Hotspot sensing hierarchical storage and computing resource scheduling method for multi-modal data

Abstract

The invention discloses a hotspot sensing hierarchical storage and computing resource scheduling method oriented to multi-modal data, and belongs to the technical field of data center resource management. The core of the method is as follows: firstly, monitoring characteristics, access modes and computing node loads of multi-modal data in real time; secondly, a comprehensive popularity evaluation model is used for dynamically calculating the comprehensive popularity value of the data object, and the model comprehensively considers the data type, the access frequency, the data size and the task priority; then, intelligently scheduling the data to different performance levels of the hierarchical storage system according to the popularity value; and finally, when a calculation task is scheduled, synchronously sensing the heat and the position of required data, and carrying out collaborative scheduling by taking optimal data locality and load balance as targets. According to the method, storage and computing resources are managed cooperatively through the unified heat index, the problem of resource scheduling disjunction in a multi-modal data environment is effectively solved, and the system throughput rate, the response speed and the resource utilization efficiency are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data center resource management technology, and in particular to a hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data. Background Technology

[0002] In today's era of big data and artificial intelligence, the data generated by applications exhibits significant multimodal characteristics. Scenarios such as smart cities and autonomous driving require the simultaneous processing of multiple formats, including text, images, videos, and time-series data. This data deluge places immense pressure on data center infrastructure. However, existing storage and computing resource scheduling methods suffer from severe technological disconnects, making it difficult to efficiently address this challenge. At the storage level, traditional tiered storage systems primarily determine data hotness based on simple access frequency, lacking a fine-grained understanding of the inherent differences in access patterns and performance requirements of multimodal data (such as large-capacity video versus small-scale logs). This often results in high-speed storage resources being ineffectively occupied by cold data, while hot data faces I / O bottlenecks. At the computing level, mainstream schedulers (such as Kubernetes and YARN) typically manage computing resources independently, with insufficient coordination with storage systems. This can easily lead to computing tasks being scheduled to nodes far from the required data, causing high network overhead for data migration. This fragmented management of storage and computing resources, and the lack of dynamic collaborative awareness of "data hotspots" and "computing hotspots," have become key bottlenecks restricting the overall performance improvement of the system.

[0003] The existing data center architecture is increasingly revealing systemic deficiencies in its storage and computing resource scheduling when processing multimodal data, manifesting in two major micro-bottlenecks. First, there is a mismatch between the static allocation of storage resources and dynamic demand. Tiered storage based on fixed strategies cannot accurately characterize the dynamic "heat" determined by data type and access patterns, causing hot data to potentially linger on low-speed storage media, directly slowing down data access. Second, and more critically, is the severe decoupling between computing task scheduling and data layout. Current systems generally adopt a passive mode of "computation finds data" or "data waits for computation." When allocating tasks, the resource scheduler only focuses on computing resources such as CPU and memory, ignoring the specific location of the underlying data and the performance differences of the storage media. This not only generates significant data movement overhead but also makes storage I / O and network bandwidth prone to becoming bottlenecks in system performance during peak data processing periods, failing to meet the quality of service requirements of low-latency, high-throughput services. Therefore, developing a new method that can collaboratively perceive data hotspots and perform integrated scheduling is crucial for improving resource utilization and application performance. Summary of the Invention

[0004] This invention specifically relates to a hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data. By using a unified heat index to collaboratively manage storage and computing resources, it effectively solves the problem of resource scheduling disconnect in multimodal data environments, significantly improving system throughput, response speed, and resource utilization efficiency. To achieve the above objectives, the specific technical solution of this invention's hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data is as follows: A hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data includes the following steps: S1. Data and system status monitoring steps: Real-time collection of characteristic information of multimodal data, data access mode information, and resource load information of computing nodes; S2, Comprehensive popularity value calculation steps: Based on the information collected in step S1, a comprehensive popularity evaluation model is used to dynamically calculate the real-time comprehensive popularity value of each data object; S3. Storage Tier Decision and Data Scheduling Steps: Based on the comprehensive heat value calculated in step S2, the data object is automatically scheduled to the corresponding tier in the hierarchical storage system composed of storage media with different performance. S4. Computation Task Co-scheduling Steps: When scheduling computation tasks, the overall heat value of the data objects required by the task and their storage level are simultaneously perceived. With the goals of optimal data locality and system load balancing, the computation task is scheduled to be executed on the most suitable computing node.

[0005] Furthermore, in step S1, the feature information of the multimodal data includes data type and data size; the data access pattern information includes access frequency and recent access time; and the resource load information of the computing node includes CPU utilization, memory utilization, and network I / O load.

[0006] Furthermore, in step S2, when calculating the comprehensive popularity value, the comprehensive popularity evaluation model takes into account at least the data type weight, the normalized access frequency, the data size, and the priority of the associated calculation tasks.

[0007] Furthermore, the comprehensive popularity assessment model described in step S2 calculates the comprehensive popularity value H of the data object using the following formula: H = α * W_type + β * F_freq + γ * (1 / S_size) + δ * P_task Where W_type is a preset weight factor based on data type, F_freq is the normalized access frequency, S_size is the data size, P_task is the priority weight of the associated computation task, and α, β, γ, and δ are configurable weight coefficients for each dimension, and α+β+γ+δ=1.

[0008] Furthermore, the storage tier decision in step S3 specifically involves: presetting a popularity threshold range for each tier of the hierarchical storage system; comparing the overall popularity value of a data object with the threshold range to determine the storage tier to which it should be placed; wherein, the data object with the highest popularity value is placed in the storage tier with the fastest access speed.

[0009] Furthermore, step S3 also includes a dynamic data migration mechanism: periodically recalculating the overall popularity value of all data objects. When the current popularity value of a data object does not match the popularity threshold range corresponding to its current storage level, the system automatically triggers a data migration operation to migrate it to a storage level that matches the current popularity value.

[0010] Furthermore, in step S4, with the goal of achieving optimal data locality and system load balancing, the scheduler prioritizes selecting computing nodes that have already locally stored copies of the required data; when no node satisfies data locality, the computing node with the shortest network topology distance, the lowest network overhead, and the lowest current combined CPU and memory load is selected.

[0011] Furthermore, in step S4, when scheduling computing tasks, tasks whose required data is determined to be of high demand are given a higher scheduling priority, and computing resources are allocated to them first.

[0012] Furthermore, the method also includes a feedback optimization step S5: monitoring the actual execution efficiency of the completed computation task. If the task execution efficiency is lower than the expected threshold due to data access delay or network transmission bottleneck, the information is fed back to the comprehensive heat evaluation model in step S2 for dynamic adjustment of relevant parameters or future scheduling strategies for similar tasks and data.

[0013] Furthermore, in the feedback optimization step S5, based on the feedback of task execution efficiency, the configurable weight coefficients α, β, γ, and δ in the comprehensive heat evaluation model are dynamically adjusted, or the heat value of the data object causing the bottleneck is increased, so as to prompt the system to make better decisions in subsequent scheduling.

[0014] The beneficial effects of this invention: Compared with existing technologies, the "Hotspot Awareness Hierarchical Storage and Computing Resource Scheduling Method for Multimodal Data" provided by this invention brings significant benefits. First, its core advantage lies in achieving global collaborative optimization and refined management of storage and computing resources. By constructing a unified hotspot awareness model, this invention can dynamically and accurately characterize the comprehensive "heat value" of multimodal data and intelligently schedule data to the most suitable storage level accordingly. More importantly, it breaks down the barriers of fragmented resource management in traditional architectures, enabling collaborative scheduling based on data hotspot locations and computing task requirements, achieving "computing power close to data" or "data close to computing power," thereby minimizing data network transmission overhead, significantly reducing task latency, effectively alleviating I / O bottlenecks, and comprehensively improving system data processing throughput and response efficiency.

[0015] Secondly, this invention effectively enhances the system's resilience in handling dynamic loads, ultimately achieving the combined value of reducing total cost of ownership and improving business efficiency. The system can perceive the dynamic migration of data and computing hotspots in real time and automatically and quickly adjust resource allocation strategies, thereby ensuring the service quality of critical applications and the overall robustness of the system during business peaks or load fluctuations. This intelligent and adaptive resource scheduling capability greatly improves the utilization efficiency of hardware resources (such as high-speed storage, computing units, and network bandwidth), avoiding resource idleness and waste. This not only directly reduces infrastructure operation and maintenance costs but also provides high-performance, highly reliable support for upper-layer businesses by accelerating data processing flows, possessing significant economic value and practical significance. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the working principle of the present invention; Figure 2 This is a diagram showing the interaction between system components of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description is provided in conjunction with preferred embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0018] like Figure 1-2As shown in the accompanying drawings and specific embodiments, the present invention will be described in great detail below. The core of the present invention lies in constructing a unified and intelligent resource management and scheduling framework. Through real-time, multi-dimensional hotspot perception, it collaboratively decides on the allocation of storage resources and the scheduling of computing tasks, thereby maximizing system performance in a multimodal data environment. A typical implementation architecture includes a global resource manager, a status monitoring agent, a metadata service, a heat assessment unit, a storage scheduler, and a computing scheduler. The status monitoring agent is deployed on each computing / storage node and is responsible for collecting local data access indicators and node resource usage; the heat assessment unit calculates the real-time comprehensive heat value of data objects based on the collected information; the storage scheduler is responsible for the placement and migration of data in the hierarchical storage system; the computing scheduler is responsible for the final scheduling decision of computing tasks; and the metadata service maintains the metadata and real-time status of all data objects. These components work collaboratively through the global resource manager to form a complete closed-loop system.

[0019] A hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data includes the following steps: S1, Data and system status monitoring step: Real-time acquisition of feature information, data access pattern information, and resource load information of computing nodes for multimodal data; S2, Comprehensive heat value calculation step: Based on the information acquired in step S1, a comprehensive heat evaluation model is used to dynamically calculate the real-time comprehensive heat value of each data object; S3, Storage level decision and data scheduling step: Based on the comprehensive heat value calculated in step S2, the data object is automatically scheduled to the corresponding level in a hierarchical storage system composed of storage media with different performance; S4, Computing task collaborative scheduling step: When scheduling computing tasks, the comprehensive heat value of the data object required by the task and its storage level are simultaneously sensed, and the computing task is scheduled to be executed on the most suitable computing node with the goals of optimal data locality and system load balancing. In step S1, the feature information of the multimodal data includes data type and data size; the data access pattern information includes access frequency and recent access time; the resource load information of the computing node includes CPU utilization, memory utilization, and network I / O load. In step S2, the comprehensive popularity assessment model, when calculating the comprehensive popularity value, considers at least the data type weight, normalized access frequency, data size, and the priority of associated computation tasks. The comprehensive popularity assessment model in step S2 calculates the comprehensive popularity value H of the data object using the following formula: H = α * W_type + β * F_freq + γ * (1 / S_size) + δ * P_task, where W_type is a preset weight factor based on data type, F_freq is the normalized access frequency, S_size is the data size, P_task is the priority weight of the associated computation task, and α, β, γ, and δ are configurable weight coefficients for each dimension, and α+β+γ+δ=1. The storage tier decision in step S3 specifically involves: setting a preset popularity threshold range for each tier of the hierarchical storage system; comparing the comprehensive popularity value of the data object with the threshold range to determine the storage tier to which it should be placed; wherein, the data object with the highest popularity value is placed in the storage tier with the fastest access speed.

[0020] Step S3 also includes a dynamic data migration mechanism: periodically recalculating the overall popularity value of all data objects. When the current popularity value of a data object does not match the popularity threshold range corresponding to its current storage level, the system automatically triggers a data migration operation to migrate it to a storage level that matches the current popularity value. Step S4, with the goal of optimal data locality and system load balancing, prioritizes the selection of computing nodes that already locally store copies of the required data. When no node satisfies data locality, the computing node with the shortest network topology distance, the lowest network overhead, and the lowest current combined CPU and memory load is selected. In Step S4, when scheduling computing tasks, tasks whose required data is determined to be of high popularity are given a higher scheduling priority, and computing resources are allocated to them preferentially. This method also includes a feedback optimization step S5: monitoring the actual execution efficiency of completed computing tasks. If the task execution efficiency is lower than the expected threshold due to data access latency or network transmission bottlenecks, this information is fed back to the overall popularity evaluation model in Step S2 for dynamic adjustment of relevant parameters or future scheduling strategies for similar tasks and data. In the feedback optimization step S5, based on the feedback of task execution efficiency, the configurable weight coefficients α, β, γ, and δ in the comprehensive heat evaluation model are dynamically adjusted, or the heat value of the data object causing the bottleneck is increased, so as to prompt the system to make better decisions in subsequent scheduling.

[0021] Example 1: Systematic Application in Smart City Video Surveillance Platform Taking a large-scale smart city video surveillance and analysis platform as an application scenario, this platform needs to continuously process multimodal data such as video streams, sensor data, and logs. In the data and system status monitoring step (S1), the status monitoring agents deployed on each node continuously collect the characteristic information (such as type and size), data access pattern information (such as access frequency and recent access time) of the multimodal data, and the resource load information of the computing nodes (such as CPU, memory, and network I / O utilization), and report this real-time data. In the comprehensive popularity value calculation step (S2), the popularity evaluator periodically performs preprocessing such as normalization on the collected raw data, and uses the comprehensive popularity evaluation model to calculate the real-time popularity value of each data object. This model comprehensively considers multiple dimensions such as data type weight, normalized access frequency, data size (taking the reciprocal), and associated task priority, and calculates the quantified popularity value H through the weighted formula (H = αW_type + βF_freq + γ(1 / S_size) + δP_task), and then updates it to the metadata service.

[0022] In the storage tier decision and data scheduling step (S3), the storage scheduler schedules data objects to the appropriate tier in the hierarchical storage system based on preset popularity threshold ranges (e.g., H≥0.8 is placed in memory, 0.5≤H<0.8 is placed in NVMe SSD, 0.2≤H<0.5 is placed in SATA SSD, and H<0.2 is placed in HDD). For example, a video data block with a popularity value of H=0.6053 will be decided to be placed in the NVMe SSD tier. Simultaneously, the system runs a dynamic data migration mechanism, periodically recalculating popularity values. When changes in data popularity cause a mismatch between its current tier and the popularity threshold range, a data migration operation is automatically triggered, such as migrating log files with increased popularity from HDD to SATA SSD. In the collaborative scheduling step (S4) of computing tasks, when a computing task is submitted, the computing scheduler first parses its data dependencies and queries the metadata service to obtain the real-time heat value and physical location of the required data blocks. Then, it filters and scores nodes with the goals of optimal data locality and system load balancing, giving priority to nodes that have the required data stored locally. If this is not possible, it comprehensively compares network transmission overhead and real-time node load. At the same time, tasks with high overall heat of the required data are given higher scheduling priority. Finally, the task is scheduled to be executed on the most suitable computing node based on the scoring results.

[0023] Example 2: Application of Feedback Optimization in Online Machine Learning Platforms This embodiment focuses on demonstrating how the feedback optimization mechanism works in a distributed model training scenario. At the initial run of the task, the system schedules according to the baseline process of Embodiment 1. In the feedback optimization step (S5), the system continuously monitors the actual execution efficiency of completed computational tasks, such as the iteration speed of each Worker process. When the system analyzes and finds that a certain Worker process is slow in iteration due to frequent data reading from remote nodes (i.e., the task execution efficiency is lower than the expected threshold due to network transmission bottlenecks), this bottleneck information (including a list of specific data blocks involved) is fed back to the heat evaluator. Subsequently, the system dynamically adjusts based on this feedback: one approach is to directly increase the heat value of the specific data block causing the bottleneck (e.g., temporarily increase ΔH=0.2), so that it is preferentially migrated to a faster storage layer or affects the data locality decision of the computational task in subsequent storage scheduling; another approach is to progressively adjust the configurable weight coefficients in the comprehensive heat evaluation model (e.g., fine-tuning the values ​​of α, β, γ, and δ), so that the model is more inclined to avoid such network bottlenecks during long-term learning. After several training cycles and feedback adjustments, the system can automatically optimize data layout and task scheduling strategies, which significantly improves the execution efficiency of subsequent similar tasks, thus demonstrating the system's self-learning and adaptive capabilities.

[0024] Working principle and operation process of the present invention The core working principle of this invention lies in using a unified "comprehensive heat value" as the core metric to break down the barriers between storage management and computation scheduling, enabling global optimization decisions across resources. The system dynamically quantifies the value, activity, and business importance of data into a calculable numerical value. Based on this value, it intelligently places data in the appropriate storage tier and schedules computation tasks to the most suitable location. The ultimate goal is to achieve optimal resource allocation in a multimodal data environment, ensuring "fast access to hot data and priority execution of hot tasks." The entire operating mechanism is a dynamic closed-loop system encompassing real-time monitoring, intelligent computing, collaborative scheduling, and feedback optimization.

[0025] System operation begins with comprehensive and continuous status monitoring. Status monitoring agents deployed on each node of the cluster are responsible for collecting real-time characteristic information of multimodal data (such as data type and size), dynamic access pattern information (such as access frequency and recent access time), and real-time resource load information of computing nodes (such as CPU, memory, and network I / O utilization). These massive real-time metrics are periodically aggregated to the global resource manager, providing a solid data foundation for subsequent intelligent decision-making.

[0026] After data acquisition is complete, the system's popularity evaluator begins operation. It standardizes the raw data and applies a comprehensive popularity evaluation model to calculate the real-time popularity value for each data object. This model is a multi-dimensional weighted formula that comprehensively considers the inherent weight of data type, normalized access frequency, data size (reflecting the system's focus on small data blocks in reciprocal form), and the priority of associated computational tasks. The calculated popularity values ​​are updated in real-time to the metadata service, becoming a shared and crucial basis for guiding resource scheduling throughout the system.

[0027] Based on the popularity value, the storage scheduler begins intelligent data layout management. The system presets corresponding popularity threshold ranges for different performance storage tiers (such as memory, high-speed SSDs, and large-capacity HDDs). The storage scheduler compares the current popularity value of a data object with these ranges, automatically determines its appropriate storage tier, and triggers necessary data migration operations. Simultaneously, a dynamic migration mechanism runs periodically, continuously adjusting the data layout based on changes in data popularity values ​​to ensure that high-popularity data is always located on high-speed storage media, thereby achieving efficient utilization of storage resources.

[0028] When a computation task is submitted, the computation scheduler performs heat-aware collaborative scheduling. It first parses the task's data requirements and retrieves the real-time heat and physical location of the required data blocks from the metadata service. Then, the scheduler makes decisions based on the core objectives of "optimal data locality" and "system load balancing": it prioritizes nodes that already have the data stored locally; if none are available, it comprehensively evaluates network transmission overhead (which is positively correlated with data heat) and the current load of the target node to select a globally optimal node. For tasks accessing high-heat data, the system assigns them higher scheduling priority to ensure that critical business processes are handled promptly.

[0029] Finally, the system achieves self-learning and continuous improvement through a feedback optimization mechanism. The system monitors the actual execution efficiency of completed tasks. If it finds that performance is substandard due to data access latency or network transmission bottlenecks, it feeds this information back to the heat assessment evaluator. The evaluator then makes dynamic adjustments, such as temporarily increasing the heat of the data block causing the bottleneck or progressively fine-tuning the weight parameters in the heat assessment model. This closed-loop feedback mechanism allows the system to learn from historical experience, continuously optimize its scheduling strategy, and adapt to dynamically changing workloads, demonstrating strong adaptive capabilities.

[0030] In summary, this invention constructs an intelligent, collaborative, and adaptive resource management ecosystem through a series of interconnected and cyclical steps: "monitoring—computation—storage scheduling—computation scheduling—feedback." It effectively links data popularity with the allocation of computing resources, ultimately significantly improving system throughput, response speed, and resource utilization efficiency in complex multimodal data application scenarios.

[0031] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for hotspot-aware hierarchical storage and computing resource scheduling for multimodal data, characterized in that, Includes the following steps: S1. Data and system status monitoring steps: Real-time collection of characteristic information of multimodal data, data access mode information, and resource load information of computing nodes; S2, Comprehensive popularity value calculation steps: Based on the information collected in step S1, a comprehensive popularity evaluation model is used to dynamically calculate the real-time comprehensive popularity value of each data object; S3. Storage Tier Decision and Data Scheduling Steps: Based on the comprehensive heat value calculated in step S2, the data object is automatically scheduled to the corresponding tier in the hierarchical storage system composed of storage media with different performance. S4. Computation Task Co-scheduling Steps: When scheduling computation tasks, the overall heat value of the data objects required by the task and their storage level are simultaneously perceived. With the goals of optimal data locality and system load balancing, the computation task is scheduled to be executed on the most suitable computing node.

2. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 1, characterized in that, In step S1, the feature information of the multimodal data includes data type and data size; the data access pattern information includes access frequency and recent access time; and the resource load information of the computing node includes CPU utilization, memory utilization, and network I / O load.

3. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 2, characterized in that, In step S2, when calculating the comprehensive popularity value, the comprehensive popularity evaluation model takes into account at least the data type weight, the normalized access frequency, the data size, and the priority of the associated calculation tasks.

4. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 3, characterized in that, The comprehensive popularity assessment model described in step S2 calculates the comprehensive popularity value H of the data object using the following formula: H = α * W_type + β * F_freq + γ * (1 / S_size) + δ * P_task Where W_type is a preset weight factor based on data type, F_freq is the normalized access frequency, S_size is the data size, P_task is the priority weight of the associated computation task, and α, β, γ, and δ are configurable weight coefficients for each dimension, and α+β+γ+δ=1.

5. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 1, characterized in that, The storage tier decision in step S3 specifically involves: presetting a popularity threshold range for each tier of the hierarchical storage system; comparing the overall popularity value of a data object with the threshold range to determine the storage tier to which it should be placed; wherein, the data object with the highest popularity value is placed in the storage tier with the fastest access speed.

6. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 5, characterized in that, Step S3 also includes a dynamic data migration mechanism: periodically recalculate the overall popularity value of all data objects. When the current popularity value of a data object does not match the popularity threshold range corresponding to its current storage level, the system automatically triggers a data migration operation to migrate it to a storage level that matches the current popularity value.

7. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 1, characterized in that, In step S4, with the goal of achieving optimal data locality and system load balancing, the scheduler prioritizes selecting computing nodes that have already locally stored copies of the required data. When no node satisfies data locality, the scheduler selects the computing node with the shortest network topology distance, the lowest network overhead, and the lowest current combined CPU and memory load.

8. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 1, characterized in that, In step S4, when scheduling computing tasks, tasks whose required data is determined to be of high demand are given a higher scheduling priority, and computing resources are allocated to them first.

9. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 1, characterized in that, The method also includes a feedback optimization step S5: monitoring the actual execution efficiency of the completed computation task. If the task execution efficiency is lower than the expected threshold due to data access delay or network transmission bottleneck, the information is fed back to the comprehensive heat evaluation model in step S2 for dynamic adjustment of relevant parameters or future scheduling strategies for similar tasks and data.

10. The hotspot-aware hierarchical storage and computing resource scheduling method for multimodal data according to claim 9, characterized in that, In the feedback optimization step S5, based on the feedback of task execution efficiency, the configurable weight coefficients α, β, γ, and δ in the comprehensive heat evaluation model are dynamically adjusted, or the heat value of the data object causing the bottleneck is increased, so as to prompt the system to make better decisions in subsequent scheduling.

Images