Method and system for efficiently executing computing tasks in multi-mode intelligent computing network

Abstract

The invention belongs to the technical field of resource elastic scaling, and discloses an efficient execution method and system for a computing task in a multi-mode intelligent computing network, and the method comprises the steps: collecting the resource use time sequence data of the task in real time, wherein the resource use time sequence data comprises a current time step length and T-1 historical time step lengths adjacent to the current time step length; inputting the time sequence data into a trained resource demand prediction model to obtain resource use time sequence data in L prediction time steps; wherein L is the time window length of the task load stability determined by the resource volatility and modal characteristics of the task; determining the optimal resource configuration required by the task in the L prediction time steps by using the resource use time sequence data in the L prediction time steps; and comparing the optimal resource configuration with a preset elastic scaling strategy rule, and determining whether resource elastic scaling needs to be carried out or not. Furthermore, the invention further provides a resource recommendation mode for a new task, and the resource allocation accuracy and stability and the response speed of the system can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of resource elastic scaling, and more specifically, relates to a method and system for efficiently executing computing tasks in a multi-modal intelligent connection computing network. Background Art

[0002] To meet the intelligent connection computing requirements, it is necessary to build a multi-modal intelligent connection computing network architecture that can meet the application requirements of various vertical industries. Among them, when dealing with complex and changeable vertical industry intelligent connection computing service scenarios, how to improve the resource utilization rate as much as possible on the premise of providing continuous and reliable resource guarantee capabilities for intelligent connection computing tasks, and ensure efficient and transparent execution within the network is a key issue. A flexible and reliable resource elastic scaling method is an effective way to solve this problem.

[0003] Through resource elastic scaling, the quality of service can be guaranteed, the performance requirements of intelligent connection computing tasks can be met, and at the same time, the resource utilization rate can be improved, and the efficient and transparent execution of intelligent connection computing tasks can be ensured. Usually, resource elastic scaling can be divided into two types: horizontal scaling and vertical scaling. Horizontal scaling expands or shrinks by increasing or decreasing the number of task execution instances, and vertical scaling expands or shrinks by increasing or decreasing the resource capacity of a single instance.

[0004] With the gradual progress of cloud computing technology, most resource orchestration systems of architectures already have basic elastic scaling capabilities. For example, Kubernetes provides three elastic scaling methods: Horizontal Pod Autoscaler that dynamically adjusts the number of Pod replicas based on a preset resource usage threshold, Vertical Pod Autoscaler that dynamically adjusts the Pod resource request and limit based on the actual resource usage, and Cluster Autoscaler that adjusts the number of cluster nodes based on the Pod resource requirements. In addition, the patent with the publication number CN115981863A configures a scaling policy model according to the virtual machine service characteristics, judges whether the expansion or contraction policy threshold is met based on the recent performance data results of the virtual machine. If it is met, a scaling process confirmation request is initiated. After the request is passed, an elastic scaling command is sent to the resource management platform to execute elastic contraction. However, due to the latency of resource elastic scaling execution by the resource orchestration system and the need to wait for the request review to pass before elastic scaling, there is an elastic hysteresis problem in the process of this reactive trigger of resource elastic scaling based on the instance runtime load and the preset policy threshold, which cannot respond to the resource requirements during the load peak caused by sudden traffic in a timely manner, reducing the accuracy of resource allocation and the response speed of the system.

[0005] The patent with the publication number CN116610454A proposes a hybrid cloud resource elastic scaling system and operation method based on the MADDPG algorithm, which is used for horizontal and vertical hybrid elastic scaling of cloud resources. It can respond to load changes in a short time based on cluster load monitoring data and provide sufficient resources to meet application requirements. However, in a multi-modal intelligent computing network environment, the large-scale task data and resource environment data will increase the data processing difficulty and computing overhead of this method. More importantly, when the task load fluctuates frequently, this method frequently adjusts resource allocation for resource elastic scaling, increasing the system operation cost, and the execution task experiences multiple resource changes in a short time, which will affect the stability of task execution. Summary of the Invention

[0006] In view of the above defects or improvement requirements of the prior art, the present invention provides an efficient execution method and system for computing tasks in a multi-modal intelligent computing network, aiming to improve the accuracy of resource allocation, the response speed of the system, and the stability of resource allocation.

[0007] To achieve the above object, the present invention provides an efficient execution method for computing tasks in a multi-modal intelligent computing network, including:

[0008] Real-time collect the resource usage time series data of the task including the current time step and the adjacent T-1 historical time steps, and input it into the trained resource demand prediction model to obtain the resource usage time series data within L prediction time steps; where L is the time window length of task load stability determined by the resource volatility and modal characteristics of the task, and the resource volatility is determined based on the historical resource usage time series data of the task; the task is a computing task in a multi-modal intelligent computing network;

[0009] Use the resource usage time series data within the L prediction time steps to determine the optimal resource configuration required by the task within the L prediction time steps; compare the optimal resource configuration with the preset elastic scaling policy rules, and when resource elastic scaling is required, initiate an elastic scaling request to the downstream resource orchestration system to perform coordinated scheduling of horizontal and vertical scaling.

[0010] Further, it further includes:

[0011] Record the optimal resource configuration and the corresponding task modal characteristics in the task record table;

[0012] When a new task arrives, obtain the modal characteristics of the new task according to the request information of the new task, match the modal characteristics of the new task with the modal characteristics of each task in the task record table, and based on the similarity matching result, recommend the optimal resource configuration in the task record table as the resource configuration specification of the new task Among them, the task record table includes the optimal resource allocation required for the task within L prediction time steps and the corresponding task modal features;

[0013] Calculate the resource allocation specification recommended for the new task The difference coefficient from the pre-defined configuration specification R of the new task When the difference coefficient α does not exceed the preset threshold, adopt the recommended resource allocation specification As the final recommended result of the new task resource allocation, otherwise, adopt the pre-defined configuration specification R of the new task as the final recommended result;

[0014] Feed the final recommended result back to the task scheduler of the downstream resource orchestration system for resource scheduling, and obtain the historical resource usage time series data of the new task, which is used as the training sample set to train the resource demand prediction model corresponding to the new task. Use the trained resource demand prediction model corresponding to the new task to predict the resource demand and perform elastic scaling based on the predicted resource demand.

[0015] Furthermore, perform similarity matching between the modal features of the new task and the modal features of each task in the task record table. Based on the similarity matching result, recommend the optimal resource allocation in the task record table as the resource allocation specification of the new task Including:

[0016] Calculate the weighted cosine similarity between the modal features of the new task and the modal features of the tasks in the task list with the same modal category as the new task, and use the weighted cosine similarity as the similarity score;

[0017] Cluster the modal features and corresponding resource volatility of the tasks in the task list whose similarity scores exceed the threshold τ, and take the optimal resource allocation of the task closest to the center of the largest cluster as the resource allocation specification recommended for the new task

[0018] Furthermore, the calculation method of the resource volatility is:

[0019]

[0020] Among them, CV is the resource volatility, and σ and μ represent the standard deviation and mean of the historical resource usage time series data of the task respectively.

[0021] Furthermore, the trained resource demand prediction model is obtained by collecting historical resource usage time series data of tasks in real time as a training sample set and training the resource demand prediction model with the training sample set; wherein, the training sample includes resource usage time series data of T historical time steps and resource usage time series data of L prediction time steps, and the resource usage time series data of one time step is the data obtained by splicing the resource usage data of multiple sampling time intervals ΔT in time series and then compressing the data.

[0022] Furthermore, the resource types of the tasks include: CPU usage, memory usage, network bandwidth usage; data compression of each task resource includes: using the 95th percentile to represent the CPU usage within one time step, and using the peak value to represent the memory usage and network bandwidth usage within one time step;

[0023] Determining the optimal resource configuration required by the task within the L prediction time steps using the resource usage time series data within the L prediction time steps includes:

[0024] The optimal resource configuration R of CPU CPU = P 95 (Y CPU )×(1 + λ CPU ), where λ CPU is the reservation coefficient of CPU, Y CPU represents the CPU usage within the L prediction time steps, and P 95 (·) represents the operation of taking the 95th percentile;

[0025] The optimal resource configuration R of memory Mem = Max(Y Mem )×(1 + λ Mem ), where λ Mem is the reservation coefficient of memory, Y Mem represents the memory usage within the L prediction time steps, and Max(·) represents the operation of finding the peak value;

[0026] For the optimal resource configuration R of bandwidth BW = Max(Y BW )×(1 + λ BW ), where λ BW is the reservation coefficient of bandwidth resource, and Y BW represents the bandwidth usage within the L prediction time steps;

[0027] Among them, the reservation coefficients of various resources decrease as the overall resource load of the multi-modal intelligent connection computing network increases.

[0028] Furthermore, the resource demand prediction model includes a Prophet network and an LSTM;

[0029] The predicted value of the Prophet network is:

[0030]

[0031] where g(t) is the trend term of the training sample, s(t) is the seasonal term of the training sample, and h(t) is the holiday effect of the training sample;

[0032] LSTM is used to model the residual r that the Prophet network fails to capture t to obtain the predicted residual where y t is the input training sample;

[0033] The predicted value of the resource demand prediction model is:

[0034] The present invention also provides an efficient execution system for computing tasks in a multi-modal intelligent connection computing network, including a computer-readable storage medium and a processor;

[0035] The computer-readable storage medium is used to store executable instructions;

[0036] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the efficient execution method for computing tasks in the multi-modal intelligent connection computing network described in any one of the above.

[0037] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the efficient execution method for computing tasks in the multi-modal intelligent connection computing network described in any one of the above.

[0038] The present invention also provides a computer program product, including a computer program, and when the computer program runs on a computer, it causes the computer to execute the efficient execution method for computing tasks in the multi-modal intelligent connection computing network described in any one of the above.

[0039] Generally speaking, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

[0040] (1) The elastic scaling policy in the method for efficient execution of computing tasks in the multi-modal intelligent connection computing network of the present invention is a prediction-based proactive elastic scaling policy. Based on historical resource data, the time series prediction algorithm is used to accurately predict the changing trend of future resource requirements, and the resource elastic scaling operation is actively triggered based on the prediction results to adjust the instance resource allocation in advance, effectively avoiding the elastic lag problem existing in the process of responsive triggering of resource elastic scaling, and significantly improving the accuracy of resource allocation and the response speed of the system.

[0041] The proactive elastic scaling policy of the present invention adopts a stable resource elastic scaling mechanism for a long time period (L prediction time steps). Specifically, by comprehensively analyzing historical resource usage data, the volatility of the business load is effectively identified, and combined with the modal characteristics of the computing tasks, the time window length of the stability of each computing task load is determined, and the time window length of the stability of the computing task load is used as the prediction window. Taking the prediction results within the time window length of the stability of each computing task load as a unit, a resource specification adjustment is performed once. This combination of long-term prediction models accurately plans the resource allocation strategy. Compared with the short-term elastic scaling strategy, it can reduce the system jitter caused by frequent resource adjustment and balance the stability of resource allocation and the demand for dynamic adjustment.

[0042] (2) Further, the resource specification configuration recommendation mechanism based on task modal feature matching of the present invention records task feature information during the computing task distribution process and records the resource allocation results of the elastic scaling policy. When a new task arrives, based on the task modal feature matching and the recorded resource allocation results of the elastic scaling policy, the optimal resource configuration is recommended for new tasks with the same feature modality, avoiding resource waste caused by reserved resources in the instance specification configuration and effectively improving the initial resource utilization rate of the instance.

[0043] (3) Preferably, the resource usage time series data for one time step in the present invention is the data obtained by splicing the resource usage data of multiple sampling time intervals ΔT according to the time series and then compressing the data. Based on different resource metric types, different characteristic values are used to represent the characteristics of the resource metric within one time step. For example, the 95th percentile is used to represent the CPU resource usage within one time step, and the peak value is used to represent the memory resource usage and network bandwidth resource usage within one time step, solving the problems of complex and changeable computing tasks, huge historical resource data collected, high data processing complexity, and excessive computing overhead in large-scale environments such as cloud-network integration in multi-modal intelligent connection computing networks.

[0044] (4) Preferably, a combined model of Prophet model and LSTM is adopted. The Prophet model is used to capture the long-term trend, periodicity, and holiday effects of resource requirements, and the LSTM is used to model the prediction residuals of the Prophet model to capture non-linear fluctuations such as sudden traffic, etc., improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of an efficient execution method for computing tasks in a multi-modal intelligent connection computing network in an embodiment of the present invention;

[0046] Figure 2 Flowchart of an efficient execution method for computing tasks in a multi-modal intelligent connection computing network in an embodiment of the present invention;

[0047] Figure 3 Flowchart of building a combined prediction model in an embodiment of the present invention;

[0048] Figure 4 System architecture diagram of an efficient execution method for computing tasks in a multi-modal intelligent connection computing network in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0050] Embodiment 1

[0051] As Figure 1 , Figure 2 shown, an embodiment of the present invention provides an efficient execution method for computing tasks in a multi-modal intelligent connection computing network, mainly including:

[0052] S1. Real-time collect the historical resource usage time series data of computing tasks in the multi-modal intelligent connection computing network to construct a data set for predicting task resource requirements; wherein, the training samples in the data set include the resource usage time series data of T historical time steps and the resource usage time series data of L prediction time steps.

[0053] S2. Use the training samples in the dataset to train the resource demand prediction model. The trained model is used to infer and predict the resource demand within the next L prediction time steps for the computing task. Here, L is the length of the time window during which the computing task load is stable. The resource volatility is calculated based on the historical resource usage time series data of the computing task, and the length of the time window during which the computing task load is stable is determined by combining the modal characteristics of the computing task.

[0054] S3. Input the resource usage time series data of the computing task for T time steps into the trained resource demand prediction model to obtain the resource usage time series data within the next L prediction time steps. Here, the T time steps include the current time step and the T - 1 historical time steps connected to it.

[0055] S4. Calculate the optimal resource configuration required for the computing task within the L prediction time steps using the resource usage time series data within the L prediction time steps. Compare the optimal resource configuration with the preset elastic scaling policy rules to determine whether resource elastic scaling is required. If so, initiate an elastic scaling request to the downstream resource orchestration system, and through the API, call the resource orchestration system to perform coordinated scheduling of horizontal and vertical scaling to ensure resource reliability and reduce resource waste. When performing coordinated scheduling of horizontal and vertical scaling, prefer horizontal expansion to handle sudden loads, and gradually reduce resources after the load drops to ensure a smooth transition of resource allocation and avoid task interruption caused by frequent jitters.

[0056] In S1, the historical resource usage time series data of each computing task in the multi-modal intelligent connection computing network is collected in real time to construct a dataset for task resource demand prediction, including:

[0057] Use the network-aware resource monitoring tool to collect the computing, storage, and network resource usage data of the intelligent connection computing task at fixed time intervals in real time and record the modal feature information of the task. Complete the work of data cleaning and feature extraction, and store the extracted features and time series data in the time series database to form a resource prediction dataset. Specifically, at every interval of ΔT time, the time series data of each resource usage of the computing task is collected in real time, mainly including: the actual CPU usage CPU usage 、the actual memory usage Mem usage 、the actual network bandwidth usage BW usage and other computing, storage, and network resource metrics.

[0058] Concatenate the time series data collected at N time interval points in time series as the time series data of a large time window (i.e., the above-mentioned one time step).

[0059] Considering that the resource elastic scaling method relies on collecting historical resource data and making analysis decisions, in a large-scale environment such as a multi-modal intelligent connection computing network where cloud and network are integrated, the computing tasks are complex and changeable, the collected historical resource data is huge, the data processing complexity is relatively high, and the computing overhead is too large. In the embodiments of the present invention, before constructing a data set for predicting task resource requirements based on the obtained historical resource usage time series data, feature extraction is also included for the obtained historical resource usage time series data to achieve data compression, specifically including:

[0060] Based on different resource metric types, different feature values are used to represent the features of the resource metric within one time step; in the embodiments of the present invention, the 95th percentile is used to represent the CPU resource usage within one time step, and the peak value is used to represent the memory resource usage and network bandwidth resource usage within one time step.

[0061] The historical resource usage time series data represented by feature values and the modal features of the task are stored in a time series database, thereby forming a resource prediction data set. That is, each training sample in the data set is the resource usage time series data of each time step represented by feature values.

[0062] In S2, preferably, a combined model of Prophet model and LSTM (Long Short-Term Memory network) is adopted. The Prophet model is used to capture the long-term trend, periodicity and holiday effect of resource requirements, and the LSTM is used to model the prediction residuals of the Prophet model to capture non-linear fluctuations, such as burst traffic. After completing the training of the combined model, the trained model is deployed as an online prediction service to predict the resource requirement sequence of the future time window and support incremental learning to adapt to the change of resource requirement patterns.

[0063] Specifically, the input sample of the model is the time series data y t , including the resource usage time series data of T historical time steps including the current time step t, fitting the Prophet model with historical data to obtain the trend, seasonality and holiday effect, and obtaining the predicted value of the Prophet model

[0064]

[0065] Among them, g(t) is the trend term, represented by a logistic growth model, s(t) is the seasonal term, represented by a Fourier series, h(t) is the holiday effect, and the residual r that the Prophet fails to capture can be calculated after Prophet prediction t :

[0066]

[0067] In the embodiments of the present invention, the residual r is modeled by LSTM t and the residual sequence R of the past T time steps t =[r t-T ,r t-T+1 ,…,r t-1 is used as the input to obtain the residual prediction values of the future L time steps The predicted values of the Prophet model and the residual prediction values of LSTM are added to obtain the final predicted value

[0068]

[0069] The modeling flow chart of the combined prediction model in the embodiments of the present invention is as shown in Figure 3 In the embodiments of the present invention, the resource prediction data set is divided into a training set, a validation set, and a test set according to 7:2:1 to train the combined model. The trained model is deployed as an online prediction service, and supports real-time receiving of the latest monitoring data and incrementally updating the model parameters to adapt to the changes in the resource demand pattern.

[0070] In other embodiments, the ARIMA model can also be selected to replace the Prophet model, and the GRU can be selected to replace the LSTM network.

[0071] Preferably, in S2, based on the historical resource usage time series data of each computing task, the coefficient of variation (CV) is used to quantitatively calculate the resource volatility of the computing task, and the class resource volatility of the computing task is obtained where σ and μ respectively represent the standard deviation and the mean of the historical resource usage time series data of the computing task.

[0072] In the embodiments of the present invention, according to the modal characteristics of the task such as modal category, SLA constraint, priority level, etc., and the resource volatility of the task, the time window length L for different task loads to be stable is determined. Specifically, for low-fluctuation tasks (determined by expert experience according to resource volatility and modal characteristics), that is, tasks with relatively stable resource loads for a long time, a long time window is used for resource demand adjustment to reduce the frequency of task resource adjustment. For high-fluctuation tasks, the task load changes frequently, and maintaining the same resource configuration for a long time does not meet the requirements of task load changes. A short time window is used for resource demand adjustment to better dynamically adjust resources.

[0073] In S2, based on the request information of the computing task, record the modal characteristics of the computing task, mainly including the task modal category, task SLA (Service Level Agreement) constraints, input data scale, priority, etc.

[0074] In S4, use the resource usage time series data within the next L prediction time steps to calculate the optimal resource configuration required for the computing task within the next L prediction time steps, including:

[0075] For the optimal resource configuration R of the CPU CPU = P 95 (Y CPU ) × (1 + λ CPU ), where λ CPU is the reservation coefficient of CPU resources, Y CPU represents the CPU usage within the next L prediction time steps, and the 95th percentile is used to represent the CPU resource usage in L prediction time steps, P 95 (·) represents the 95th percentile operation;

[0076] For the optimal resource configuration R of memory Mem = Max(Y Mem ) × (1 + λ Mem ), where λ Mem is the reservation coefficient of memory resources, Y Mem represents the memory resource usage within the next L prediction time steps, and Max(·) represents the peak operation;

[0077] For the optimal resource configuration R of bandwidth BW = Max(Y BW ) × (1 + λ BW ), where λ BW is the reservation coefficient of bandwidth resources, Y BW represents the bandwidth resource usage within the next L prediction time steps, and the peak value is used; the reservation coefficients of each resource decrease as the overall resource load of the multi-modal intelligent connection computing network increases.

[0078] Considering that during the process of instance resource configuration, more resource buffers are usually reserved to cope with possible high-load situations. This configuration method also brings a problem, that is, when a new task arrives, the initial resource utilization rate of the instance is often low. This is because the reserved resources are not fully utilized at the initial stage of the task, resulting in resource idleness and waste, affecting the overall resource utilization efficiency. Therefore, the method in the embodiments of the present invention further includes step S5:

[0079] S51. Bind the optimal resource allocation required for the computing tasks within L prediction time steps (the recommended value of the resource allocation specification for the next time window of the task) to the task modal features, and record the task modal features, optimal resource allocation, resource volatility, and the current resource allocation of the task in the task record table. In the embodiments of the present invention, the key fields of the task record table are shown in Table 1. The disk configuration of the task is an initially set value, which is obtained by counting the actual usage.

[0080] Table 1 Partial key fields of the task record table

[0081]

[0082]

[0083] S52. When a new task (a task that has not been executed) arrives, record the modal features of the new task according to the request information of the new task, match the modal features of the new task with the modal features of each task in the task record table, and recommend a suitable resource allocation specification for the new task based on the similarity matching result. Specifically, it includes:

[0084] Calculate the weighted cosine similarity between the modal features of the new task and the modal features of the tasks in the task list with the same modal category as the new task, and use the weighted cosine similarity as the similarity score.

[0085] Cluster the modal features and resource volatility of the tasks in the task list whose similarity scores exceed the threshold τ (generally, 0.85 can be taken), and take the optimal resource allocation of the task closest to the center of the largest cluster as the resource allocation specification recommended for the new task. In the embodiments of the present invention, the maximum cluster center value resource allocation data of the tasks in these same-category modalities is selected by the K-means clustering method as the resource allocation specification recommended for the new task.

[0086] S53. Calculate the difference coefficient between the resource allocation specification recommended for the new task and the pre-defined configuration specification R of the new task. When the difference coefficient α does not exceed the preset threshold, adopt the recommended resource allocation specification as the final recommended result. Otherwise, adopt the pre-defined configuration specification R of the new task as the final recommended result. In the embodiments of the present invention, when α ≤ 0.5, the recommended resource allocation specification is adopted as the final recommended result. When α > 0.5, the new task-defined configuration specification R is adopted as the final recommended result.

[0087] S54. Feed the final recommendation result (resource allocation optimization result) back to the task scheduler of the downstream resource orchestration system for resource scheduling, that is, the process of allocating resources to new tasks based on the resource allocation optimization result. After a period of time, based on the historical resource usage time series data of the new tasks, a dataset for predicting the resource requirements of the new tasks is constructed, and the resource requirement prediction model for the new tasks is trained with the training samples in the dataset. The trained model is used to infer and predict the resource requirements of the new tasks within the next L prediction time steps; and elastic scaling is performed based on the predicted resource requirements.

[0088] The method and system for efficiently executing computing tasks in the multi-modal intelligent connection computing network in the embodiments of the present invention can, on the premise of providing resource guarantee capabilities for intelligent connection computing tasks, improve the resource utilization rate as much as possible and ensure efficient and transparent execution within the task network.

[0089] Embodiment 2

[0090] The embodiments of the present invention provide a system for efficiently executing computing tasks in a multi-modal intelligent connection computing network, including:

[0091] Data acquisition and preprocessing module: Real-time collect the computing, storage, and network resource usage data of intelligent connection computing tasks at fixed time intervals through a network-aware resource monitoring tool, record the modal feature information of the tasks, complete the work of data cleaning and feature extraction, and store the extracted features and time series data in a time series database to form a resource prediction dataset.

[0092] Resource requirement prediction module: Adopt a combined model of Prophet and LSTM. Capture the long-term trend, periodicity, and holiday effect of resource requirements through the Prophet model, and model the prediction residuals of the Prophet through the LSTM model to capture non-linear fluctuations, such as bursty traffic. Complete the training work of the combined model, deploy the trained model as an online prediction service to predict the resource requirement vector sequence in the future time window, and support incremental learning to adapt to changes in resource requirement patterns.

[0093] Resource requirement planning module: Calculate the resource volatility characteristics of computing tasks, obtain the prediction time window, generate a reasonable elastic scaling strategy based on the prediction result and task modal characteristics, update the result to the task record table, and execute the elastic scaling instruction by calling the downstream resource orchestration system through the API, coordinating horizontal and vertical scaling operations to ensure a smooth transition of resource allocation.

[0094] Task Modal Feature Matching and Recommendation Module: When a new task arrives, extract the task modal feature vector, calculate the weighted cosine similarity with similar tasks of the same category of modality in the task record table, perform K-means clustering on the similar tasks, obtain the maximum cluster center value as the recommended configuration, make reasonable configuration corrections to the user-defined configuration, and dynamically optimize the recommendation model based on the data collection results and the system operation status. The system architecture is as Figure 4 shown.

[0095] The related technical solutions are the same as above and will not be elaborated here.

[0096] Embodiment 3

[0097] The embodiment of the present invention provides an efficient execution system for computing tasks in a multi-modal intelligent connection computing network, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the efficient execution method for computing tasks in the multi-modal intelligent connection computing network in Embodiment 1 above.

[0098] The related technical solutions are the same as above and will not be elaborated here.

[0099] Embodiment 4

[0100] The embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the efficient execution method for computing tasks in the multi-modal intelligent connection computing network in Embodiment 1 above.

[0101] Specifically, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0102] The related technical solutions are the same as above and will not be elaborated here.

[0103] Embodiment 5

[0104] The embodiment of the present invention provides a computer program product, including a computer program. When the computer program runs on a computer, it causes the computer to execute the efficient execution method for computing tasks in the multi-modal intelligent connection computing network in Embodiment 1 above.

[0105] The related technical solutions are the same as above and will not be elaborated here.

[0106] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An efficient execution method for computing tasks in a multimodal intelligent connection computing network, characterized in that, Including: The resource usage time series data of the current time step and the adjacent T - 1 historical time steps of the real - time acquisition task is input into the trained resource demand prediction model to obtain the resource usage time series data within L prediction time steps; where L is the time window length of task load stability determined by the resource volatility and modal characteristics of the task, and the resource volatility is determined based on the historical resource usage time series data of the task; the task is a computing task in a multi - modal intelligent connection computing network. Determine the optimal resource configuration required by the task within the L prediction time steps using the resource usage time series data within the L prediction time steps; compare the optimal resource configuration with the preset elastic scaling policy rules, and when resource elastic scaling is required, initiate an elastic scaling request to the downstream resource orchestration system to enable it to execute the coordinated scheduling of horizontal and vertical scaling.

2. The efficient execution method of computing tasks in the multimodal intelligent connection computing network according to claim 1, wherein Also including: When a new task arrives, obtain the modal characteristics of the new task according to the request information of the new task, perform similarity matching between the modal characteristics of the new task and the modal characteristics of each task in the task record table, and based on the similarity matching result, recommend the best resource configuration in the task record table as the resource configuration specification of the new task Wherein, the task record table includes the best resource configuration required by the task within L prediction time steps and the corresponding task modal characteristics; Calculate the resource configuration specification recommended for the new task Difference coefficient from the pre-defined configuration specification R of the new task When the difference coefficient α does not exceed the preset threshold, adopt the recommended resource configuration specification As the final recommended result of the new task resource configuration, otherwise, adopt the pre-defined configuration specification R of the new task as the final recommended result; Feed the final recommendation result back to the task scheduler of the downstream resource orchestration system for resource scheduling, and obtain the historical resource usage time series data of the new task, use it as a training sample set to train the resource demand prediction model of the new task, use the trained resource demand prediction model to predict the resource demand of the new task, and perform elastic scaling based on the predicted resource demand.

3. The efficient execution method of computing tasks in the multimodal intelligent connection computing network according to claim 2, characterized in that Perform similarity matching between the modal features of the new task and the modal features of each task in the task record table. Based on the similarity matching results, recommend the optimal resource configuration in the task record table as the resource configuration specification for the new task. including: Calculate the weighted cosine similarity between the modal characteristics of the new task and the modal characteristics of the tasks with the same modal category in the task list, and use the weighted cosine similarity as the similarity score. Cluster the task modality features and corresponding resource volatility in the task list with similarity scores exceeding the threshold τ, and take the optimal resource configuration of the task closest to the center of the cluster with the largest distance as the resource configuration specification recommended for the new task.

4. The method for efficient execution of computing tasks in a multimodal intelligent connection computing network according to any one of claims 1-3, characterized in that The calculation method of the resource volatility is: Where CV is the resource volatility, and σ and μ respectively represent the standard deviation and mean of the historical resource usage time series data of the task.

5. The method for efficient execution of computing tasks in a multimodal intelligent connection computing network according to any one of claims 1-3, characterized in that, The trained resource demand prediction model is obtained by collecting the historical resource usage time series data of the task in real - time, using it as a training sample set, and training the resource demand prediction model with the training sample set; where the training sample includes the resource usage time series data of T historical time steps and the resource usage time series data of L prediction time steps, and the resource usage time series data of one time step is the data obtained by splicing the resource usage data of multiple sampling time intervals ΔT according to time series and then compressing the data.

6. The method for efficient execution of computing tasks in the multimodal intelligent connection computing network according to claim 5, characterized in that, The resource types of the task include: CPU usage, memory usage, network bandwidth usage; data compression of each task resource includes: using the 95th percentile to represent the CPU usage within one time step, and using the peak value to represent the memory usage and network bandwidth usage within one time step. Determine the optimal resource configuration required by the task within the L prediction time steps using the resource usage time series data within the L prediction time steps, including: Optimal Resource Allocation R of CPU CPU = P 95 (Y CPU ) × (1 + λ CPU ), where λ CPU is the reservation coefficient of the CPU, and Y CPU represents the CPU usage within L prediction time steps, and P 95 (·) represents the 95th percentile operation; Optimal Resource Allocation R of Memory Mem = Max(Y Mem ) × (1 + λ Mem ), where λ Mem is the reserved coefficient of memory, and Y Mem represents the memory usage within L predicted time steps. Max(·) represents the peak operation; For the optimal bandwidth resource allocation R BW = Max(Y BW ) × (1 + λ BW ), where λ BW is the reservation coefficient of bandwidth resources, and Y BW represents the bandwidth usage within L predicted time steps; Among them, the reservation coefficient of various resources decreases as the overall resource load of the multi - modal intelligent connection computing network increases.

7. The method for efficient execution of computing tasks in the multi-modal intelligent connection computing network according to claim 6, characterized in that, The resource demand prediction model includes a Prophet network and an LSTM. Predicted value of Prophet network is as follows: Where g(t) is the trend term of the training sample, s(t) is the seasonal term of the training sample, and h(t) is the holiday effect of the training sample. The LSTM is used to model the residual r that the Prophet network fails to capture t to obtain the predicted residual where y t is the input training sample; The predicted value of the resource demand prediction model is as follows:

8. An efficient execution system for computing tasks in a multi-modal intelligent connection computing network, characterized in that, Including a computer - readable storage medium and a processor; The computer - readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method for efficiently executing a computing task in the multimodal intelligent connection computing network according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for efficiently executing a computing task in the multimodal intelligent connection computing network according to any one of claims 1-7.

10. A computer program product, characterized in that, It includes a computer program, which when running on a computer causes the computer to execute the method for efficiently executing a computing task in the multimodal intelligent connection computing network according to any one of claims 1-7.

Images