Multi-branch spatio-temporal feature fusion prediction method for heterogeneous GPU cluster task completion time

Abstract

The invention relates to the field of heterogeneous computing, in particular to a multi-branch spatio-temporal feature fusion prediction method for task completion time of a heterogeneous GPU cluster, and aims to accurately describe an internal rule of the task completion time in the heterogeneous GPU cluster. The method comprises the following steps: constructing a spatial relationship network diagram based on task characteristics and GPU resource characteristics in a heterogeneous GPU cluster; based on the spatial relation network diagram, spatial features of the tasks are extracted through spatial branches; meanwhile, based on the time sequence of the task, time features of the task are extracted through a time branch; carrying out fusion processing on the spatial features and the time features to obtain fusion features; and based on the fusion features, taking a full connection layer as an output layer, and mapping the fusion features to the predicted task completion time through an activation function by the output layer through linear transformation.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous computing, and in particular to a multi-branch spatiotemporal feature fusion prediction method for task completion time of heterogeneous GPU clusters. Background Technology

[0002] Driven by the wave of 6G full cloudification technology, network architecture is evolving towards deep collaboration between cloud, edge, and device, giving rise to diverse intelligent applications such as holographic communication and intelligent sensing. These applications continuously generate a large number of computationally intensive tasks such as machine learning and deep learning, while simultaneously placing stringent requirements on low latency and high reliability in computing power supply. Heterogeneous Graphics Processing Unit (GPU) clusters, due to their flexible and powerful parallel computing capabilities, have been widely adopted and have become a key computing infrastructure supporting the aforementioned 6G intelligent services. In actual production environments, heterogeneous GPU clusters typically consist of various GPUs with different architectures and performance levels. However, due to differences in computing performance, hardware architecture, and specifications among different types of GPUs, the execution time of tasks varies significantly across different types of GPUs, leading to low GPU resource utilization and overall wasted computing power within the cluster. Therefore, building a high-precision, highly generalizable task completion time prediction model has become a crucial prerequisite for achieving efficient scheduling of heterogeneous GPU clusters.

[0003] Currently, there are numerous research findings on resource prediction for computing clusters. Early research primarily focused on homogeneous computing cluster designs, and the proposed methods failed to adequately consider the significant differences in hardware performance and architecture among heterogeneous nodes. Subsequently, researchers began to focus on heterogeneous environments, introducing linear regression and machine learning methods to predict key indicators such as CPU, GPU, memory utilization, and task completion time. However, these methods often struggle to effectively capture the complex long-term dependencies and nonlinear dynamic changes in cloud workloads during the modeling process. In reality, due to the iterative and repetitive nature of jobs, there are significant temporal correlations between multiple tasks of the same user. Spatially, the current resource state of the cluster is affected by running tasks, which in turn affects the completion time of subsequent iterative or cyclical tasks. Although some studies have attempted to model from a single perspective, most methods have failed to systematically integrate the spatiotemporal dual correlations, which limits their prediction accuracy and generalization ability in heterogeneous environments.

[0004] Therefore, there is an urgent need for a task completion time prediction method that can deeply integrate spatiotemporal features to accurately depict the inherent laws of task execution in heterogeneous GPU clusters, thereby providing a reliable basis for high-efficiency computing resource scheduling in the 6G era. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-branch spatiotemporal feature fusion prediction method for task completion time in heterogeneous GPU clusters, aiming to accurately characterize the inherent laws of task execution in heterogeneous GPU clusters, thereby providing a reliable basis for high-efficiency computing resource scheduling in the 6G era.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a multi-branch spatiotemporal feature fusion prediction method for task completion time in heterogeneous GPU clusters, comprising: S1, constructing a spatial relationship network graph based on task features and GPU resource characteristics in the heterogeneous GPU cluster. Spatial relationship network graph is composed of edge set and node set Composition, represented as This describes the spatial dependencies between tasks; task characteristics include at least the task name, submission time, execution duration, and resource usage; GPU resource characteristics include GPU types such as NVIDIA Tesla T4, P100, V100, V100M32, and MISC; node set. For task set edge set S1. Represent the association between tasks executed using the same type of GPU; S2. Based on the spatial relationship network graph, extract the spatial features of the task through the spatial branch; simultaneously, based on the time series of the task, extract the temporal features of the task through the temporal branch; the temporal branch includes a bidirectional gated recurrent unit and a multi-head gated attention mechanism; the spatial branch includes a gated attention network; S3. Fuse the spatial features and temporal features to obtain fused features; based on the fused features, use the fully connected layer as the output layer, and after linear transformation, map the fused features to the predicted task completion time through an activation function.

[0007] Spatial Relationship Network Diagram in Step S1 The construction process includes: using an adjacency matrix A matrix representing the connection relationships between nodes; an adjacency matrix. It is by Composed of nodes A symmetric square matrix of order 1, when the nodes With nodes When there is a connectivity relationship, , denoted as having an edge When node With nodes When there is no connection between them, .

[0008] In step S2, the gated attention network updates the hidden state of nodes based on a multi-head attention mechanism. It assigns learnable weights to each attention head through a gating mechanism, dynamically adjusts the information aggregation process according to the local graph structure, adapts to the complexity of the association between nodes, and generates more refined node representations to improve the spatial feature extraction effect. At the same time, the gating mechanism can suppress the interference of irrelevant neighbor nodes to enhance the robustness of the model.

[0009] The process of extracting spatial features using spatial branches includes: given an input feature matrix Adjacency matrix of graph connection ,in It is the number of task nodes. It is the feature dimension of each node; it captures nodes in parallel through multiple independent attention heads. With neighboring nodes Different spatial dependencies between them are dynamically weighted for the output of each attention head through a gating mechanism to suppress redundant spatial information and enhance the contribution of effective features; the gating-filtered... The initial spatial features are spliced ​​together to obtain the spatial features of the nodes.

[0010] In step S2, the temporal branch uses a cascaded bidirectional gated loop unit multi-head gated attention mechanism to extract temporal features. The bidirectional gated loop unit captures the bidirectional dependencies of the temporal data, and the multi-head gated attention mechanism performs fine-grained screening and weighting based on the basic temporal features extracted by the bidirectional gated loop unit to improve feature quality and focus on key feature nodes.

[0011] The bidirectional gated recurrent unit includes a forward GRU unit and a backward GRU unit; the process of extracting temporal features from the temporal branch includes: given the input matrix of the temporal features, ,in, For sequence length, The input feature dimension is defined as follows: the input matrix is ​​fed in parallel into the forward GRU unit and the backward GRU unit to obtain the forward hidden state sequence and the backward hidden state sequence, respectively; the forward and backward hidden states at corresponding time steps are concatenated into vectors to form a bidirectional fused hidden state sequence to comprehensively capture the temporal dependency features in the task execution process; an attention mechanism is added to the bidirectional fused hidden state sequence to calculate the importance weight of the features at each time step to enhance the expression of key temporal features and suppress redundant information; the attention-weighted features are dynamically adjusted through a gating unit to adaptively strengthen the temporal relationships sensitive to the prediction target, and the enhanced temporal features are output.

[0012] Before step S1, a data preprocessing step is also included: the data preprocessing step includes data filtering, data anomaly detection and feature encoding; data filtering will filter out the task data that has been completed according to the task status; data anomaly detection will remove the values ​​of CPU, GPU and memory utilization exceeding 1; feature encoding will encode the username, task name and GPU type, and after standardizing the numerical features, the initial features with temporal relationship are obtained.

[0013] The fusion process in step S3 is achieved by splicing, which splices spatial features and temporal features according to dimensions to generate high-dimensional fusion features that combine spatiotemporal information.

[0014] The performance of the prediction results is evaluated using mean absolute error and root mean square error. Mean absolute error measures the average deviation between the predicted value and the true value, while root mean square error emphasizes the impact of larger deviations to reflect the actual accuracy of the prediction model.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application provides a multi-branch spatiotemporal feature fusion prediction method for task completion time in heterogeneous GPU clusters. It constructs a multi-feature fusion architecture with parallel spatial and temporal branches, effectively solving the prediction bias problem caused by traditional prediction models focusing only on a single feature dimension and ignoring spatiotemporal correlations. By deeply fusing spatial and temporal features, the prediction model can more comprehensively analyze the multiple factors affecting task completion time in heterogeneous environments. Compared to traditional prediction models based on linear regression or machine learning, this method significantly reduces prediction errors and improves the reliability and practicality of prediction results when dealing with complex heterogeneous cluster environments, providing a more accurate basis for subsequent resource scheduling and elastic scaling decisions.

[0016] 2. This invention constructs a spatial relationship network diagram. It fully considers the spatial dependencies between tasks and the impact of different types of GPUs in a heterogeneous GPU cluster on task completion time, breaking through the limitations of traditional methods that only consider the time sequence relationship of tasks based on homogeneous clusters.

[0017] 3. The temporal branch proposed in this invention uses a bidirectional gated recurrent unit (BI-GRU) cascaded with a multi-head gated attention mechanism (MGA) to extract temporal features. The bidirectional gated recurrent unit can capture the bidirectional dependencies of temporal data, and the multi-head gated attention mechanism further integrates the global context through the attention layer. At the same time, it embeds the gated unit, which changes the weight allocation method of the attention mechanism, so that it can dynamically adjust the feature weights according to the input data, improve the feature quality and focus on key feature nodes.

[0018] 4. This invention verifies the effectiveness and robustness of the model using a publicly available dataset of real-world heterogeneous GPU cluster trajectories from Alibaba. The output task completion time prediction results provide direct evidence for resource scheduling decisions in heterogeneous GPU clusters, helping the scheduling system achieve precise matching between tasks and GPU resources. This avoids resource waste caused by lightweight tasks occupying high-performance GPUs and prevents response latency caused by heavy-load tasks being assigned to low-performance GPUs, effectively improving cluster resource utilization. It can meet the low-latency scheduling requirements of 6G fully cloud-based networks, ultimately improving the overall operating performance of the cluster and user experience. Attached Figure Description

[0019] Figure 1 This is a flowchart of a multi-branch spatiotemporal feature fusion prediction method for task completion time of heterogeneous GPU clusters provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the construction of a graph structure provided in an embodiment of this application; Figure 3 This is a time branch model diagram provided in an embodiment of this application; Figure 4 This is a spatial branch model diagram provided in an embodiment of this application; Figure 5 This is a schematic diagram of a system model provided in an embodiment of this application; Figure 6 This is an experimental result diagram provided in an embodiment of this application; Figure 7 This is a visualization analysis diagram provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] For example, such as Figure 1 As shown in the embodiments of this application, a multi-branch spatiotemporal feature fusion prediction method for task completion time of heterogeneous GPU clusters is provided, including: S1. Constructing a spatial relationship network diagram based on task characteristics and GPU resource characteristics in heterogeneous GPU clusters. .

[0022] For example, refer to Figure 2 Spatial Relationship Network Diagram By edge set and node set Composition, represented as This primarily describes the spatial dependencies between tasks. Node set For task set edge set This indicates the association between tasks executed using the same type of GPU. Task characteristics include at least the task name, submission time, execution duration, and resource usage. GPU resource characteristics refer to different types of GPUs, such as NVIDIA Tesla T4, P100, V100, V100M32, and MISC.

[0023] Before step S1, a data preprocessing step is included: the data preprocessing step includes data filtering, data anomaly detection, and feature encoding. Data filtering will filter out task data that has been completed based on the task status; data anomaly detection will remove values ​​where CPU, GPU, and memory utilization exceed 1; feature encoding will appropriately encode the username, task name, and GPU type, and standardize the numerical features to ensure consistent scaling, finally obtaining initial features with temporal relationships; For continuous attributes such as GPU utilization and memory usage, normalization is used to eliminate dimensional differences. The normalization formula is as follows: in: for The normalized value, and The first The maximum and minimum values ​​of the data information for each task.

[0024] S2. Based on the spatial relationship network graph, extract the spatial features of the task through spatial branches; at the same time, based on the time series of the task, extract the temporal features of the task through temporal branches. For example, refer to Figure 3 In step S2, the temporal branch is a cascaded structure of a Bidirectional Gated Recurrent Unit (BI-GRU) and a Multi-Gated Attention (MGA) mechanism. This embodiment captures the bidirectional dependencies of time-series data using the BI-GRU, then focuses on the dynamic changes of key time nodes using the MGA, ultimately outputting temporal features whose weights can be dynamically adjusted. Constructing a multi-gated attention mechanism in the temporal branch allows attention to key information in the data features. Simultaneously, the multi-gated attention mechanism can adaptively allocate weights, focusing on the time step most relevant to the current prediction, thus making the BI-GRU more flexible in capturing long-term dependencies in the sequence.

[0025] GRU effectively mitigates long-sequence dependency issues through its gating mechanism of update and reset gates. Its "bidirectional" design captures both past-to-future and future-to-past temporal relationships, avoiding the limitation of unidirectional recurrent neural networks (RNNs) that can only focus on historical information. Furthermore, compared to Long Short-Term Memory (LSTM) networks, GRU has fewer parameters and lower computational cost, enabling it to extract fundamental temporal features from long-sequence data with lower computational overhead, providing more refined input for subsequent attention mechanisms. The calculation formula for GRU is: in, Indicates an update to the door. Indicates the candidate hidden state. This indicates the hidden state of the previous time step. Indicates the current hidden state. This represents the pointwise product operator.

[0026] Update Gate The calculation formula is in, This represents the Sigmoid function. This represents the weight matrix.

[0027] Candidate hidden state The calculation formula is in, Represents the hyperbolic tangent function. This indicates that the door is being reset.

[0028] Reset door The calculation formula is Reference Figure 3 The bidirectional gated recurrent unit consists of forward GRU units and backward GRU units. The forward GRU units calculate the hidden layer states sequentially from front to back, and the backward GRU units calculate the hidden layer states sequentially from back to front. Represented as in, This indicates forward GRU cell computation. This indicates backward GRU cell computation.

[0029] Backward GRU unit The calculation formula is In step S2, the bidirectional gated recurrent unit captures the dependencies between time-series data from the past to the future and from the future to the past through forward and backward computations, respectively, and outputs basic time-series features that fuse bidirectional time-series information as input to the multi-head gated attention mechanism. The multi-head gated attention mechanism performs refined filtering and weighting on the basic time-series features extracted by the bidirectional gated recurrent unit to improve feature quality and focus on key feature nodes.

[0030] As one possible implementation, the process of extracting temporal features using a temporal branch includes: given the input matrix of temporal features as... ,in, For sequence length, The input feature dimension is defined; the input matrix is ​​fed in parallel into the forward GRU unit and the backward GRU unit to obtain the forward hidden state sequence and the backward hidden state sequence, respectively; the forward and backward hidden states at corresponding time steps are concatenated as vectors to form a bidirectional fused hidden state sequence to comprehensively capture the temporal dependency features during task execution; the bidirectional fused hidden state at each time step is then used as the input feature dimension. A separate linear layer performs a projection transformation, mapping features from different time steps to a unified semantic space. Then, scaled dot-product attention is used to calculate the importance weights of features at each time step, enhancing the representation of key temporal features and suppressing redundant information. The attention-weighted features are then dynamically adjusted through a gating unit. The gating output passes through another linear layer for further dimensionality adjustment of the calibrated features to adaptively strengthen temporal relationships sensitive to the prediction target. Finally, the output passes through a fully connected layer. By aggregating all the information, the final time characteristics are obtained.

[0031] Multi-head gating attention mechanisms dynamically allocate the weights of attention heads by embedding gating units. Multi-head attention is composed of multiple single-head attention heads, each H emphasizing different relevances. The linear mapping of H is divided into a query matrix. Keyword matrix AND-value matrix Each element is projected onto multiple subspaces through different linear transformations, generating multiple subsets of queries, keywords, and values.

[0032] The key to the attention mechanism lies in data fusion through scaled dot product attention, which effectively alleviates the vanishing and exploding gradient problems, further stabilizing the training process of deep neural networks. Scaled dot product attention, as an important component of multi-head attention, is defined by the following formula for independently calculating attention weights in each subspace: in, This indicates the projection dimension, used for scaling to prevent the inner product value from becoming too large. express The transpose of .

[0033] Attention mechanism The calculation formula is in, , , Representing the query matrix Keyword matrix Sum matrix In the The weight matrix during attention head training. Indicates the number of keys in the subspace.

[0034] Gated weights capture the importance of different attention heads through a learnable non-linear function. The calculation formula is in, These are the weights of the fully connected layer.

[0035] Finally, through the multi-head gating attention mechanism, the input features are projected into multiple subspaces, resulting in attention vectors for each subspace. These vectors are then combined, and the correlation between embedding methods is analyzed from multiple perspectives to improve model performance. Multi-head gating attention is defined as follows: For example, refer to Figure 4 In step S2, the spatial branch is a gated attention network (GaAN) structure. The gated attention network updates the hidden state of nodes based on a multi-head attention mechanism, assigns learnable weights to each attention head through a gating mechanism, dynamically adjusts the information aggregation process according to the local graph structure, adapts to the complexity of the association between nodes, and generates more refined node representations to improve the spatial feature extraction effect. At the same time, the gating mechanism can suppress the interference of irrelevant neighbor nodes and enhance the robustness of the model.

[0036] For example, the process of extracting spatial features using spatial branches includes: given an input feature matrix Adjacency matrix of graph connection ,in It is the number of task nodes. It is the feature dimension of each node; it captures nodes in parallel through multiple independent attention heads. With neighboring nodes Different spatial dependencies Through gating mechanism Weights are dynamically assigned to the output of each attention head to suppress redundant spatial information and enhance the contribution of effective features; the gating-filtered... The initial spatial features are concatenated and finally passed through a fully connected layer. The aggregated features are subjected to nonlinear transformation to further fuse information and obtain the spatial features of the nodes.

[0037] The multi-head attention mechanism in Gated Attention Networks (GaAN) employs both key-value attention and dot-product attention. To extract more important and critical information, the attention mechanism allows the model to assign different weights to the input features. The formula for calculating the feature output is as follows: in, For the number of attention heads, Represents a node and its neighboring nodes Between Each weighted attention coefficient, For splicing characters, These represent different transformation parameters.

[0038] Weighted attention coefficient The calculation formula is in, This represents the dot product of vectors.

[0039] Learnable weights are assigned to each attention head through a gating mechanism. The specific process is as follows: in, This indicates a node. With neighboring nodes Between the first The weight of each head, Convolutional networks, node selection With neighboring nodes The network is constructed using the maximum and average values ​​between them. This represents the sigmoid activation function.

[0040] S3. The spatial and temporal features are fused to obtain the fused features. Based on the fused features, the fully connected layer is used as the output layer. After linear transformation, the fused features are mapped to the predicted task completion time through an activation function.

[0041] The fusion process in step S3 is achieved by splicing, which splices spatial features and temporal features according to dimensions to generate a high-dimensional fusion feature vector that combines spatiotemporal information.

[0042] As one possible approach, the performance of the prediction results can be evaluated using mean absolute error and root mean square error. Mean absolute error is used to measure the average deviation between the predicted value and the true value, while root mean square error is used to emphasize the impact of larger deviations, thus reflecting the actual accuracy of the prediction model more objectively.

[0043] Mean Absolute Error (MAE) is a commonly used regression evaluation metric that measures the average deviation between predicted and actual values. Root Mean Square Error (RMSE) can more accurately reflect the overall deviation between predicted and actual values, and more objectively reflect the prediction accuracy.

[0044] The formulas for calculating RMSE and MAE are as follows: in, and Indicates the actual value and the predicted value. This represents the number of test data.

[0045] This application provides a parallel multi-branch feature fusion task completion time prediction model, MSTP (Multi-Branch Spatiotemporal Feature Fusion Prediction Algorithm for TaskCompletion Time Prediction in Heterogeneous GPU Clusters), such as... Figure 5 As shown, this model addresses the need for heterogeneous computing power collaboration in a 6G fully cloud-based network. It extracts task sequence features through multiple branches to achieve accurate prediction of task completion time. The model comprises three core modules: a data preprocessing module, a multi-branch feature extraction module, and a completion time prediction module. The data preprocessing module performs data preprocessing before step S1, the multi-branch feature extraction module performs steps S1 and S2, and the time prediction module performs step S3.

[0046] This application uses Alibaba's publicly available heterogeneous GPU cluster real trajectory dataset to verify the effectiveness and robustness of the parallel multi-branch feature fusion task completion time prediction network model, and compares its performance with other baseline models to evaluate its merits.

[0047] To verify the effectiveness and applicability of the model provided in this application embodiment, the model training was carried out in a Linux Ubuntu 18.04 system environment, using an NVIDIA GeForce GTX 4080 GPU and Python 3.8 as the software environment.

[0048] We selected a subset of the cluster-trace-GPU-v2020 dataset from Alibaba's PAI (Platform of Artificial Intelligence) heterogeneous GPU dataset. This dataset covers training and inference tasks executed by various types of GPUs, all implemented using state-of-the-art machine learning algorithms. It provides rich and practical data support for research on heterogeneous GPU cluster scheduling. The data in the dataset is mainly distributed across seven files, providing information such as task execution status, CPU, GPU, and memory resource utilization, task start time, and end time.

[0049] To verify the performance advantages of the model in a 6G fully cloud-based network scenario, the MSTP model provided in this application was compared with models such as LSTM, BI-GRU, GNN, and GCN on the cluster-trace-GPU-v2020 dataset. The experimental results are as follows: Figure 6As shown. Experimental results show that the root mean square error (RMSE) and mean absolute error (MAE) obtained by the method provided in this application exhibit significant advantages, with both indicators at extremely low levels. This indicates that the method provided in this application can accurately capture the core influencing factors of task completion time in heterogeneous GPU clusters, and the average deviation between the predicted results and the true values ​​is extremely small. Specifically, compared with LSTM and BI-GRU models that only extract temporal features, the fusion mechanism fully considers the importance of spatial feature information to task completion time, enhancing its feature representation ability. Compared with GNN and GCN models that only extract spatial features, the method provided in this application adds a gated attention mechanism to the traditional GCN model. By assigning learnable weights to each attention head through the gating mechanism, it can better adapt to the complexity of the association between nodes and generate more refined node representations to improve the spatial feature extraction effect. Compared with gradient boosting decision tree (GBDT) and classification regression tree (CART) models, the method provided in this application considers the repetitiveness of the job and the influence of the cluster's resource status on the resource usage of the current task. The model provided in this application consistently outperforms other models. This result fully demonstrates that in the complex heterogeneous environment of a 6G fully cloud-based network, the modeling approach of fusing spatiotemporal features across multiple branches can effectively enhance the comprehensiveness of feature representation, thereby significantly optimizing the prediction of task completion time.

[0050] The embodiments of this application visualize and analyze experimental results, such as Figure 7 As shown. Figure 7 A visualization of the results of task completion time prediction using the model provided in the embodiments of this application on the cluster-trace-GPU-v2020 dataset is shown. Figure 7 (a) The vertical axis represents task execution time, and the horizontal axis represents the number of task samples. Specifically, the blue dotted curve represents the actual value of task execution time, while the red square curve represents the model's predicted execution time. Figure 7 (b) is a visualization of the error distribution. As can be seen from the graph, most errors are controlled within the range of [-0.075, 0.075]. Figure 7 It can be intuitively observed that the prediction results of the method provided in this application are basically close to the true values, which indicates that the method provided in this application has excellent prediction performance.

[0051] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0052] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-branch spatiotemporal feature fusion prediction method for task completion time in heterogeneous GPU clusters, characterized in that, include: S1. Constructing a spatial relationship network diagram based on task characteristics and GPU resource characteristics in heterogeneous GPU clusters. The spatial relationship network graph consists of an edge set and node set Composition, represented as This is used to describe the spatial dependencies between tasks; the task characteristics include at least the task name, submission time, execution duration, and resource usage; the GPU resource characteristics include GPU types of NVIDIA Tesla T4, P100, V100, V100M32, and MISC; the node set For task set The set of edges S2. Based on the spatial relationship network graph, extract spatial features of the tasks through spatial branches; simultaneously, based on the time series of the tasks, extract temporal features of the tasks through temporal branches; the temporal branches include bidirectional gated recurrent units and multi-head gated attention mechanisms; the spatial branches include gated attention networks; S3. Fuse the spatial features and temporal features to obtain fused features; Based on the fusion features, the fully connected layer is used as the output layer. After linear transformation, the fusion features are mapped to the predicted task completion time through an activation function.

2. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 1, characterized in that, Spatial Relationship Network Diagram in Step S1 The construction process includes: using an adjacency matrix A matrix representing the connection relationships between nodes; an adjacency matrix. It is by Composed of nodes A symmetric square matrix of order 1, when the nodes With nodes When there is a connectivity relationship, , denoted as having an edge When node With nodes When there is no connection between them, .

3. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 1, characterized in that, The gated attention network described in step S2 updates the hidden state of nodes based on a multi-head attention mechanism. It assigns learnable weights to each attention head through a gating mechanism, dynamically adjusts the information aggregation process according to the local graph structure, adapts to the complexity of the association between nodes, and generates more refined node representations to improve the spatial feature extraction effect. At the same time, the gating mechanism can suppress the interference of irrelevant neighbor nodes to enhance the robustness of the model.

4. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 2, characterized in that, The process of extracting spatial features using spatial branches includes: given an input feature matrix Adjacency matrix of graph connection ,in It is the number of task nodes. It is the feature dimension of each node; it captures nodes in parallel through multiple independent attention heads. With neighboring nodes Different spatial dependencies between them are dynamically weighted for the output of each attention head through a gating mechanism to suppress redundant spatial information and enhance the contribution of effective features; the gating-filtered... The initial spatial features are spliced ​​together to obtain the spatial features of the nodes.

5. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 1, characterized in that, In step S2, the time branch uses a bidirectional gated loop unit cascaded multi-head gated attention mechanism to extract time features. The bidirectional gated loop unit captures the bidirectional dependencies of time series data, and the multi-head gated attention mechanism performs fine-tuning and weighting based on the basic time series features extracted by the bidirectional gated loop unit to improve feature quality and focus on key feature nodes.

6. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 5, characterized in that, The bidirectional gated recurrent unit includes a forward GRU unit and a backward GRU unit; the process of extracting time features from the time branch includes: given the input matrix of the time features, ,in, For sequence length, The input feature dimension is defined as follows: the input matrix is ​​fed in parallel into the forward GRU unit and the backward GRU unit to obtain the forward hidden state sequence and the backward hidden state sequence, respectively; the forward and backward hidden states at corresponding time steps are concatenated into vectors to form a bidirectional fused hidden state sequence to comprehensively capture the temporal dependency features in the task execution process; an attention mechanism is added to the bidirectional fused hidden state sequence to calculate the importance weight of the features at each time step to enhance the expressive power of key temporal features and suppress redundant information; the attention-weighted features are dynamically adjusted through a gating unit to adaptively strengthen the temporal relationships sensitive to the prediction target, and the enhanced temporal features are output.

7. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 1, characterized in that, The data preprocessing steps preceding step S1 include data filtering, data anomaly detection, and feature encoding. Data filtering will filter out task data that is already completed based on the task status; Data anomaly detection removes values ​​where CPU, GPU, and memory utilization exceed 1. Feature encoding encodes the username, task name, and GPU type, and after standardizing the numerical features, it yields initial features with temporal relationships.

8. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 1, characterized in that, The fusion process in step S3 is achieved by splicing, which splices spatial features and temporal features according to dimensions to generate high-dimensional fusion features that combine spatiotemporal information.

9. The multi-branch spatiotemporal feature fusion prediction method for heterogeneous GPU cluster task completion time according to claim 1, characterized in that, The performance of the prediction results is evaluated using mean absolute error and root mean square error. The mean absolute error measures the average deviation between the predicted value and the true value, while the root mean square error reflects the actual accuracy of the prediction model by emphasizing the impact of larger deviations.

Images