Middleware service multi-granularity load prediction and intelligent resource pooling performance optimization method and system

Abstract

The invention relates to the technical field of middleware service load prediction and resource optimization, in particular to a middleware service multi-granularity load prediction and intelligent resource pooling performance optimization method and system. The system comprises a local load feature extraction module, a global load distribution modeling unit, an adaptive resource scheduling engine and a resource utilization rate monitoring feedback loop. The load data processing efficiency is improved through a sliding window mechanism and a distributed architecture, the load prediction accuracy is enhanced by utilizing a hierarchical neural network and dynamic weight adjustment, and efficient resource allocation is realized in combination with task priority analysis and heterogeneous resource adaptation. The resource utilization rate and the task processing efficiency in the complex distributed system can be remarkably improved, and the method has wide application prospects and practical value.

Description

Technical Field

[0001] This invention belongs to the field of information technology and computer science, specifically a method and system for performance optimization of middleware service multi-granularity load prediction and intelligent resource pooling. Background Technology

[0002] With the development of cloud computing and distributed systems, middleware services are playing an increasingly prominent role in resource scheduling and load management. While existing resource pooling technologies can achieve a certain degree of resource allocation optimization, they still have shortcomings in multi-granularity load prediction and intelligent resource pooling performance optimization. Dynamically changing load demands in complex scenarios place higher demands on resource utilization efficiency.

[0003] A search revealed patent document CN117992239B, which discloses a resource management and allocation method, an intelligent computing cloud operating system, and a computing platform, published on June 18, 2024. This technology achieves effective management and utilization of intelligent platform resources through a resource pooling allocation model and multi-level resource pool configuration, improving system resource utilization. However, this technical solution lacks a multi-granularity load prediction mechanism for middleware services. Its resource pooling strategy primarily relies on static classification processing and logical resource configuration, which has limitations in responding to real-time dynamic load changes. Furthermore, this solution does not adequately consider the correlation and priority between tasks during resource allocation, potentially affecting resource allocation efficiency, especially in high-concurrency scenarios, where the ability for fine-grained resource scheduling needs improvement.

[0004] A search revealed that patent document CN115994023B discloses a method, device, terminal, and medium for scheduling resource information in a visual intelligence system, published on June 6, 2023. This technology reduces the complexity of task and resource management by structuring and quantifying device resource and task information to create a resource pool and implement task scheduling. However, this technical solution is primarily geared towards specific application scenarios in visual intelligence systems. Its load prediction method is suitable for video analysis tasks but is difficult to extend to multi-granularity load prediction for general middleware services. Furthermore, this solution lacks a comprehensive consideration of global load distribution in resource pooling performance optimization, potentially leading to uneven resource utilization and limited dynamic adaptation and optimization capabilities in heterogeneous resource environments.

[0005] The aforementioned problems indicate that existing resource pooling technologies still have significant shortcomings in multi-granularity load prediction and intelligent resource pooling performance optimization for middleware services. Therefore, this invention provides a method and system for multi-granularity load prediction and intelligent resource pooling performance optimization for middleware services. The aim is to improve the accuracy and dynamic adaptability of resource allocation by introducing a multi-granularity load prediction model and an intelligent resource pooling strategy, thereby meeting the needs of complex distributed systems for efficient and flexible resource management. Summary of the Invention

[0006] This invention provides a method and system for multi-granularity load prediction and intelligent resource pooling performance optimization of middleware services, aiming to solve the problems of insufficient responsiveness to dynamically changing load demands, low resource allocation efficiency, and uneven resource utilization in existing technologies for complex distributed systems. By introducing a local load feature extraction module, a global load distribution modeling unit, and an adaptive resource scheduling engine, combined with a task priority analysis mechanism and heterogeneous resource adaptation strategy, accurate prediction of middleware service load and intelligent resource pooling management are achieved.

[0007] The technical solution of the present invention includes the following components: First, the local load feature extraction module collects real-time data streams from the middleware service during operation and divides the load data into multiple granular levels based on a time-series segmentation algorithm. This module uses a sliding window mechanism to segment the load data, and each segment generates a corresponding load feature vector after feature extraction. These feature vectors contain key information such as load intensity, volatility, and periodicity, providing foundational data support for subsequent global load distribution modeling. The local load feature extraction module adopts a distributed architecture design, consisting of multiple parallel computing nodes. Each node is responsible for processing load data within a specific time period, thereby improving data processing efficiency.

[0008] Secondly, the global load distribution modeling unit receives load feature vectors from the local load feature extraction module and constructs a global load distribution map using a deep learning model. This unit employs a hierarchical neural network structure, where the input layer receives the load feature vectors, the hidden layer abstracts and expresses the load features through a non-linear activation function, and the output layer generates the global load distribution matrix. The global load distribution matrix reflects the load status of each resource node in the current system and their interrelationships. Furthermore, the global load distribution modeling unit introduces a dynamic weight adjustment mechanism, automatically adjusting the weight values ​​of each node based on real-time load changes to enhance the model's adaptability to dynamic load variations.

[0009] Furthermore, the adaptive resource scheduling engine makes resource allocation decisions based on a global load distribution matrix. This engine categorizes and prioritizes currently pending tasks through a task priority analysis mechanism, giving priority to high-priority tasks. The task priority analysis mechanism comprehensively considers the urgency of tasks, resource requirements, and their correlation with other tasks, forming a multi-dimensional task priority scoring model. The adaptive resource scheduling engine also introduces a heterogeneous resource adaptation strategy, formulating differentiated allocation rules for different types of computing resources (such as CPU, GPU, and memory) to ensure flexibility and efficiency in resource allocation.

[0010] Furthermore, the present invention also includes a resource utilization monitoring feedback loop for real-time monitoring of resource allocation effectiveness and dynamic adjustment of the allocation strategy. This loop collects resource utilization data through sensor components and transmits the data to a data analysis module for processing. The data analysis module determines whether the current resource allocation is reasonable based on preset threshold conditions. If the resource utilization rate is found to be lower than the set threshold, a reallocation process is triggered. The reallocation process recalculates the resource allocation scheme by calling an adaptive resource scheduling engine until the resource utilization rate meets the requirements.

[0011] The technical effects of this invention are reflected in the following aspects: First, the local load feature extraction module significantly improves the efficiency and accuracy of load data processing through a sliding window mechanism and distributed architecture design. This design enables the system to quickly capture load change trends in complex scenarios, providing high-quality data support for subsequent global load distribution modeling.

[0012] Second, the global load distribution modeling unit adopts a hierarchical neural network structure and a dynamic weight adjustment mechanism, which enhances the system's responsiveness to dynamic load changes. The hierarchical neural network structure, through multi-level abstraction of load characteristics, can more comprehensively reflect the global load distribution status; the dynamic weight adjustment mechanism automatically optimizes model parameters based on real-time load changes, further improving the model's prediction accuracy.

[0013] Third, the adaptive resource scheduling engine combines a task priority analysis mechanism and a heterogeneous resource adaptation strategy to achieve refined and flexible resource allocation. The task priority analysis mechanism ensures that high-priority tasks are processed in a timely manner by scoring tasks from multiple dimensions; the heterogeneous resource adaptation strategy formulates differentiated allocation rules based on the characteristics of different resource types to avoid resource waste and allocation conflicts.

[0014] Fourth, the resource utilization monitoring feedback loop effectively improves the rationality of resource allocation through real-time monitoring and dynamic adjustment mechanisms. This loop can promptly identify problems of low resource utilization and optimize resource utilization through reallocation processes, thereby ensuring the overall performance stability of the system.

[0015] In summary, this invention addresses the shortcomings of existing technologies in multi-granularity load prediction and intelligent resource pooling performance optimization through the collaborative work of a local load feature extraction module, a global load distribution modeling unit, an adaptive resource scheduling engine, and a resource utilization monitoring feedback loop. This invention not only improves the accuracy and dynamic adaptability of resource allocation but also significantly enhances resource utilization and task processing efficiency in complex distributed systems, demonstrating broad application prospects and practical value. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system architecture of the present invention, illustrating the logical relationship and collaborative working method between the local load feature extraction module, the global load distribution modeling unit, the adaptive resource scheduling engine, and the resource utilization monitoring feedback loop.

[0017] Figure 2 This is a flowchart of the local load feature extraction module, which details the process of segmenting load data using the sliding window mechanism and the steps for generating load feature vectors.

[0018] Figure 3 This diagram illustrates the task priority analysis mechanism of the adaptive resource scheduling engine, showcasing the multi-dimensional input parameters of the task priority scoring model and the output flow of priority ranking results.

[0019] The attached figures are labeled as follows: 1. Local load feature extraction module; 2. Global load distribution modeling unit; 3. Adaptive resource scheduling engine; 4. Resource utilization monitoring feedback loop; 5. Sliding window mechanism; 6. Load feature vector; 7. Task priority scoring model; 8. Heterogeneous resource adaptation strategy. Detailed Implementation

[0020] This invention provides a method and system for middleware service multi-granularity load prediction and intelligent resource pooling performance optimization, the specific implementation of which combines... Figure 1 , Figure 2 and Figure 3 The accompanying drawings are described in detail. Figure 1 This is a schematic diagram of the system architecture of the present invention, illustrating the logical relationship and collaborative working method between the local load feature extraction module 1, the global load distribution modeling unit 2, the adaptive resource scheduling engine 3, and the resource utilization monitoring feedback loop 4. Figure 2 The flowchart of the local load feature extraction module 1 describes in detail the process of segmenting the load data using the sliding window mechanism 5, as well as the steps for generating the load feature vector 6. Figure 3This diagram illustrates the task priority analysis mechanism of the Adaptive Resource Scheduling Engine 3, showcasing the multi-dimensional input parameters and output flow of the priority ranking results for the Task Priority Scoring Model 7.

[0021] In practical implementation, the local load feature extraction module 1, as the first part of the system, is responsible for collecting real-time data streams from the middleware service during operation and dividing the load data into multiple granular levels based on a time series segmentation algorithm. This module uses a sliding window mechanism 5 to segment the load data. Specifically, the sliding window mechanism 5 divides the load data into several consecutive time segments according to a preset time interval. The length of each time segment is determined by system requirements, typically on the order of seconds or minutes. The sliding window mechanism 5 moves with a fixed step size to ensure a certain overlap between each time segment, capturing the continuous characteristics of load changes. After feature extraction, the load data within each time segment generates a corresponding load feature vector 6, which contains key information such as load intensity, volatility, and periodicity. This information extraction is accomplished by calculating statistical indicators within each time segment, such as mean, variance, peak value, and frequency domain features. The local load feature extraction module 1 adopts a distributed architecture design, consisting of multiple parallel computing nodes, each responsible for processing load data within a specific time period. These nodes are connected via high-speed communication links to ensure efficient data transmission between nodes. The output of each node is aggregated into the global load distribution modeling unit 2 to form a complete set of load features.

[0022] The global load distribution modeling unit 2 receives the load feature vector 6 from the local load feature extraction module 1 and constructs a global load distribution map using a deep learning model. This unit employs a hierarchical neural network structure, where the input layer receives the load feature vector 6, the hidden layers abstractly represent the load features using non-linear activation functions, and the output layer generates the global load distribution matrix. The hierarchical neural network structure includes multiple hidden layers, with the number of neurons in each hidden layer dynamically adjusted according to the complexity of the load features to ensure the model can fully express the multi-level information of the load features. The global load distribution matrix reflects the load status of each resource node in the current system and their interrelationships; its generation depends on the input of the load feature vector 6 and the training results of the neural network. To enhance the model's adaptability to dynamic load changes, the global load distribution modeling unit 2 introduces a dynamic weight adjustment mechanism. This mechanism automatically adjusts the weight values ​​of each node by monitoring real-time load changes, enabling the model to quickly respond to load fluctuations. The specific implementation of the dynamic weight adjustment mechanism involves calculating the load change rate and combining it with historical load data to dynamically update the weight parameters of each node in the neural network. The output of the global load distribution matrix is ​​passed to the adaptive resource scheduling engine 3 for subsequent resource allocation decisions.

[0023] The Adaptive Resource Scheduling Engine 3 makes resource allocation decisions based on a global load distribution matrix. Its core function is to classify and sort currently pending tasks through a task priority analysis mechanism. The specific implementation of the task priority analysis mechanism is as follows: Figure 3 As shown, tasks are first scored across multiple dimensions using Task Priority Scoring Model 7. The input parameters of Task Priority Scoring Model 7 include the task's urgency, resource requirements, and relevance to other tasks. Urgency is assessed based on the difference between the task's deadline and the current time; a smaller difference indicates higher urgency. Resource requirements are assessed based on the amount of CPU, GPU, memory, and other resources needed; higher resource requirements result in higher priority. Task relevance is assessed based on the task's dependencies on other tasks; stronger dependencies result in higher priority. The output of Task Priority Scoring Model 7 is a comprehensive score, obtained by weighted summation of the input parameters. A higher score indicates higher task priority. The Adaptive Resource Scheduling Engine 3 sorts tasks based on the output of Task Priority Scoring Model 7 and prioritizes high-priority tasks. Furthermore, the Adaptive Resource Scheduling Engine 3 introduces a heterogeneous resource adaptation strategy 8, which establishes differentiated allocation rules for different types of computing resources. The specific implementation of the heterogeneous resource adaptation strategy 8 involves setting different allocation weights based on the characteristics of different resource types. For example, CPU resources are prioritized for CPU-intensive tasks, and GPU resources are prioritized for GPU-intensive tasks. The allocation rules of heterogeneous resource adaptation strategy 8 are implemented by dynamically adjusting weight parameters to ensure the flexibility and efficiency of resource allocation.

[0024] The resource utilization monitoring feedback loop 4 is used to monitor the resource allocation effect in real time and dynamically adjust the allocation strategy. This loop collects resource utilization data through sensor components deployed on various resource nodes of the system to collect data such as CPU utilization, GPU utilization, and memory usage in real time. The collected data is transmitted to the data analysis module via a high-speed communication link. The data analysis module determines whether the current resource allocation is reasonable based on preset threshold conditions. The preset threshold conditions include upper and lower limits for resource utilization, such as triggering a reallocation process when CPU utilization is below 20% or above 80%. If the resource utilization is found to be below the set threshold, the reallocation process is triggered. The reallocation process recalculates the resource allocation scheme by calling the adaptive resource scheduling engine 3 until the resource utilization meets the requirements. The specific implementation of the reallocation process is to adjust the weight parameters of the task priority scoring model 7, recalculate the task priority, and reallocate resources according to the new priority results. The dynamic adjustment mechanism of the resource utilization monitoring feedback loop 4 ensures that the system can maintain the rationality of resource allocation under frequent load changes.

[0025] In practical applications, the technical solution of this invention can be applied to complex distributed systems such as cloud computing platforms and big data processing centers. Taking a cloud computing platform as an example, the local load feature extraction module 1 is deployed in the cloud server cluster and is responsible for collecting load data of virtual machine instances. The global load distribution modeling unit 2 is deployed on the central control node and is responsible for constructing a global load distribution map. The adaptive resource scheduling engine 3 is deployed on the resource management node and is responsible for making resource allocation decisions based on the global load distribution matrix. The resource utilization monitoring feedback loop 4 is deployed on the monitoring node and is responsible for monitoring resource usage in real time and triggering the reallocation process. Through the collaborative work of the above modules, the system can achieve accurate prediction of middleware service load and intelligent resource pooling management. For example, if the system detects a significant increase in the load intensity of a virtual machine instance within a certain period, the local load feature extraction module 1 captures this change through the sliding window mechanism 5 and generates a corresponding load feature vector 6. The global load distribution modeling unit 2 updates the global load distribution matrix based on the load feature vector 6, and the adaptive resource scheduling engine 3 reallocates resources based on the updated matrix to ensure that high-load instances obtain sufficient computing resources. Meanwhile, the resource utilization monitoring feedback loop 4 monitors resource usage in real time. If it detects that the utilization rate of certain resource nodes is too low, it triggers a reallocation process to further optimize resource utilization efficiency. Through the above implementation method, the present invention can effectively solve the problems of insufficient responsiveness to dynamically changing load demands, low resource allocation efficiency, and uneven resource utilization in existing technologies in complex distributed systems, thereby improving the overall performance and stability of the system.

[0026] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention will be further explained below in conjunction with a specific application scenario.

[0027] In the actual operation of the cloud computing platform, the local load feature extraction module 1 is deployed on multiple nodes of the cloud server cluster, responsible for collecting real-time load data of virtual machine instances. This data includes key indicators such as CPU utilization, memory usage, and network bandwidth consumption. Using a sliding window mechanism 5, module 1 divides the load data into multiple time segments, each segment being 1 minute long with a 30-second step size to ensure overlap between adjacent segments, thereby capturing the continuous characteristics of load changes. Subsequently, the load data within each time segment undergoes statistical analysis to generate a load feature vector 6 containing information on load intensity, volatility, and periodicity. For example, if the load intensity of a virtual machine instance increases significantly within a certain time period, the sliding window mechanism 5 can quickly capture this change and generate the corresponding load feature vector 6, serving as the foundational data for subsequent global load distribution modeling.

[0028] The global load distribution modeling unit 2 is deployed on the central control node. It receives the load feature vector 6 from the local load feature extraction module 1 and constructs a global load distribution matrix through a hierarchical neural network structure. This matrix reflects the load status of each resource node in the current system and their interrelationships. Specifically, the input layer receives the load feature vector 6, the hidden layer abstracts the load features using a non-linear activation function, and the output layer generates the global load distribution matrix. To enhance the model's adaptability to dynamic load changes, the global load distribution modeling unit 2 introduces a dynamic weight adjustment mechanism. This mechanism dynamically updates the weight parameters of each node in the neural network by calculating the load change rate and combining it with historical load data. For example, when a sudden increase in the load of a resource node is detected, the dynamic weight adjustment mechanism automatically increases the weight value of that node, enabling the model to quickly respond to load fluctuations and update the global load distribution matrix.

[0029] The adaptive resource scheduling engine 3 is deployed on the resource management node and makes resource allocation decisions based on the global load distribution matrix. First, the task priority scoring model 7 scores the currently pending tasks from multiple dimensions. The input parameters of the scoring model include the urgency of the task, its resource requirements, and its relevance to other tasks. For example, a task approaching its deadline has a higher urgency score; a task requiring a large amount of GPU resources has a higher resource requirements score; and a task dependent on other tasks has a higher relevance score. The scoring model generates a comprehensive score by weighted summing of these input parameters and sorts the tasks according to the score. Subsequently, the adaptive resource scheduling engine 3 prioritizes high-priority tasks and formulates differentiated resource allocation rules based on the heterogeneous resource adaptation strategy 8. For example, CPU resources are prioritized for CPU-intensive tasks, and GPU resources are prioritized for GPU-intensive tasks. This strategy is implemented by dynamically adjusting weight parameters to ensure the flexibility and efficiency of resource allocation.

[0030] Resource utilization monitoring feedback loop 4 is deployed on the monitoring node, collecting real-time utilization data of each resource node through sensor components, including CPU utilization, GPU utilization, and memory usage. The collected data is transmitted to the data analysis module via a high-speed communication link. The module determines whether the current resource allocation is reasonable based on preset threshold conditions. For example, when the CPU utilization of a resource node is below 20% or above 80%, a reallocation process is triggered. The reallocation process calls the adaptive resource scheduling engine 3 to recalculate the resource allocation scheme until the resource utilization meets the requirements. Specifically, the reallocation process adjusts the weight parameters of the task priority scoring model 7, recalculates task priorities, and reallocates resources based on the new priority results. For example, when the utilization of a resource node is too low, the system will lower the priority of tasks on that node and reallocate resources to higher-priority tasks, thereby optimizing resource utilization efficiency.

[0031] Through the above steps, this invention enables accurate prediction of middleware service load and intelligent resource pooling management. For example, if the system detects a significant increase in the load intensity of a virtual machine instance within a certain time period, the local load feature extraction module 1 captures this change through the sliding window mechanism 5 and generates a corresponding load feature vector 6. The global load distribution modeling unit 2 updates the global load distribution matrix based on the load feature vector 6, and the adaptive resource scheduling engine 3 reallocates resources according to the updated matrix to ensure that high-load instances receive sufficient computing resources. Simultaneously, the resource utilization monitoring feedback loop 4 monitors resource usage in real time. If it detects that the utilization of certain resource nodes is too low, it triggers a reallocation process to further optimize resource utilization efficiency. Therefore, this invention, through the collaborative work of the local load feature extraction module 1, the global load distribution modeling unit 2, the adaptive resource scheduling engine 3, and the resource utilization monitoring feedback loop 4, solves the problems of insufficient responsiveness to dynamically changing load demands, low resource allocation efficiency, and uneven resource utilization in existing technologies in complex distributed systems, thereby improving the overall performance and stability of the system.

[0032] All content not described in detail in this specification is prior art known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited; conventional equipment can be used. Electrical control components not mentioned in this technical solution are not shown in the figures because they are prior art, and will not be described further here.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A middleware service multi-granularity load prediction and intelligent resource pooling performance optimization system, characterized in that It comprises a local load feature extraction module (1), a global load distribution modeling unit (2), an adaptive resource scheduling engine (3), and a resource utilization monitoring feedback loop (4). The local load feature extraction module (1) processes the real-time data stream in the middleware service running process by a sliding window mechanism (5) and generates a load feature vector (6). The global load distribution modeling unit (2) receives the load feature vector (6) and generates a global load distribution matrix through a hierarchical neural network structure. The adaptive resource scheduling engine (3) allocates resources based on the global load distribution matrix, combined with a task priority scoring model (7) and a heterogeneous resource adaptation strategy (8). The resource utilization monitoring feedback loop (4) collects resource usage data through a sensor component and dynamically adjusts the resource allocation strategy.

2. The middleware service multi-granularity load prediction and intelligent resource pooling performance optimization system of claim 1, wherein The local load feature extraction module (1) adopts a distributed architecture design, which is composed of multiple parallel computing nodes. Each node is responsible for processing the load data in a specific time period and aggregating the results to the global load distribution modeling unit (2).

3. The middleware service multi-granularity load prediction and intelligent resource pooling performance optimization system of claim 1, wherein The global load distribution modeling unit (2) introduces a dynamic weight adjustment mechanism. This mechanism dynamically updates the weight parameters of each node in the hierarchical neural network by calculating the load change rate combined with historical load data to respond to load fluctuations.

Images