Heterogeneous resource scheduling cloud optimization method and system based on space-time task intensity evaluation

Abstract

The invention discloses a heterogeneous resource scheduling cloud optimization method based on spatio-temporal task intensity evaluation, which comprises the following steps of: firstly, performing calculation intensity evaluation on spatio-temporal tasks in combination with spatio-temporal task characteristics to obtain appropriate calculation resources required by the spatio-temporal tasks; secondly, preference information presented by the space-time task in the aspect of heterogeneous resource use is mined, an adaptive weight attention mechanism is designed, and a space-time task adaptive decision network (ST-ADN) is constructed; on the basis, a decision-making module in a reinforcement learning deep double-Q network (DDQN) is replaced by an ST-ADN network, and a space-time task adaptive scheduling model (DDQN-STADN) based on the DDQN network is constructed, so that the adaptive capability of the space-time task scheduling model in coping with different space-time task and heterogeneous resource combinations is remarkably improved; and meanwhile, the heterogeneous resource allocation efficiency is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of high-performance geographic computing, and in particular to a heterogeneous resource scheduling cloud optimization method and system based on spatiotemporal task intensity evaluation. Background Art

[0002] With the rapid development of technologies such as aerospace, sensors, wireless communications, and computers, data on the perception of geographic environment and human activities continues to accumulate, and spatiotemporal data presents obvious "big data" characteristics, and the demand for computing power has also increased. Based on the concepts of "on-demand allocation" and "shared resources", cloud environments have become an important technical means for spatiotemporal task processing and application deployment with their powerful computing power, elastic scalability, and efficient resource management. However, the current cloud computing heterogeneous resource scheduling still faces many challenges, such as unreasonable resource allocation, unbalanced load, and low resource utilization, which limits the great potential of cloud computing. Spatiotemporal tasks are a special type of computing tasks that involve multiple dimensions such as time and space. Their computing intensity will change dynamically with changes in time, space range, and data volume. At the same time, spatiotemporal tasks have obvious preference characteristics for the use of heterogeneous resources. Different types of spatiotemporal tasks rely on different hardware resource combinations (such as CPU, GPU, storage devices, etc.) to achieve high-performance computing, which further increases the complexity of heterogeneous resource scheduling. Traditional, general scheduling methods often cannot cope with complex and changeable spatiotemporal tasks, resulting in adverse consequences such as task execution delays and resource waste. Therefore, how to design an adaptive heterogeneous resource balancing scheduling strategy based on the spatiotemporal characteristics of tasks and resource status is a key link in intelligent spatiotemporal high-performance computing.

[0003] At present, most of the existing cloud computing resource scheduling methods focus on the overall resource allocation balance or single-dimensional resource optimization, and lack heterogeneous resource scheduling optimization strategies for spatiotemporal tasks. For example, many heuristic algorithms, meta-heuristic algorithms, random algorithms and machine learning algorithms are widely used to seek the optimal scheduling approximate solution. Algorithms with lower time complexity, such as the longest processing time (LPT), first come first serve (FCFS), round robin (RR), and greedy algorithms, are also often used to solve tasks with a single scenario. However, when facing the problem of resource heterogeneity, some existing methods cannot dynamically adjust resource weights according to the actual needs of the task, making it difficult to achieve efficient resource utilization and load balancing, and even more unable to meet the growing high-performance processing needs of spatiotemporal tasks. In recent years, reinforcement learning has shown certain application potential in the field of resource scheduling. It continuously optimizes decision-making strategies through interactive learning between intelligent agents and the environment to adapt to dynamically changing tasks and resource conditions. However, traditional reinforcement learning methods also have certain limitations when dealing with spatiotemporal task scheduling. Due to the complex spatiotemporal characteristics of tasks, existing reinforcement learning algorithms find it difficult to handle multi-dimensional spatiotemporal features, and do not fully incorporate characteristics such as the spatiotemporal task preference for heterogeneous resources. As a result, it is difficult to fully capture the complexity of spatiotemporal tasks when building decision models and make scientific resource scheduling decisions, which limits the full utilization of reinforcement learning advantages in spatiotemporal task resource scheduling. Summary of the invention

[0004] In view of the many limitations of existing related technologies in spatiotemporal task resource scheduling, the present invention provides a heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation. First, the computing intensity of spatiotemporal tasks is evaluated in combination with the characteristics of spatiotemporal tasks to obtain the appropriate computing resources required for spatiotemporal tasks; secondly, the preference information presented by spatiotemporal tasks in the use of heterogeneous resources is mined, an adaptive weighted attention mechanism is designed, and a spatiotemporal task adaptive decision network (ST-ADN) is constructed. On this basis, the decision module in the reinforcement learning deep double Q network (DDQN) network is replaced with the ST-ADN network, and a spatiotemporal task adaptive scheduling model (DDQN-STADN) based on the DDQN network is constructed, which significantly improves the adaptive ability of the spatiotemporal task scheduling model in dealing with different spatiotemporal tasks and heterogeneous resource combinations, and effectively improves the efficiency of heterogeneous resource allocation.

[0005] According to one aspect of the present invention, a heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation is provided, comprising: S1, analyze the computational intensity of spatiotemporal tasks and evaluate the resource requirements of the tasks; S2. Define the concept and expression model of spatiotemporal tasks, classify spatiotemporal tasks based on their preference for heterogeneous resources, and design a variety of scheduling optimization indicators; S3, design an adaptive weighted attention mechanism based on heterogeneous resource usage preference information and build a spatiotemporal task adaptive decision network; S4. Replace the decision module in the deep dual-Q network with the spatiotemporal task adaptive decision network designed in S3, and build a spatiotemporal task adaptive scheduling model based on the deep dual-Q network.

[0006] As a further technical solution, the method further includes: Train and optimize the spatiotemporal task adaptive scheduling model based on deep dual-Q network.

[0007] As a further technical solution, the S1 further includes: Analyze the characteristics of spatiotemporal tasks to obtain the time series length, time frequency, spatial resolution, spatial range, data scale and data imbalance characteristics of spatiotemporal data; The spatiotemporal task characteristics are converted into multi-dimensional feature vectors, and the raster data and vector data are represented as the input form of the convolutional neural network through the spatiotemporal feature fusion method; Based on different convolutional neural network models, the computational intensity of spatiotemporal tasks is modeled, and its performance is evaluated through the designed accuracy indicators and model time cost.

[0008] As a further technical solution, the S2 further includes: Define spatiotemporal tasks as a task flow consisting of a series of geographic algorithms that process information in the temporal and spatial dimensions; According to resource requirements, space-time tasks are divided into Montage space-time tasks, SIPHIT space-time tasks, Epigenomics space-time tasks and CUDA space-time tasks; Design an expression model for spatiotemporal tasks, using a directed acyclic graph to represent the spatiotemporal task flow, covering four dimensions: time range, spatial range, data source, and subject tasks; Design computing resource optimization indicators, including time resource indicators, computing resource occupancy indicators and system energy consumption indicators.

[0009] As a further technical solution, the S3 further includes: Design the input layer, combine the spatiotemporal task characteristics, resource pool status and spatiotemporal information to define the model input; Normalize the numerical features of the spatiotemporal tasks and embed the discrete features to form a unified multi-dimensional input vector; Design a four-layer hidden layer structure, use ReLU activation function, L2 regularization and Dropout mechanism to handle the complexity of spatiotemporal tasks; Design an adaptive weighted attention mechanism to dynamically adjust resource priorities through the self-attention mechanism; Design the output layer to output the resource selection probability distribution as the final decision module.

[0010] As a further technical solution, the S4 further includes: Design the network architecture of the deep dual Q network, including the main Q network and the target Q network, and use the Q value update mechanism to make decisions; Design a reward function that combines the time to complete the spatiotemporal task, resource utilization, and load balancing to form a comprehensive reward; The spatiotemporal task adaptive decision network is combined with the deep double-Q network, and the spatiotemporal task scheduling strategy is optimized based on the deep double-Q network.

[0011] As a further technical solution, the method further includes: Collect spatiotemporal task scheduling data, perform intensity estimation and annotation, normalize the data, and divide it into training set, validation set, and test set; Based on the Q-value update mechanism, the spatiotemporal task adaptive scheduling model is trained and the hyperparameters are adjusted to optimize the scheduling strategy. Perform model comparison and verify model performance.

[0012] According to one aspect of the present invention, a heterogeneous resource scheduling cloud optimization system based on spatiotemporal task intensity evaluation is provided, comprising: The computational intensity assessment module is used to analyze the computational intensity of spatiotemporal tasks and assess the resource requirements of the tasks; The spatiotemporal task resource scheduling module is used to define the concept and expression model of spatiotemporal tasks, classify spatiotemporal tasks based on their preference for heterogeneous resources, and design a variety of scheduling optimization indicators; The spatiotemporal task adaptive decision network module is used to design an adaptive weighted attention mechanism based on heterogeneous resource usage preference information and build a spatiotemporal task adaptive decision network; The reinforcement learning scheduling module is used to replace the decision module in the deep dual-Q network with the spatiotemporal task adaptive decision network designed by S3, and to build a spatiotemporal task adaptive scheduling model based on the deep dual-Q network.

[0013] According to one aspect of the present invention specification, there is provided a heterogeneous resource scheduling cloud optimization device based on spatiotemporal task intensity assessment, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity assessment.

[0014] According to one aspect of the present specification, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions enable the computer to execute the steps of the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention first evaluates the computing intensity of spatiotemporal tasks in combination with the characteristics of spatiotemporal tasks to obtain the appropriate computing resources required for spatiotemporal tasks; secondly, it mines the preference information presented by spatiotemporal tasks in the use of heterogeneous resources, designs an adaptive weighted attention mechanism, and constructs a spatiotemporal task adaptive decision network; on this basis, the decision module in the reinforcement learning deep dual-Q network is replaced with the ST-ADN network, and a spatiotemporal task adaptive scheduling model based on the DDQN network is constructed, which significantly improves the adaptability of the spatiotemporal task scheduling model in dealing with different spatiotemporal tasks and heterogeneous resource combinations, and effectively improves the efficiency of heterogeneous resource allocation.

[0016] 2. The present invention can replace traditional heuristic algorithms, meta-heuristic algorithms and other means to achieve high-performance scheduling of spatiotemporal task computing resources. The present invention has higher efficiency, accuracy and flexibility in complex spatiotemporal task scheduling and resource allocation, and is particularly suitable for large-scale and dynamically changing computing environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction is given below to the drawings used in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 It is an overall flow chart of an embodiment of the present invention.

[0019] Figure 2 This is the spatiotemporal task computing intensity evaluation process of an embodiment of the present invention.

[0020] Figure 3 It is a spatiotemporal task adaptive decision network diagram of an embodiment of the present invention.

[0021] Figure 4 This is a spatiotemporal task adaptive scheduling model structure based on DDQN in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the 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 in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention are arbitrarily combined with each other to form a new technical solution. This combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that this combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0023] The embodiment of the present invention provides a heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation, which is implemented on the open earth engine computing platform. Figure 1 The specific steps are as follows: Step 1: Spatiotemporal task computational intensity evaluation method.

[0024] like Figure 2 The specific implementation process of the embodiment is described as follows: Step 1.1: Analysis of spatiotemporal task characteristics. In the stage of spatiotemporal task characteristic analysis, the first step is to obtain the data type and basic information of the spatiotemporal task, classify the data involved in the spatiotemporal task according to the basic type, and clarify its basic data type, including raster data and vector data. Secondly, by analyzing the time series length, time frequency, spatial resolution, spatial range, data scale and data imbalance of spatiotemporal data, the spatiotemporal task is comprehensively modeled. In addition, the spatiotemporal task characteristics are further analyzed to clarify the task types, including data preprocessing tasks (such as data cleaning, format conversion), data analysis tasks (such as calculating the area of ​​geographic regions, counting the number of specific geographic elements) and data visualization tasks. Finally, the complexity is quantified according to the calculation steps of the spatiotemporal task. The calculation steps are more and the complexity is relatively high. In terms of computational memory consumption, if the data analysis task involves spatiotemporal operations on large-scale data, a large memory may be required to store intermediate results, quantify the complexity of each task, and determine the dependencies between tasks and data transmission requirements. Through this process, the dynamic changes and resource consumption characteristics of the spatiotemporal task during execution can be effectively captured, laying the foundation for subsequent computational intensity modeling.

[0025] Step 1.2: Representation of spatiotemporal task sample characteristics: In the stage of representing spatiotemporal task sample characteristics, firstly, multiple characteristics of spatiotemporal tasks (such as time dimension, space dimension, resource consumption, etc.) are converted into a unified multidimensional feature vector. Taking traffic flow analysis as an example, the time dimension can be converted into the characteristic values ​​of different time periods (such as peak hours and non-peak hours), the space dimension can be converted into the characteristic values ​​of different road sections, and the resource consumption dimension can be converted into the estimated value of the computing resources required to process the traffic data of the corresponding road section (such as CPU time, memory usage). These characteristic values ​​are combined into a multidimensional feature vector and a specific numerical encoding rule is adopted. Secondly, a multimodal fusion method of spatiotemporal characteristics is designed to analyze the neural network that can characterize the raster and vector data structures. Considering the characteristics of convolutional neural networks, for raster data, it is divided or resampled to a specified image size, such as 256×256, according to different granularity types. If the original image resolution is high, downsampling operations can be performed; if the resolution is low, upsampling operations such as interpolation can be performed to reach the target size. For vector data, such as road network vector data, it is divided and rasterized into a specified image size. For example, the road network is divided into a certain grid, and the information such as the road length and the number of road nodes in each grid is counted and converted into raster data as sample features for input into the convolutional neural network.

[0026] Step 1.3: Modeling the computational intensity of spatiotemporal tasks: In the modeling stage of the computational intensity of spatiotemporal tasks, we first select models and compare their performance. We input samples into different convolutional neural networks, such as LeNet, AlexNet, VGG, ResNet, DenseNet, etc., and compare the performance of different convolutional neural network models, including accuracy and performance indicators. Since it is a regression problem, the accuracy indicator can be evaluated by goodness of fit and mean square error. The performance indicator refers to the time cost of using the model, which is used to evaluate the impact of model prediction on the allocation of computing resources, and is mainly determined by the model size, network depth, and number of parameters. For the model size, LeNet is relatively small, while the VGG network may be large, and the model size affects its reading time. In terms of network depth and number of parameters, deep networks such as ResNet have more network layers and parameters, and their model prediction time is relatively long. By comparing the performance of different models in terms of accuracy and performance, we determine their applicability in different spatiotemporal task scenarios. Secondly, we perform balanced processing and computational intensity prediction on the data. The input data is evenly decomposed, and the large-scale image data is divided into multiple small blocks according to certain rules. The decomposition granularity is resampled or rasterized, and the decomposition granularity is represented as the input of the convolutional neural network. Finally, the optimal model is read and the computational intensity of the decomposition granularity is predicted, such as the computational resources required to process each image block (CPU core number, memory usage, processing time, etc.).

[0027] Step 2: Design of spatiotemporal task resource scheduling model.

[0028] Embodiment Specific implementation scheme is: Step 2.1: Spatiotemporal tasks: refers to a task flow consisting of a series of geographic algorithms involving information processing in time and space dimensions. These tasks are designed to operate and analyze data related to the geographical environment, human activities, etc. and with spatiotemporal attributes.

[0029] Step 2.2: Classification of spatiotemporal tasks. According to the resource scheduling requirements of spatiotemporal tasks, they are divided into: (1) Montage-type spatiotemporal tasks: Most of the spatiotemporal tasks in this category have low CPU and memory utilization, but high I / O utilization. Taking the mosaic task of processing large-scale remote sensing image data as an example, this type of task usually requires reading a large number of image files (high I / O utilization) and stitching them together according to certain geographic coordinate rules. (2) SIPHIT-type spatiotemporal tasks: Most of the spatiotemporal tasks in this category have low CPU utilization, but high memory utilization. Taking the spatiotemporal data spatial index construction task as an example, during the construction process, it is necessary to store a large amount of geographic spatial data structure information in memory (high memory utilization), while the CPU calculation is mainly for the construction and maintenance operations of the data structure, and the calculation intensity is relatively low. (3) Epigenomics spatiotemporal tasks: Most of the spatiotemporal tasks in this category require high CPU utilization and low I / O utilization. Taking global climate model simulation as an example, a large number of numerical calculations are required for atmospheric temperature, humidity, air pressure and other variables based on meteorological physics equations, while the input and output of data are relatively small (low I / O utilization). (4) CUDA spatiotemporal tasks: Most spatiotemporal tasks in this category require high GPU utilization. In the application of deep learning to image recognition tasks, a large amount of calculations rely on the parallel computing capabilities of the GPU.

[0030] Step 2.3: Spatiotemporal task expression model: Based on the analysis of the semantic connotation and dependency of spatiotemporal tasks, we abstract them into a directed acyclic graph and define the expression model of spatiotemporal tasks covering four dimensions: time range, spatial range, data source, and subject task. The expression model is:

[0031] in, It is the unique identifier of the spatiotemporal task flow graph; The time information of the spatiotemporal task flow graph records the start and end time of different data, the duration of each spatiotemporal task, the time interval between the dependencies of each spatiotemporal task, the periodicity of data, etc. The spatial information in the spatiotemporal task flow diagram records the actual geographical location and coverage of the spatiotemporal task; For data information in the spatiotemporal task flow diagram, record data type, data format, metadata, data access rights, etc.; It is the subject task information in the spatiotemporal task flow graph; is the set of directed edges between all spatiotemporal task nodes in the spatiotemporal task flow graph, where express The output is Each task node in the set can have multiple input nodes at the same time, and can also serve as the input of multiple task nodes at the same time, thereby representing the merging and branching relationship between task nodes.

[0032] Step 2.4: Calculate resource optimization indicators.

[0033] (1) Time resource indicators.

[0034] 1) Execution time: refers to the sum of the computation delay and transmission delay of the spatiotemporal task. represents the computational latency of spatiotemporal tasks, Indicates the transmission delay, express The execution time is calculated as follows:

[0035] 2) Transmission delay: the time it takes for spatiotemporal data to be transmitted between nodes, such as and There exists rely The transmission delay indicates that the data is transmitted from Execution node transfers to The time required to execute the node.

[0036] (2) Calculate resource usage indicators.

[0037] The CPU usage is calculated as follows:

[0038] 2) Memory usage, calculated as follows:

[0039] 3) Bandwidth occupancy rate, calculated as follows:

[0040] 4) GPU occupancy rate, calculated as follows: According to the classification of spatiotemporal tasks, adaptive weight allocation of computing resources is performed. The allocation formula is as follows: (3) System energy consumption index: time energy consumption and resource consumption The total energy.

[0041]

[0042] Step 3: Construction of spatiotemporal task adaptive decision network (ST-ADN).

[0043] like Figure 3 As shown, the specific implementation scheme of the embodiment is: Step 3.1: Input layer design. The input layer combines the characteristics of spatiotemporal tasks and resource usage information, so that the model comprehensively considers the dynamic changes of spatiotemporal tasks and resources. The main contents include: spatiotemporal task characteristics (CPU / GPU / memory / disk requirements, spatiotemporal task types, spatiotemporal task priorities, spatiotemporal task execution time), resource pool status (resource remaining amount of each node, load of each resource, resource priority), spatiotemporal information (time interval, data source, task execution status), etc. The input dimension is N; Specifically, for each spatiotemporal task, first collect its task feature information, such as the task may require higher GPU resources for graphics rendering, its task priority is medium, and the expected execution time is long. Monitor and record the resource pool status during task execution. The resource pool contains multiple computing nodes, each with different CPU, GPU, memory and disk capacity. In addition, obtain spatiotemporal information such as data information, time information, task execution status, etc. of the spatiotemporal task. Integrate these spatiotemporal task characteristics, resource pool status and spatiotemporal information to form input data, and its input dimension N is determined according to the specific amount of information collected.

[0044] Step 3.2: Spatiotemporal feature processing and embedding: First, normalize the collected numerical features (such as CPU requirements, GPU requirements, resource load, etc.) to make the scale of spatiotemporal feature information consistent. For the collected continuous numerical features, such as CPU requirements (0 - 100 cores), GPU requirements (0 - 5 GPUs), resource load (0 - 100%), etc., the normalization formula is used to unify their scales to the 0 - 1 range. Secondly, for discrete spatiotemporal features such as task types (preprocessing, data analysis, etc.), the Embedding layer is used to embed each spatiotemporal feature information into a low-dimensional vector space. Finally, the data information of the spatiotemporal task is separately encoded (time interval, spatial range, data source, etc.) to capture the impact of different spatiotemporal data on computing resources, helping the model to make appropriate scheduling strategies for spatiotemporal tasks. For time intervals, such as a week, it can be encoded as [0, 7], indicating the range of days from the start time to the end time; if the spatial range is a certain city area, it can be encoded according to its geographic coordinate range, such as [minimum longitude, maximum longitude, minimum latitude, maximum latitude]; if the data source has satellite images and geographic databases, it can be encoded as [1, 0] to indicate that there is a satellite image data source and no other specific type of data source (assuming the encoding rule is like this). These encoded information are spliced ​​with the normalized numerical features and the embedded discrete features into a large vector as the input of the spatiotemporal task adaptive decision network.

[0045] Step 3.3: Hidden layer design: First, according to the complexity of the spatiotemporal task, the hidden layer is designed as four layers based on the deep neural network (DNN). The first hidden layer introduces a fully connected layer, and the number of neurons is set to 256 to fully learn various feature information in the input vector. The ReLU activation function is used. For the input vector element x, the output of the ReLU function is max (0, x), which can help the network learn nonlinear features and avoid the gradient disappearance problem, so that the network can effectively update the weights during the back propagation process. At the same time, the Dropout mechanism is added to randomly set the output of some neurons to 0. Assuming that the Dropout rate is 0.2, 20% of the neurons are temporarily ignored during each training to prevent the network from overfitting and improve the generalization ability of the model. The second hidden layer continues to use a fully connected structure and reduces the number of neurons to 128. This layer also uses ReLU as the activation function to maintain the nonlinear learning ability of the network. The L2 regularization technique is applied, and its regularization term is half of the square of the L2 norm of the weight vector w, that is, 0.5 * ||w||^2. Adding this regularization term to the loss function can constrain the size of the weight, making the model more concise and avoiding overfitting caused by excessive weights. The third hidden layer further reduces the dimension, making the spatiotemporal features more compact and abstract. The ReLU function is selected to maintain the network's high efficiency and good gradient propagation characteristics during training. The fourth hidden layer sets the number of neurons to 10 based on the spatiotemporal task decision requirements and resource allocation options. This layer uses the Softmax activation function. For the input vector z, the output of the Softmax function is e^z_i / sum (e^z_j) , where i represents the i-th element in the vector and j represents the sum of all elements in the vector. This converts the output into the probability distribution of 10 allocation strategies.

[0046] Step 3.4: Design an adaptive weighted attention mechanism: Dynamically adjust resource priorities based on the characteristics of spatiotemporal tasks, and use the self-attention mechanism to calculate the relationship between tasks and resources. When spatiotemporal tasks require higher memory and CPU resources, the self-attention mechanism calculates the attention scores of the task to each resource based on these demand characteristics of the task. Based on these attention scores, appropriate weights are assigned to computing resources to ensure that the network can assign tasks to the most appropriate resource types and quantities.

[0047] Step 3.5: Output layer design: This layer is the final decision module of the spatiotemporal task adaptive decision network. The output layer will select the allocation strategy with the largest resource selection probability distribution from the probability distribution of the 10 allocation strategies output by the fourth hidden layer according to the task requirements and the status of the resource pool, thereby improving the execution efficiency of the entire spatiotemporal task and the efficiency of resource utilization.

[0048] Step 4: Construct a spatiotemporal task adaptive scheduling model (DDQN-STADN) based on reinforcement learning.

[0049] like Figure 4 As shown, the specific implementation scheme of the embodiment is: Step 4.1: DDQN network architecture design: First, the main Q network is designed. This network accepts the characteristics of the spatiotemporal task and the state of the resource pool as input, passes through several hidden layers, and outputs the Q value of each resource configuration. Secondly, the target Q network is designed. The target Q network is the same as the main Q network, but its weights are updated with delay. The output of the target Q network is used to calculate the target Q value. Finally, the Q value update design selects the action through the main Q network, and then uses the target Q network to calculate the target Q value, perform error feedback and update the weight of the main Q network.

[0050] Specifically, the spatiotemporal task features received by the main Q network include task type (such as image classification, change detection, etc.), data scale (number of image pixels, number of bands, etc.), task priority (urgency), and required CPU, GPU, memory, and disk resources. The resource pool status information covers the number of idle CPU cores, idle GPU memory, remaining memory capacity, available disk space, and current resource load of each node (such as CPU utilization, memory utilization, etc.) of each computing node, and outputs the Q value of each resource configuration after passing through several hidden layers. Secondly, the target Q network is designed. The target Q network is the same as the main Q network, but its weights are updated lazily. When calculating the target Q value, the target Q network outputs the target Q value based on the current state and the action selected by the main Q network, providing a basis for error calculation. Finally, in each training step, the main Q network first selects the resource configuration action with the maximum Q value based on the current spatiotemporal task characteristics and resource pool status. Let the current state be s, the action selected by the main Q network be a, the next state be s', the reward be r, and the target Q value calculation formula be:

[0051] Where γ is a discount factor, which usually ranges from 0 to 1 and is used to weigh current rewards against future rewards.

[0052] Step 4.2: Reward design: Based on the completion time of the spatiotemporal task, resource utilization, and resource load balance, the rewards for each indicator are set as follows.

[0053] 1) Completion time reward: For time-space tasks, the specified completion time is The actual completion time is , if the task can be completed within the specified time, a positive reward will be given, and the completion time reward function is designed as:

[0054] in It is a positive number, indicating the reward coefficient for early completion time.

[0055] 2) Resource Utilization Rate Reward: Assume that the resource utilization rate is , the target resource utilization is , the resource utilization reward is designed as:

[0056] in A negative number indicates the penalty coefficient for resource utilization that deviates from the target utilization.

[0057] 3) Resource load balancing reward: Assume there are n computing nodes in total, and the resource load of the i-th node is , average load resource for: The resource load balancing reward function is designed as: in It is a positive number, indicating the penalty coefficient for load imbalance. is the load threshold.

[0058] 4) Comprehensive reward function: .

[0059] Step 4.3: Adaptive decision network combined with DDQN: First, add the spatiotemporal task adaptive decision network to the expression decision process module of DDQN, and pass the spatiotemporal task characteristics and resource pool status to the trained spatiotemporal task adaptive decision network. Secondly, based on DDQN, the strategy optimization is performed. The main Q network selects the optimal resource allocation strategy under the current state according to the current spatiotemporal task characteristics and resource pool status, combined with the preliminary decision results of the adaptive decision network. Then, according to the execution of the spatiotemporal task and the reward feedback, the Q value function is updated. If the task completion time is short, the resource utilization rate is reasonable and the load is balanced after executing the strategy, a higher reward is obtained according to the reward function, and the Q value function will be updated in a direction that is favorable to the strategy, so that in similar states, the model is more inclined to choose this strategy. Finally, the feedback reward of the spatiotemporal task scheduling is calculated after each scheduling, and the weight of the Q network is updated. By repeating this process continuously, the Q network can gradually learn a resource allocation strategy that is more suitable for different spatiotemporal tasks and resource pool states, and improve the scheduling efficiency and performance of the entire system.

[0060] Step 5: Training and application of spatiotemporal task adaptive scheduling model.

[0061] Embodiment Specific implementation scheme is: Step 5.1: Create a dataset for spatiotemporal task scheduling and heterogeneous resource allocation in cloud environments: First, collect a large amount of spatiotemporal task scheduling data and evaluate the spatiotemporal task computing intensity. Label the data according to the evaluation results as a supervisory signal for the training model, and normalize the collected task features and resource status data so that the data ranges of different features are in a similar range. For the number of CPU cores required for the task, if its value range is between 1 and 100, it is converted to the range of 0 to 1 through a specific normalization formula. Then, randomly divide the sample data into training set, validation set, and test set according to 6:2:2. During the division process, ensure that each subset has a certain representativeness in terms of task type, resource requirement type, etc.

[0062] Step 5.2: Model training and optimization: First, train the spatiotemporal task adaptive scheduling model based on QQDN. The main Q network in the model receives the normalized spatiotemporal task features and resource pool status as input, and outputs the Q value of each resource configuration after processing through multiple hidden layers. During the training process, the spatiotemporal task scheduling is performed according to the resource configuration action selected by the main Q network in the current state, and the corresponding rewards are obtained. The scheduling strategy is continuously improved using the Q value function update mechanism. According to the error between the target Q value and the Q value output by the main Q network, the weight of the main Q network is updated through the back propagation algorithm. Secondly, set a suitable discount factor (γ) to control the impact of future rewards, and a suitable learning rate to ensure the stability of the training process. Secondly, during the training process, the hyperparameters of the DDQN model, such as learning rate, batch size, discount factor, etc., are continuously adjusted to keep the model in the optimal resource allocation state.

[0063] Step 5.3: Model evaluation and verification: After training, the performance of the model is evaluated by comparing the experimental results of different scheduling strategies. The key evaluation indicators include task completion time, resource utilization, resource load balancing, etc. The spatiotemporal task adaptive scheduling model is compared with traditional scheduling strategies (such as first-come-first-served, shortest job first, etc.), as well as scheduling strategies based on specific algorithms proposed by other scholars, such as scheduling strategies based on genetic algorithms. Through these comparisons, the advantages and innovations of this model are verified, proving that it has better performance and application value in the field of geographic high-performance computing.

[0064] The implementation basis of each embodiment of the present invention is implemented by programmed processing through a device with a processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this reality, on the basis of the above embodiments, an embodiment of the present invention provides a heterogeneous resource scheduling cloud optimization system based on spatiotemporal task intensity evaluation, which is used to execute the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation in the above method embodiment.

[0065] The system includes: a computing intensity assessment module, which is used to analyze the computing intensity of spatiotemporal tasks and evaluate the resource requirements of tasks; a spatiotemporal task resource scheduling module, which is used to define the concept and expression model of spatiotemporal tasks, classify spatiotemporal tasks in view of the usage preference characteristics of spatiotemporal tasks for heterogeneous resources, and design a variety of scheduling optimization indicators; a spatiotemporal task adaptive decision network module, which is used to design an adaptive weighted attention mechanism based on the usage preference information of heterogeneous resources and construct a spatiotemporal task adaptive decision network; a reinforcement learning scheduling module, which is used to replace the decision module in the deep dual-Q network with the spatiotemporal task adaptive decision network designed by S3, and construct a spatiotemporal task adaptive scheduling model based on the deep dual-Q network.

[0066] The heterogeneous resource scheduling cloud optimization system based on spatiotemporal task intensity evaluation provided by the embodiment of the present invention is aimed at the scenario where the existing related technologies have many limitations in the spatiotemporal task resource scheduling. It adopts the aforementioned several modules, firstly, combines the spatiotemporal task characteristics to evaluate the computing intensity of the spatiotemporal task, and obtains the appropriate computing resources required for the spatiotemporal task; secondly, it mines the preference information presented by the spatiotemporal task in the use of heterogeneous resources, designs an adaptive weighted attention mechanism, and constructs a spatiotemporal task adaptive decision network. On this basis, the decision module in the reinforcement learning deep dual Q network is replaced with the ST-ADN network, and a spatiotemporal task adaptive scheduling model based on the DDQN network is constructed, which significantly improves the adaptive ability of the spatiotemporal task scheduling model in dealing with different spatiotemporal tasks and heterogeneous resource combinations, and effectively improves the efficiency of heterogeneous resource allocation.

[0067] It should be noted that the system embodiments provided by the present invention are not only used to implement the methods in the above-mentioned method embodiments, but also used to implement the methods in other method embodiments provided by the present invention. The only difference lies in the setting of corresponding functional modules, and the principles thereof are basically the same as the principles of the above-mentioned system embodiments provided by the present invention. As long as technical personnel in this field refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned system embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and on the premise of ensuring the practicality of the technical solutions, they will improve the modules in the above-mentioned system embodiments to obtain corresponding system class embodiments for implementing the methods in other method class embodiments.

[0068] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention also provides a heterogeneous resource scheduling cloud optimization device based on spatiotemporal task intensity assessment, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity assessment.

[0069] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable the computer to execute the steps of the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity assessment.

[0070] In summary, the present invention discloses a heterogeneous resource scheduling cloud optimization method based on spatio-temporal task intensity evaluation, and the method includes: S1, spatio-temporal task computing intensity evaluation, accurately estimating the computing resources of spatio-temporal tasks; S2, defining the concept and expression model of spatio-temporal tasks, classifying spatio-temporal tasks in view of the usage preference characteristics of spatio-temporal tasks for heterogeneous resources, and designing a variety of scheduling optimization indicators; S3, fully integrating information such as spatio-temporal task characteristics and resource pool status, designing an adaptive weighted attention mechanism based on heterogeneous resource usage preference information, and constructing an adaptive decision-making network for spatio-temporal tasks (Adaptive decision-making network for spatio-temporal tasks, ST-ADN); S4, taking into account the complexity, dynamics and resource heterogeneity of spatio-temporal tasks, the decision-making function of the reinforcement learning deep double Q network (double deep Q network, DDQN) is transformed. In this process, the decision module in the DDQN network is replaced with the spatiotemporal task adaptive decision network designed by S3, so as to construct a spatiotemporal task adaptive scheduling model (DDQN-STADN) based on the DDQN network; S5, through reasonable model training and optimization, the model learns efficient spatiotemporal task scheduling strategies, and the performance advantages of the model are verified through comprehensive evaluation and comparison. This invention can effectively improve the efficiency of spatiotemporal task processing, optimize resource utilization and load balancing, and is suitable for the field of high-performance geographic computing.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation, characterized in that: include: S1, analyze the computational intensity of spatiotemporal tasks and evaluate the resource requirements of the tasks; S2. Define the concept and expression model of spatiotemporal tasks, classify spatiotemporal tasks based on their preference for heterogeneous resources, and design a variety of scheduling optimization indicators; S3, design an adaptive weighted attention mechanism based on heterogeneous resource usage preference information and build a spatiotemporal task adaptive decision network; S4. Replace the decision module in the deep dual-Q network with the spatiotemporal task adaptive decision network designed in S3, and build a spatiotemporal task adaptive scheduling model based on the deep dual-Q network.

2. The heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation according to claim 1 is characterized in that: The method further comprises: Train and optimize the spatiotemporal task adaptive scheduling model based on deep dual-Q network.

3. The heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation according to claim 1 is characterized in that: The S1 further comprises: Analyze the characteristics of spatiotemporal tasks to obtain the time series length, time frequency, spatial resolution, spatial range, data scale and data imbalance characteristics of spatiotemporal data; The spatiotemporal task characteristics are converted into multi-dimensional feature vectors, and the raster data and vector data are represented as the input form of the convolutional neural network through the spatiotemporal feature fusion method; Based on different convolutional neural network models, the computational intensity of spatiotemporal tasks is modeled, and its performance is evaluated through the designed accuracy indicators and model time cost.

4. The heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation according to claim 1 is characterized in that: The S2 further includes: Define spatiotemporal tasks as a task flow consisting of a series of geographic algorithms that process information in the temporal and spatial dimensions; According to resource requirements, space-time tasks are divided into Montage space-time tasks, SIPHIT space-time tasks, Epigenomics space-time tasks and CUDA space-time tasks; Design an expression model for spatiotemporal tasks, using a directed acyclic graph to represent the spatiotemporal task flow, covering four dimensions: time range, spatial range, data source, and subject tasks; Design computing resource optimization indicators, including time resource indicators, computing resource occupancy indicators and system energy consumption indicators.

5. The heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation according to claim 1 is characterized in that: The S3 further includes: Design the input layer, combine the spatiotemporal task characteristics, resource pool status and spatiotemporal information to define the model input; Normalize the numerical features of the spatiotemporal tasks and embed the discrete features to form a unified multi-dimensional input vector; Design a four-layer hidden layer structure, use ReLU activation function, L2 regularization and Dropout mechanism to handle the complexity of spatiotemporal tasks; Design an adaptive weighted attention mechanism to dynamically adjust resource priorities through the self-attention mechanism; Design the output layer to output the resource selection probability distribution as the final decision module.

6. The heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation according to claim 1 is characterized in that: The S4 further comprises: Design the network architecture of the deep dual Q network, including the main Q network and the target Q network, and use the Q value update mechanism to make decisions; Design a reward function that combines the time to complete the spatiotemporal task, resource utilization, and load balancing to form a comprehensive reward; The spatiotemporal task adaptive decision network is combined with the deep double-Q network, and the spatiotemporal task scheduling strategy is optimized based on the deep double-Q network.

7. The heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation according to claim 2 is characterized in that: The method further comprises: Collect spatiotemporal task scheduling data, perform intensity estimation and annotation, normalize the data, and divide it into training set, validation set, and test set; Based on the Q-value update mechanism, the spatiotemporal task adaptive scheduling model is trained and the hyperparameters are adjusted to optimize the scheduling strategy. Perform model comparison and verify model performance.

8. A heterogeneous resource scheduling cloud optimization system based on spatiotemporal task intensity evaluation, characterized in that: include: The computational intensity assessment module is used to analyze the computational intensity of spatiotemporal tasks and assess the resource requirements of the tasks; The spatiotemporal task resource scheduling module is used to define the concept and expression model of spatiotemporal tasks, classify spatiotemporal tasks based on their preference for heterogeneous resources, and design a variety of scheduling optimization indicators; The spatiotemporal task adaptive decision network module is used to design an adaptive weighted attention mechanism based on heterogeneous resource usage preference information and build a spatiotemporal task adaptive decision network; The reinforcement learning scheduling module is used to replace the decision module in the deep dual-Q network with the spatiotemporal task adaptive decision network designed by S3, and to build a spatiotemporal task adaptive scheduling model based on the deep dual-Q network.

9. A heterogeneous resource scheduling cloud optimization device based on spatiotemporal task intensity evaluation, characterized in that: It includes a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, which enable the computer to execute the steps of the heterogeneous resource scheduling cloud optimization method based on spatiotemporal task intensity evaluation as described in any one of claims 1 to 7.

Images