is used to calculate the attention score of edge feature enhancement:

[0043]

[0044]

[0045] in, represents vector concatenation, is specific to the edge type Learnable attention vectors for

[0046] For each attention head, normalize the attention weights of neighboring nodes with the same type of edges:

[0047]

[0048] in, is a node v By type The set of neighbors connected by the edges;

[0049] Introducing edge features to directly modulate message content:

[0050]

[0051] in, and is specific to the edge type and attention heads k The learnable weight matrix, edge feature term Allowing network characteristics to directly modulate the content of information being conveyed;

[0052] For each central node v , aggregate multi-head attention messages from different edge types and different neighbor nodes:

[0053]

[0054] Among them, ⊕ represents weighted average;

[0055] Capturing temporal dependencies through the gated recurrent unit GRU:

[0056]

[0057]

[0058]

[0059]

[0060] Among them, σ is the sigmoid function, ⊙ represents element-wise multiplication, 、 、 、 、 、 is a learnable parameter;

[0061] Aggregate the spatial information at each time step Directly participate in time status updates;

[0062] According to the final representation of each computing power node at time t , using a multi-task learning framework to predict the future H Node and edge dynamic characteristics at time steps:

[0063]

[0064] in, For the predicted future Step node dynamic characteristics, Is specific to the node type The decoder network of

[0065]

[0066] in, For the predicted future H Step edge dynamic characteristics.

[0067] In the above technical solution, preferably, the specific process of converting the computing task requested by the user into a computing demand vector includes:

[0068] According to the computing task request submitted by the user, it is converted into a standardized computing demand vector CRV:

[0069]

[0070] in, TaskID and TaskType Indicates the task code and task type, ResourceRequests Indicates the demand for different types of resources, Constraints represents the constraints of the task, Metadata Represents the meta information of the task.

[0071] In the above technical solution, preferably, a multi-objective optimization resource allocation THGATRM algorithm is used to allocate computing power resources to the computing task according to the resource status of the corresponding time window. The specific method includes:

[0072] Generate a candidate resource set for each type of resource demand corresponding to the computing task based on the node computing power resource status and computing demand vector predicted by the THGAT model;

[0073] Generating an effective resource combination that satisfies all resource requirements and interconnection constraints based on the candidate resource machines, wherein the constraints include: satisfying all resource requirements, satisfying task deadlines, satisfying budget constraints, satisfying resource availability, and satisfying network connectivity requirements;

[0074] For each effective resource combination, a multi-objective comprehensive score is calculated based on the optimization objectives and corresponding weights of minimizing task completion time, minimizing resource cost, minimizing energy consumption, and maximizing resource utilization;

[0075] Find the best start time window for the first preset number of valid resource combinations with the highest scores;

[0076] An effective resource combination with the highest multi-objective comprehensive score and a feasible time window is selected as the computing power resource allocation plan for the computing task.

[0077] In the above technical solution, preferably, the computing task scheduling method based on the network resource status prediction of the time sequence diagram further includes:

[0078] Monitor the execution status of the computing task, and perform model update and strategy optimization on the THGAT model based on the monitoring data;

[0079] Performing regular retraining and incremental learning training on the THGAT model;

[0080] The performance of the THGAT model is continuously evaluated based on preset indicators, and a model update is triggered when the prediction error exceeds a preset threshold.

[0081] The present invention further proposes a computing task scheduling system based on time sequence graph network resource state prediction, which applies the computing task scheduling method based on time sequence graph network resource state prediction disclosed in any of the above technical solutions, including:

[0082] The data acquisition and modeling module is used to collect computing resource status data of the heterogeneous computing network based on computing nodes and model it as time-series heterogeneous graph data;

[0083] A model construction and training module is used to construct a temporal heterogeneous graph attention network THGAT model and perform offline training on the THGAT model with the goal of minimizing the prediction error, using node feature prediction loss and edge feature prediction loss as loss functions;

[0084] A resource status prediction module, configured to input the time series heterogeneous graph data into the THGAT model and predict the resource status of the computing power network in a future time window;

[0085] The computing resource allocation module is used to convert the computing task requested by the user into a computing demand vector, and adopt the multi-objective optimization resource allocation THGATRM algorithm to allocate computing power resources to the computing task according to the resource status of the corresponding time window;

[0086] The computing task execution module is used to execute computing tasks according to the computing power resource allocation plan.

[0087] Compared with the existing technology, the beneficial effects of the present invention are: by modeling the graph structure of the computing power network, it can more accurately reflect the complexity of the actual computing power network, construct a time-series heterogeneous graph attention network THGAT model, and deeply integrate time and space information in the message transmission and state update process, as well as integrate edge features into the attention calculation and message transmission process, to perform multi-step prediction of the computing power resources of the computing power network, thereby improving the accuracy of the prediction, and allocating the computing demand vector of the computing task and the predicted resources based on multi-objective optimization, thereby realizing efficient allocation and optimization of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 A flow chart of a computing task scheduling method based on timing diagram network resource status prediction disclosed in an embodiment of the present invention;

[0089] Figure 2 A schematic diagram of the processing flow of the temporal heterogeneous graph attention network THGAT model disclosed in one embodiment of the present invention;

[0090] Figure 3 A schematic diagram of the processing flow of the edge feature enhanced attention mechanism disclosed in one embodiment of the present invention;

[0091] Figure 4 A schematic diagram of the processing flow of the THGATRM algorithm for multi-objective optimization resource allocation disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0092] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0093] The present invention will be described in further detail below with reference to the accompanying drawings:

[0094] like Figure 1 As shown, a computing task scheduling method based on timing diagram network resource status prediction provided by the present invention includes:

[0095] The computing resource status data of the heterogeneous computing network is collected based on computing nodes and modeled as time-series heterogeneous graph data;

[0096] A temporal heterogeneous graph attention network (THGAT) model was constructed, and offline training was performed on the THGAT model with the goal of minimizing the prediction error, using node feature prediction loss and edge feature prediction loss as loss functions.

[0097] Input the time series heterogeneous graph data into the THGAT model to predict the resource status of the computing power network in the future time window;

[0098] The computing task requested by the user is converted into a computing demand vector, and the multi-objective optimization resource allocation THGATRM algorithm is used to allocate computing power resources to the computing task according to the resource status of the corresponding time window;

[0099] Execute computing tasks according to the computing power resource allocation plan.

[0100] In this implementation, by modeling the graph structure of the computing power network to more accurately reflect the complexity of the actual computing power network, a time-series heterogeneous graph attention network THGAT model is constructed. By deeply integrating time and space information in the process of message transmission and state update, and integrating edge features into the attention calculation and message transmission process, the computing power resources of the computing power network are predicted in multiple steps, which improves the accuracy of the prediction. The computing demand vector of the computing task and the predicted resources are allocated based on multi-objective optimization, thereby achieving efficient allocation and optimization of resources.

[0101] In the above embodiment, preferably, the sequence of the time-series heterogeneous graph data is , any time t The computing power network heterogeneous graph is ;

[0102] in, A node collection represents all computing resource units in the computing network, such as GPUs, CPU groups, TPUs, etc.

[0103] Node type mapping function, mapping each node to a predefined type set ;

[0104] For the moment t The edge set represents the network connection between resource nodes;

[0105] Edge type mapping function, mapping each edge to a predefined set of types ;

[0106] For the moment t The feature set of all nodes;

[0107] For the moment t The feature set of all edges.

[0108] Specifically, the heterogeneous graph explicitly distinguishes different types of computing resources and network connections, and can more accurately reflect the complexity of the real computing power network.

[0109] In the above embodiment, preferably, the node features include static features and dynamic features. , the static characteristics are

[0110]

[0111] in, is the embedding representation of the node type, is the peak computing power vector, such as . is the total memory or video memory capacity (GB), is the memory bandwidth (GB / s), is the usage cost per unit time, It is a location identifier used for network delay estimation.

[0112] Dynamic Features Changes over time:

[0113]

[0114] in, is a multi-dimensional load indicator, such as . For real-time available computing power, is the available memory or video memory, is the instantaneous power consumption (W), is the temperature (℃), is the task queue length or estimated waiting time, The health status is represented by a numerical code {healthy, warning, error}.

[0115] Edge features are used to describe the state of network connections between nodes, including static features and dynamic features. , the static characteristics are:

[0116]

[0117] in, is the embedding representation of the edge type, is the theoretical maximum bandwidth (Gbps), is the base delay (ms), is the number of network hops, is the data transfer cost ($ / GB).

[0118] The dynamic features are:

[0119]

[0120] in, is the real-time available bandwidth (Gbps), is the instantaneous end-to-end delay (ms), is the packet loss rate (%), is the delay jitter (ms), is the link utilization rate, , Indicates the link status is {normal, degraded, disconnected}.

[0121] Specifically, by dividing the above node status into static and dynamic features, and dividing the edge features into static and dynamic parts, the resource status can be captured more comprehensively.

[0122] like Figure 2 As shown, in the above embodiment, preferably, the THGAT model includes a heterogeneous graph encoder, a relationship-aware attention layer with edge feature enhancement, a temporal information fusion module, and a multi-step prediction decoder;

[0123] Heterogeneous graph encoder is used to process features of different types of nodes and edges;

[0124] The edge feature-enhanced relationship-aware attention layer performs relationship-aware transformation on node and edge features, uses a K-head attention mechanism to calculate the edge feature-enhanced attention score, and normalizes the attention weights of neighboring nodes with the same type of edge. At the same time, edge features are introduced to directly modulate the message content, and multi-head attention messages of different variable types and different neighboring nodes are aggregated for each central node.

[0125] The temporal information fusion module captures temporal dependencies through the gated recurrent unit (GRU), directly integrating the spatial aggregation information of each time step into the temporal state update, and capturing the evolution of the state over time.

[0126] The multi-step prediction decoder uses a multi-task learning framework to predict the dynamic features and edge dynamic features of computing power nodes based on the final representation of each computing power node at a specific moment, thereby predicting the resource status in the future time window.

[0127] Specifically, the THGAT model can achieve deep fusion of spatiotemporal information. layer, at time step ,node In the The hidden state of the layer is represented as Initial state Obtained through linear transformation of node features:

[0128]

[0129] in and Is specific to the node type The learnable parameters.

[0130] Traditional graph attention mechanisms typically only consider node features while ignoring the important information of edge features. However, the edge-feature-enhanced relationship-aware attention mechanism in this implementation explicitly incorporates edge features into the attention calculation and message passing process.

[0131] In the above embodiment, preferably, the THGAT model is trained offline using the node feature prediction loss and the edge feature prediction loss as the loss function and minimizing the prediction error as the goal. The specific process includes:

[0132] The training goal of the offline training of the THGAT model is to minimize the prediction error of the future preset time step. The loss function includes the node feature prediction loss and the edge feature prediction loss.

[0133] The loss function is:

[0134]

[0135] Among them, the node feature prediction loss is:

[0136]

[0137] The edge feature prediction loss is:

[0138]

[0139] in, and is the error weight vector, and are the weights of node and edge losses;

[0140] Set decreasing weights for the predicted time step k ,in is the attenuation factor, which makes the model focus more on the accuracy of recent predictions.

[0141] like Figure 2 As shown, in the above embodiment, preferably, the time series heterogeneous graph data is input into the THGAT model to predict the resource status of the computing power network in the future time window. Based on the structure of the THGAT model in the above embodiment, the specific process of prediction includes:

[0142] like Figure 3 As shown, linear transformation is performed on node features and edge features, and relationship-aware feature projection transformation is performed on different types of node features and edge features:

[0143]

[0144]

[0145]

[0146] Representatives in t Moment, Model l Source node of the layer u The hidden state of Representative Node u The new feature representation obtained after type-aware linear transformation, It is specifically designed for The learnable weight matrix for this type of node is is a node type specific The learnable bias vector. It is specifically designed for The learnable weight matrix for nodes of this type. A node type specific The learnable bias vector.

[0147] Representatives in t moment, connecting nodes u and v That side e The eigenvector of . is a learnable weight matrix. is a learnable bias vector specific to the edge type. The new feature representation obtained after the edge (u, v) undergoes type-aware linear transformation.

[0148] Based on the transformed node features and edge features, calculate the edge feature enhanced attention score:

[0149]

[0150]

[0151] in, represents vector concatenation, is specific to the edge type The learnable attention vector.

[0152] Furthermore, to enhance the expressiveness of the model, a K-head attention mechanism is adopted:

[0153]

[0154] For each attention head, normalize the attention weights of neighboring nodes with the same type of edges:

[0155]

[0156] in, is a node v By type The set of neighbors connected by edges.

[0157] In traditional GNNs, message passing typically only considers the features and attention weights of the source node. However, in this implementation, edge features are introduced to directly modulate the message content, making message passing more accurate:

[0158]

[0159] in, and is specific to the edge type and attention heads k The learnable weight matrix, edge feature term Allows network characteristics (e.g., bandwidth, latency) to directly modulate the information content being transferred.

[0160] For each central node v , aggregate multi-head attention messages from different edge types and different neighbor nodes:

[0161]

[0162] Here, ⊕ represents weighted average.

[0163] In the THGAT model, unlike the traditional method of processing space first and time later, a mechanism for deeply integrating spatiotemporal information is designed.

[0164] Capturing temporal dependencies through the gated recurrent unit GRU:

[0165]

[0166]

[0167]

[0168]

[0169] Among them, σ is the sigmoid function, ⊙ represents element-wise multiplication, 、 、 、 、 、 are learnable parameters.

[0170] Through this deep fusion of spatiotemporal information mechanism, the spatial aggregation information at each time step Directly participate in the time state update, realizing the deep integration of spatiotemporal information, rather than simple staged processing. When Initialize to a zero vector or initialize with a specific function.

[0171] To achieve forward-looking scheduling, completeL After the layer THGAT is calculated, the final representation of each computing power node at time t is , considering the different prediction difficulties and importance of different features, a multi-task learning framework is adopted to predict the future H Node features for each time step:

[0172]

[0173] in, For the predicted future Step node dynamic characteristics, Is specific to the node type The decoder network can be implemented as a multi-layer perceptron (MLP) or a Transformer-based sequence decoder.

[0174] Similarly, predicting the future H Edge features at time steps:

[0175]

[0176] in, To predict the future H The dynamic characteristics of the edge of the step.

[0177] In the above embodiment, preferably, the specific process of converting the computing task requested by the user into a computing demand vector includes:

[0178] The computing task request submitted by the user is converted into a standardized computing demand vector CRV to match the predicted resource status. CRV is formally defined as:

[0179]

[0180] in, TaskID and TaskType Indicates task code and task type;

[0181] Indicates the demand for different types of resources;

[0182] , describe the detailed requirements for each type of resource;

[0183] Represents the constraints of the task;

[0184] Metadata Represents the metadata of the task, such as user ID, submission time, etc.

[0185] like Figure 4As shown, in the above embodiment, preferably, a multi-objective optimization resource allocation THGATRM algorithm is used to allocate computing power resources to computing tasks according to the resource status of the corresponding time window. The specific method includes:

[0186] Based on the node computing resource status and computing demand vector predicted by the THGAT model, a candidate resource set is generated for each type of resource demand corresponding to the computing task;

[0187] Generate an effective resource combination that meets all resource requirements and interconnection constraints based on the candidate resource machines, where the constraints include: meeting all resource requirements, meeting task deadlines, meeting budget constraints, meeting resource availability, and meeting network connectivity requirements;

[0188] For each effective resource combination, a multi-objective comprehensive score is calculated based on the optimization objectives and corresponding weights of minimizing task completion time, minimizing resource cost, minimizing energy consumption, and maximizing resource utilization;

[0189] Find the best start time window for the first preset number of valid resource combinations with the highest scores;

[0190] The effective resource combination with the highest multi-objective comprehensive score and a feasible time window is selected as the computing power resource allocation plan for the computing task.

[0191] Specifically, in order to find the optimal resource allocation solution for computing tasks, a multi-objective optimization resource allocation THGATRM algorithm is designed:

[0192] Given: The future predicted by the THGAT model H Node computing power resource status at the time step and the computational requirement vector CRV of the task.

[0193] Goal: Find a resource allocation solution ,in are the resources assigned to the task, is the start time on the resource, is the estimated execution time.

[0194] Optimization goal:

[0195] 1. Minimize task completion time: ;

[0196] 2. Minimize resource costs: ;

[0197] 3. Minimize energy consumption: ;

[0198] 4. Maximize resource utilization.

[0199] Constraints:

[0200] 1. Meet all resource requirements: type, quantity, performance, memory, etc.

[0201] 2. Meeting the task deadline: ;

[0202] 3. Meeting budget constraints: ;

[0203] 4. Resource availability: Resources must be available and meet performance requirements within the allocated time period;

[0204] 5. Network connectivity meets requirements: Network connectivity between allocated resources must meet bandwidth and latency requirements.

[0205] In the specific allocation process, the THGATRM algorithm finds the approximately optimal resource allocation solution through multi-stage optimization based on the predicted resource status.

[0206] Phase 1: Resource Candidate Set Generation

[0207] For each type of resource demand , generate a set of candidate resources that meet basic performance requirements:

[0208]

[0209] Representative is Class resource requirements The generated candidate resource set, Represents a variable, representing an arbitrary resource node in the computing network. Represents the collection of all resource nodes in the entire computing network. Is the first filter condition, indicating v It must be a real resource node in the network. This is the most basic premise. Is a function that returns a node v type. This is from the resource requirements R i The expected resource type extracted from . P stands for performance, where The THGAT model is for nodes v Compute resources (nodes) in the future time window from t to t+H v ) is the predicted value of the available computing power. The THGAT model is for nodes v The predicted value of available memory / GPU memory in the future time window from t to t+H.

[0210] Phase 2: Generation of effective resource combinations

[0211] Generate candidate resource combinations that satisfy all resource requirements and connectivity constraints:

[0212]

[0213] To evaluate interconnect constraints, examine the predicted bandwidth and latency:

[0214]

[0215] Represents a pair of resource nodes, This is a set that contains all network connections required by the task. This set is generated based on the interconnect requirements (InterConnect) in the task's requirement vector (CRV). represents the predicted available bandwidth in the time window from t to t+H, MinBW Indicates the minimum bandwidth required by the task, represents the instantaneous delay predicted in the time window from t to t+H, MaxLa Represents the maximum allowed delay for a task.

[0216] Phase 3: Multi-objective scoring and ranking

[0217] A composite score is calculated for each effective resource combination, taking into account several objectives:

[0218]

[0219] in, is the weight of each goal, which is dynamically adjusted according to system strategy and task priority. Each sub-score is calculated as follows:

[0220] 1. Time rating:

[0221]

[0222] Indicates the absolute latest time point at which the task must be completed. It is the estimated completion time point when using the resource combination Comb to perform the task. Indicates the time when the user submits the task.

[0223] 2. Cost Rating:

[0224]

[0225] Budget Represents the maximum cost budget set by the user for this task. EstCost ( Comb) represents the estimated total cost of using resource combination Comb to perform the task.

[0226] 3. Energy consumption score:

[0227]

[0228] EstEnergy ( Comb ) The total power consumed by using the resource combination Comb to perform the task. MaxEnergy Represents a theoretical or maximum allowable energy consumption 。

[0229] 4. Utilization score:

[0230]

[0231] | Comb |Indicates the cardinality of the set Comb, that is, the number of resource nodes contained in the combination. AvgUtilIncrease ( v ), which is the average utilization increment.

[0232] Phase 4: Optimal time window selection and scheduling

[0233] For the top K resource combinations with the highest scores, find the optimal start time window:

[0234]

[0235] in, is the expected execution time of the task, It's in time Using resource combinations comprehensive cost.

[0236] Finally, the resource combination with the highest comprehensive score and feasible time window is selected as the allocation plan.

[0237] During implementation, the THGATRM algorithm is further explained through the following pseudo code.

[0238] Algorithm: THGATRM(CRV, THGAT_Model, CurrentTime t)

[0239] enter:

[0240] CRV: Computational Requirement Vector of the task

[0241] THGAT_Model: trained THGAT model

[0242] t: current time

[0243] Output:

[0244] ResourceAllocation: resource allocation plan, including resource set, start time and expected execution time

[0245] step:

[0246] 1. / / Get THGAT's future state prediction

[0247] FutureStates = THGAT_Model.Predict(t, H)

[0248] 2. / / Phase 1: Generate candidate sets for each resource requirement

[0249] CandidateSets = {}

[0250] for each ResourceRequest R_i in CRV.ResourceRequests:

[0251] Candidates = []

[0252] for each node v in V where τ(v) matches R_i.Type:

[0253] if MeetsRequirements(v, R_i, FutureStates):

[0254] Candidates.append(v)

[0255] CandidateSets[R_i] = Candidates

[0256] 3. / / Phase 2: Generate effective resource combinations

[0257] ValidCombinations = []

[0258] GenerateResourceCombinations(CandidateSets, CRV, FutureStates,ValidCombinations)

[0259] 4. / / Phase 3: Scoring and Ranking

[0260] ScoredCombinations = []

[0261] for each Combination Comb in ValidCombinations:

[0262] Score = ComputeScore(Comb, CRV, FutureStates)

[0263] ScoredCombinations.append((Comb, Score))

[0264] Sort ScoredCombinations by Score in descending order

[0265] 5. / / Stage 4: Selecting the optimal time window

[0266] for each (Comb, Score) in TopK of ScoredCombinations:

[0267] BestStartTime = FindBestStartTime(Comb, CRV, FutureStates)

[0268] if BestStartTime is valid:

[0269] return CreateResourceAllocation(Comb, BestStartTime, CRV)

[0270] 6. / / Handle the no-solution situation

[0271] return FallbackStrategy(CRV).

[0272] In the above embodiment, preferably, the computing task scheduling method based on the network resource status prediction of the time sequence diagram further includes:

[0273] Monitor the execution status of computing tasks and perform model updates and strategy optimization on the THGAT model based on monitoring data;

[0274] Perform regular retraining and incremental learning training on the THGAT model;

[0275] The performance of the THGAT model on preset indicators is continuously evaluated, and a model update is triggered when the prediction error exceeds a preset threshold.

[0276] During implementation, the model is retrained at regular intervals (e.g., 24 hours). Model updates are triggered when the prediction error exceeds a threshold. Incremental training is performed using the latest data to avoid catastrophic forgetting. A validation set is maintained to continuously evaluate model performance.

[0277] Among them, system performance is evaluated by the following indicators:

[0278] (1) Resource utilization: ;

[0279] in T is the total duration of the entire evaluation cycle, V It is the collection of all resource nodes in the entire computing cluster. V | represents the cardinality of the set V, that is, the total number of resource nodes in the cluster. Indicates at a specific point in time t , a specific resource node V utilization rate.

[0280] (2) Task completion time:

[0281] ;

[0282] Tasks The set of all tasks processed during the evaluation cycle,| Tasks |Indicates the total number of tasks. EndTime task The time point when the execution of a specific task ends. SubmitTime task The time point when the same task is submitted to the system by the user.

[0283] (3) Energy efficiency: ;

[0284] Refers to a single task The total amount of work involved. Workload task Refers to a single task The total amount of work involved. is the instantaneous power consumption of a specific node v at a specific time t. It means that the total energy consumed by the entire cluster during the entire evaluation period is obtained by integrating (discrete summing) the power consumption of all nodes at all time points.

[0285] (4) SLA compliance ratio:

[0286] Tasks met Represents the set of all tasks that successfully meet the SLA requirements, | Tasks met | represents the number of tasks that successfully meet the SLA, | Tasks |Indicates the total number of tasks submitted.

[0287] (5) Prediction accuracy: the mean square error (MSE) or mean absolute error (MAE) of the THGAT model’s prediction of future resource status.

[0288] The present invention further proposes a computing task scheduling system based on time sequence graph network resource state prediction, which applies the computing task scheduling method based on time sequence graph network resource state prediction disclosed in any of the above embodiments, including:

[0289] The data acquisition and modeling module is used to collect computing resource status data of the heterogeneous computing network based on computing nodes and model it as time-series heterogeneous graph data;

[0290] The model construction and training module is used to build the temporal heterogeneous graph attention network THGAT model and perform offline training on the THGAT model with the goal of minimizing the prediction error, using node feature prediction loss and edge feature prediction loss as loss functions.

[0291] The resource status prediction module is used to input the time series heterogeneous graph data into the THGAT model to predict the resource status of the computing power network in the future time window;

[0292] The computing resource allocation module is used to convert the computing tasks requested by users into computing demand vectors, and adopt the multi-objective optimization resource allocation THGATRM algorithm to allocate computing power resources to computing tasks according to the resource status of the corresponding time window;

[0293] The computing task execution module is used to execute computing tasks according to the computing power resource allocation plan.

[0294] According to the computing task scheduling system based on timing diagram network resource status prediction disclosed in the above-mentioned embodiment, the functions to be implemented by each module thereof correspond to the respective steps of the computing task scheduling method based on timing diagram network resource status prediction disclosed in the above-mentioned embodiment. During the implementation process, operations are performed with reference to the above-mentioned embodiment, which will not be repeated here.

[0295] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A computing task scheduling method based on network resource status prediction of a time sequence diagram, characterized in that: include: The computing resource status data of the heterogeneous computing network is collected based on computing nodes and modeled as time-series heterogeneous graph data; Constructing a temporal heterogeneous graph attention network (THGAT) model, and training the THGAT model offline with node feature prediction loss and edge feature prediction loss as loss functions and minimizing prediction error as the goal; Inputting the time series heterogeneous graph data into the THGAT model to predict the resource status of the computing power network in the future time window; The computing task requested by the user is converted into a computing demand vector, and the multi-objective optimization resource allocation THGATRM algorithm is used to allocate computing power resources to the computing task according to the resource status of the corresponding time window; Execute computing tasks according to the computing resource allocation plan; The multi-objective optimization resource allocation THGATRM algorithm is used to allocate computing power resources to the computing task according to the resource status of the corresponding time window. The specific method includes: Generate a candidate resource set for each type of resource demand corresponding to the computing task based on the node computing power resource status and computing demand vector predicted by the THGAT model; Generating an effective resource combination that satisfies all resource requirements and interconnection constraints based on the candidate resource machines, wherein the constraints include: satisfying all resource requirements, satisfying task deadlines, satisfying budget constraints, satisfying resource availability, and satisfying network connectivity requirements; For each effective resource combination, a multi-objective comprehensive score is calculated based on the optimization objectives and corresponding weights of minimizing task completion time, minimizing resource cost, minimizing energy consumption, and maximizing resource utilization; Find the best start time window for the first preset number of valid resource combinations with the highest scores; Selecting an effective resource combination with the highest multi-objective comprehensive score and a feasible time window as the computing power resource allocation plan for the computing task; The THGAT model includes a heterogeneous graph encoder, an edge feature enhanced relationship-aware attention layer, a temporal information fusion module, and a multi-step prediction decoder. The heterogeneous graph encoder is used to process features of different types of nodes and edges; The relationship-aware attention layer of edge feature enhancement performs relationship-aware transformation on node and edge features, adopts a K-head attention mechanism to calculate the attention score of edge feature enhancement, and normalizes the attention weights of neighbor nodes of the same type of edge. At the same time, edge features are introduced to directly modulate the message content, and multi-head attention messages of different variable types and different neighbor nodes are aggregated for each central node; The temporal information fusion module captures temporal dependencies through the gated recurrent unit (GRU), and directly uses the spatial aggregation information of each time step to update the temporal state. The multi-step prediction decoder uses a multi-task learning framework to predict the dynamic features and edge dynamic features of computing power nodes based on the final representation of each computing power node at a specific moment.

2. The computing task scheduling method based on timing diagram network resource status prediction according to claim 1 is characterized in that: The sequence of the time series heterogeneous graph data is: , any time t The computing power network heterogeneous graph is ; in, is a node set, is the node type mapping function, For the moment t The edge set of is the edge type mapping function, For the moment t The feature set of all nodes, For the moment t The feature set of all edges.

3. The computing task scheduling method based on timing diagram network resource status prediction according to claim 1 is characterized in that: The THGAT model is trained offline using node feature prediction loss and edge feature prediction loss as loss functions and minimizing prediction error as the goal. The specific process includes: The training goal of the offline training of the THGAT model is to minimize the prediction error of the future preset time step, and the loss function includes the node feature prediction loss and the edge feature prediction loss; The loss function is: ； Among them, the node feature prediction loss is: ； The edge feature prediction loss is: ； in, and is the error weight vector, and are the weights of node and edge losses; Set decreasing weights for the predicted time step k ,in is the attenuation factor.

4. The computing task scheduling method based on timing diagram network resource status prediction according to claim 3 is characterized in that: The node features include static features and dynamic features , the static characteristics are ； in, is the embedding representation of the node type, is the peak computing power vector, is the total memory or video memory capacity, is the memory bandwidth, is the usage cost per unit time, is the location identifier; Dynamic Features Changes over time: ； in, is a multi-dimensional load indicator, For real-time available computing power, is the available memory or video memory, is the instantaneous power consumption, is the temperature, is the task queue length or estimated waiting time, For health status; The edge features include static features and dynamic features , the static characteristics are: ； in, is the embedding representation of the edge type, is the theoretical maximum bandwidth, is the basic delay, is the number of network hops, Data transmission costs; The dynamic features are: ； in, For real-time available bandwidth, is the instantaneous end-to-end delay, is the packet loss rate, is delay jitter, is the link utilization rate, , Indicates the link status is {normal, degraded, disconnected}.

5. The computing task scheduling method based on timing diagram network resource status prediction according to claim 4 is characterized in that: The time series heterogeneous graph data is input into the THGAT model to predict the resource status of the computing power network in the future time window. The specific process includes: Performing linear transformation on the node features and edge features, and performing relationship-aware feature projection transformation on different types of node features and edge features; Based on the transformed node features and edge features, the K-head attention mechanism is used to calculate the attention score of edge feature enhancement: ； ； in, represents vector concatenation, is specific to the edge type Learnable attention vectors for For each attention head, normalize the attention weights of neighboring nodes with the same type of edges: ； in, is a node v By type The set of neighbors connected by the edges; Introducing edge features to directly modulate message content: ； in, and is specific to the edge type and attention heads k The learnable weight matrix, edge feature term Allowing network characteristics to directly modulate the content of information being conveyed; For each central node v , aggregate multi-head attention messages from different edge types and different neighbor nodes: ； Among them, ⊕ represents weighted average; Capturing temporal dependencies through the gated recurrent unit GRU: ； ； ； ； Among them, σ is the sigmoid function, ⊙ represents element-wise multiplication, 、 、 、 、 、 is a learnable parameter; Aggregate the spatial information at each time step Directly participate in time status updates; According to the final representation of each computing power node at time t , using a multi-task learning framework to predict the future H Node and edge dynamic characteristics at time steps: ； in, For the predicted future Step node dynamic characteristics, Is specific to the node type The decoder network of ； in, For the predicted future H Step edge dynamic characteristics.

6. The computing task scheduling method based on timing diagram network resource status prediction according to claim 5 is characterized in that: The specific process of converting the computing task requested by the user into a computing demand vector includes: According to the computing task request submitted by the user, it is converted into a standardized computing demand vector CRV: ； in, TaskID and TaskType Indicates the task code and task type, ResourceRequests Indicates the demand for different types of resources, Constraints represents the constraints of the task, Metadata Represents the meta information of the task.

7. The computing task scheduling method based on time sequence diagram network resource status prediction according to any one of claims 1 to 6, characterized in that: Also includes: Monitor the execution status of the computing task, and perform model update and strategy optimization on the THGAT model based on the monitoring data; Performing regular retraining and incremental learning training on the THGAT model; The performance of the THGAT model is continuously evaluated based on preset indicators, and a model update is triggered when the prediction error exceeds a preset threshold.

8. A computing task scheduling system based on network resource status prediction of a time sequence diagram, characterized in that: The computing task scheduling method based on timing diagram network resource status prediction according to any one of claims 1 to 7 is applied, comprising: The data acquisition and modeling module is used to collect computing resource status data of the heterogeneous computing network based on computing nodes and model it as time-series heterogeneous graph data; A model construction and training module is used to construct a temporal heterogeneous graph attention network THGAT model and perform offline training on the THGAT model with the goal of minimizing the prediction error, using node feature prediction loss and edge feature prediction loss as loss functions; A resource status prediction module, configured to input the time series heterogeneous graph data into the THGAT model and predict the resource status of the computing power network in a future time window; The computing resource allocation module is used to convert the computing task requested by the user into a computing demand vector, and adopt the multi-objective optimization resource allocation THGATRM algorithm to allocate computing power resources to the computing task according to the resource status of the corresponding time window; The computing task execution module is used to execute computing tasks according to the computing power resource allocation plan.

Images