Intelligent resource scheduling and optimizing method and device for terminal equipment

Abstract

The invention relates to the field of edge computing, and particularly provides an intelligent resource scheduling and optimizing method and device for terminal equipment, and the method comprises the following steps: S1, system initialization and resource monitoring; s2, edge-cloud collaborative prediction decision making; s3, dynamic resource scheduling and execution; and S4, performing effect feedback and closed-loop optimization. Compared with the prior art, the method has the advantages that the use condition of equipment hardware resources can be monitored in real time, resource requirements are predicted in combination with the AI algorithm, and calculation, storage and dynamic distribution and optimization of network resources are achieved.

Description

Technical Field

[0001] The present invention relates to the field of edge computing, and specifically provides a method and apparatus for intelligent resource scheduling and optimization of terminal devices. Background Art

[0002] Current terminal device resource management has the following problems:

[0003] (1) The resource allocation strategy is fixed and cannot adapt to the dynamic changes in demand in different application scenarios;

[0004] (2) Lack of comprehensive awareness of the overall status of the equipment, and resource scheduling decisions are based on local information;

[0005] (3) Extensive energy consumption management makes it difficult to achieve the best balance between performance and battery life;

[0006] (4) Resource conflicts are serious when multiple tasks are performed in parallel, affecting user experience;

[0007] (5) Traditional solutions rely on cloud-based decision-making, resulting in large response delays and insufficient privacy protection.

[0008] As terminal device performance improves and application scenarios diversify, users are placing higher demands on device resource management. Existing technologies, such as the Linux kernel scheduler and Android resource manager, can perform basic resource allocation but lack intelligence and adaptability, making it difficult to meet the demands of modern terminal devices for efficient and intelligent resource management. Summary of the Invention

[0009] The present invention aims to address the deficiencies of the above-mentioned prior art and provides a highly practical method for intelligent resource scheduling and optimization of terminal devices.

[0010] A further technical task of the present invention is to provide a rationally designed, safe and applicable intelligent resource scheduling and optimization device for terminal devices.

[0011] The technical solution adopted by the present invention to solve its technical problem is:

[0012] A method for intelligent resource scheduling and optimization of terminal equipment has the following steps: S1, system initialization and resource monitoring;

[0013] S2, edge-cloud collaborative prediction and decision-making;

[0014] S3, dynamic resource scheduling and execution;

[0015] S4. Effect feedback and closed-loop optimization.

[0016] Furthermore, in step S1, when the system starts, the hardware abstraction interface layer is first deployed on the terminal device to achieve unified access control of heterogeneous hardware resources, including dynamic adjustment of CPU or GPU frequency, real-time collection of memory and network status, and monitoring of hardware indicators through the operating system kernel interface. The collected raw data is normalized by sliding windows and filtered for outliers to form a time-series feature data stream.

[0017] Furthermore, the hardware abstraction interface layer obtains the underlying hardware parameters through the operating system kernel interface, encapsulates the dynamic voltage and frequency scaling (DVFS) function, calls the provided power management API, and performs a standardized data reporting protocol.

[0018] When monitoring the CPU, the difference between the total CPU time and the idle time is calculated by parsing the proc and stat files. The formula is:

[0019] CPU utilization = (1-Δ idle time / Δ total time) × 100%;

[0020] Distinguish load statistics between performance cores and energy efficiency cores;

[0021] When monitoring the GPU, obtain the utilization, temperature, and video memory usage indicators. For mobile devices, obtain the rendering frame rate through the AndroidSurfaceFlinger interface.

[0022] The sliding window calculates the mean and standard deviation in real time. If the current value exceeds the mean ± 3σ, it is marked as abnormal and triggers data cleaning.

[0023] Furthermore, in step S2, a lightweight AI model cluster is deployed on the edge node. After receiving the feature data reported by the terminal, the resource demand for a period of time in the future is predicted using the LSTM model. The type of emergency task is identified by combining the random forest classifier. When the local model prediction error exceeds the threshold, the federated learning mechanism is triggered. The terminal uploads the desensitized data processed with differential privacy to the cloud, which aggregates the data to generate a global model and sends it to the edge node.

[0024] At the same time, edge nodes analyze user behavior, dynamically adjust feature weights, and generate multi-objective optimization recommendations.

[0025] Furthermore, the input layer of the LSTM model is historical data of 10 time steps, the hidden layer is a 2-layer LSTM unit with 64 neurons in each layer, using a tanh activation function, and the output layer is a Softmax classifier that outputs a low, medium, or high load probability distribution;

[0026] The random forest classifier screens the top 20% features based on the Gini impurity reduction rate, with a maximum depth of 5 and a minimum number of leaf samples of 50.

[0027] Furthermore, in step S3, based on the prediction results and optimization goals, the NSGA-II algorithm is used to solve the Pareto optimal solution for resource allocation. The execution layer adjusts hardware parameters in real time through the underlying API, binds high-priority processes to the CPU large core cluster, sets the GPU rendering queue, and allocates network QoS channels.

[0028] For abnormal scenarios, frequency reduction, task migration or deep power saving strategies are triggered, and all operations are authorized through the trusted execution environment.

[0029] Furthermore, in step S4, after each resource adjustment, the system records the performance indicators and updates the local model weights through the gradient descent method. The edge nodes regularly synchronize the anonymized effect data to the cloud and participate in the global model iteration of federated learning. At the same time, the terminal automatically switches the scheduling strategy according to the changes in the user scenario, forming a continuous closed loop of monitoring → prediction → execution → feedback.

[0030] A terminal device intelligent resource scheduling and optimization device, comprising: at least one memory and at least one processor;

[0031] The at least one memory is configured to store a machine-readable program;

[0032] The at least one processor is used to call the machine-readable program to execute a terminal device intelligent resource scheduling and optimization method.

[0033] Compared with the prior art, the terminal device intelligent resource scheduling and optimization method and device of the present invention have the following outstanding beneficial effects:

[0034] This invention implements a unified resource abstraction layer across hardware platforms, solving the problem of ecological fragmentation. It is compatible with existing hardware and operating systems, has low deployment costs, and is easy to promote. It proposes a lightweight model based on federated learning; an update mechanism that balances privacy and model effectiveness; significantly improves terminal device performance and user experience, increasing average response speed by more than 30%; optimizes battery efficiency, extending device life by 15%-25%; reduces resource conflicts and system lag, improves multitasking capabilities, and adapts to different usage scenarios and user habits without manual adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1It is a flow chart of a method for intelligent resource scheduling and optimization of terminal devices;

[0037] Figure 2 It is an architectural diagram of a method for intelligent resource scheduling and optimization of terminal devices. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0039] A best embodiment is given below:

[0040] like Figure 1 、 2 As shown, a terminal device intelligent resource scheduling and optimization method in this embodiment has the following steps:

[0041] S1, system initialization and resource monitoring;

[0042] During system startup, a hardware abstraction interface layer (HAIL) is deployed on the terminal device to implement unified access control for heterogeneous hardware resources, including dynamic CPU / GPU frequency adjustment and real-time memory and network status monitoring. Hardware metrics are continuously monitored at a 500ms interval through operating system kernel interfaces (such as Linux / proc and Android Binder), covering computing load, storage I / O, network quality, and battery status. The collected raw data is normalized using a sliding window and filtered for outliers (using the 3σ principle) to form a time-series feature data stream, providing the foundation for subsequent analysis.

[0043] The hardware abstraction interface layer obtains the underlying hardware parameters through the operating system kernel interface (such as Linux sysfs and Windows WMI), encapsulates the dynamic voltage and frequency scaling (DVFS) function, calls the power management API provided by the chip manufacturer (such as Intel RAPL and ARM PSCI), implements the standardized data reporting protocol, supports a configurable sampling frequency of 100ms-1s, and the data format is:

[0044] "timestamp":"2025-03-20T14:30:00Z",

[0045] "cpu_usage":65.2,

[0046] "gpu_temp":42.3,

[0047] "network_rtt":28.5.

[0048] When monitoring the CPU, the difference between the total CPU time and the idle time is calculated by parsing the / proc / stat file. The formula is:

[0049] CPU utilization = (1-Δ idle time / Δ total time) × 100%;

[0050] Supports multi-core heterogeneous architectures (such as ARM big.LITTLE) and distinguishes between load statistics of performance cores and energy-efficiency cores.

[0051] When monitoring the GPU, call the manufacturer's SDK (such as NVIDIA NVML and AMD ROCm) to obtain indicators such as utilization, temperature, and video memory usage. For mobile GPUs (such as Adreno), obtain the rendering frame rate through the Android SurfaceFlinger interface.

[0052] Use the Linux tc toolchain or the Windows QoS API to measure bandwidth, latency, and jitter, implement active network quality detection based on UDP probes, and calculate the packet loss rate formula:

[0053] Packet loss rate = number of lost packets / number of sent packets × 100%.

[0054] A sliding window (window size = 10, step size = 1) is used to calculate the mean and standard deviation in real time. If the current value exceeds the mean ± 3σ, it is marked as an anomaly and triggers data cleaning.

[0055] S2, edge-cloud collaborative prediction and decision-making;

[0056] A lightweight AI model cluster is deployed on edge nodes. After receiving feature data reported by terminals, it uses an LSTM model to predict resource requirements for the next 5-10 seconds and combines it with a random forest classifier to identify the type of emergent tasks. When the local model prediction error exceeds a threshold, the federated learning mechanism is triggered: the terminal uploads desensitized data processed with differential privacy to the cloud, which aggregates it to generate a global model and sends it to the edge node. Simultaneously, the edge node analyzes user behavior (such as high-frequency touch and application switching paths), dynamically adjusts feature weights, and generates multi-objective optimization recommendations.

[0057] The input layer of the LSTM model is historical data of 10 time steps (CPU, GPU, and network indicators form a five-dimensional feature). The hidden layer consists of two layers of LSTM units, each with 64 neurons. The tanh activation function is used. The output layer is a Softmax classifier that outputs a low / medium / high load probability distribution. Curriculum learning is used to gradually increase the complexity of the time series.

[0058] In the random forest classifier, the top 20% features are selected based on the Gini impurity reduction rate, with a maximum depth of 5 and a minimum number of leaf samples of 50 to prevent overfitting.

[0059] Each terminal device in the federated learning mechanism uses local data to update the model parameters θ i ,

[0060] Cloud performs weighted average:

[0061]

[0062] Among them D i is the data volume of device i.

[0063] Map FP32 weights to INT8 range, using symmetric quantization formula,

[0064]

[0065] Use the teacher model (ResNet-50) to guide the training of the lightweight student model (MobileNetV3).

[0066] S3, dynamic resource scheduling and execution;

[0067] Based on the prediction results and optimization goals, the NSGA-II algorithm is used to solve the Pareto optimal solution for resource allocation. The execution layer adjusts hardware parameters in real time through the underlying API: binding high-priority processes to the CPU large core cluster, setting the GPU rendering queue, and allocating network QoS channels. For abnormal scenarios (such as temperature > 45°C or battery < 20%), frequency reduction, task migration, or deep power saving strategies (disabling Bluetooth / NFC, limiting background I / O) are triggered. All operations are authorized through the Trusted Execution Environment (TEE) to ensure security.

[0068] S4, effect feedback and closed-loop optimization;

[0069] After each resource adjustment, the system records performance indicators (frame rate stability, temperature changes, energy consumption ratio) and updates local model weights using gradient descent. Edge nodes regularly synchronize anonymized performance data to the cloud to participate in the global model iteration of federated learning. At the same time, the terminal automatically switches scheduling strategies based on changes in user scenarios (such as switching from game to standby), forming a continuous optimization closed loop of "monitoring → prediction → execution → feedback", ultimately achieving the technical effect of increasing resource utilization by 30% and extending battery life by 25%.

[0070] Among them, CPU scheduling binds processes to specified cores through cpuset and adjusts CPU Governor parameters (such as setting scaling_governor to performance);

[0071] In GPU scheduling, call nvmlDeviceSetApplicationsClocks to set the memory frequency, use VulkanAPI to allocate an exclusive rendering queue, and set traffic priority through Linux tc:

[0072] tc qdisc add dev eth0 root handle 1:htb

[0073] tc class add dev eth0 parent 1:classid 1:1htb rate 100Mbit ceil100Mbit

[0074] tc filter add dev eth0 protocol ip parent 1:0prio 1u32match ip dport80 0xffff flowid 1:1.

[0075] The above-mentioned specific implementation methods are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementation methods. Any technical solutions that conform to the above-mentioned specific implementation methods of the present invention and any appropriate changes or substitutions made thereto by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0076] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent resource scheduling and optimization of terminal equipment, characterized in that: The method has the following steps: S1, system initialization and resource monitoring; S2, edge-cloud collaborative prediction and decision-making; S3, dynamic resource scheduling and execution; S4. Effect feedback and closed-loop optimization.

2. A terminal device intelligent resource scheduling and optimization method according to claim 1, characterized in that: In step S1, when the system starts, the hardware abstraction interface layer is first deployed on the terminal device to achieve unified access control of heterogeneous hardware resources, including dynamic adjustment of CPU or GPU frequency, real-time collection of memory and network status, and monitoring of hardware indicators through the operating system kernel interface. The collected raw data is normalized by sliding windows and filtered for outliers to form a time-series feature data stream.

3. A terminal device intelligent resource scheduling and optimization method according to claim 2, characterized in that: The hardware abstraction interface layer obtains the underlying hardware parameters through the operating system kernel interface, encapsulates the dynamic voltage and frequency scaling (DVFS) function, calls the provided power management API, and performs standardized data reporting protocols; When monitoring the CPU, the difference between the total CPU time and the idle time is calculated by parsing the proc and stat files. The formula is: CPU utilization = (1-Δ idle time / Δ total time) × 100%; Distinguish load statistics between performance cores and energy efficiency cores; When monitoring the GPU, obtain the utilization, temperature, and video memory usage indicators. For mobile devices, obtain the rendering frame rate through the AndroidSurfaceFlinger interface. The sliding window calculates the mean and standard deviation in real time. If the current value exceeds the mean ± 3σ, it is marked as abnormal and triggers data cleaning.

4. A terminal device intelligent resource scheduling and optimization method according to claim 3, characterized in that: In step S2, a lightweight AI model cluster is deployed on the edge node. After receiving the feature data reported by the terminal, the LSTM model is used to predict the resource demand for a period of time in the future. The random forest classifier is combined to identify the type of sudden task. When the local model prediction error exceeds the threshold, the federated learning mechanism is triggered. The terminal uploads the desensitized data processed with differential privacy to the cloud. The cloud aggregates and generates a global model and sends it to the edge node. At the same time, edge nodes analyze user behavior, dynamically adjust feature weights, and generate multi-objective optimization recommendations.

5. A terminal device intelligent resource scheduling and optimization method according to claim 4, characterized in that: The input layer of the LSTM model is historical data of 10 time steps, the hidden layer is a 2-layer LSTM unit with 64 neurons in each layer, using the tanh activation function, and the output layer is a Softmax classifier that outputs a low, medium, or high load probability distribution; The random forest classifier screens the top 20% features based on the Gini impurity reduction rate, with a maximum depth of 5 and a minimum number of leaf samples of 50.

6. A terminal device intelligent resource scheduling and optimization method according to claim 5, characterized in that: In step S3, based on the prediction results and optimization goals, the NSGA-II algorithm is used to solve the Pareto optimal solution for resource allocation. The execution layer adjusts hardware parameters in real time through the underlying API, binds high-priority processes to the CPU large-core cluster, sets the GPU rendering queue, and allocates network QoS channels. For abnormal scenarios, frequency reduction, task migration or deep power saving strategies are triggered, and all operations are authorized through the trusted execution environment.

7. A terminal device intelligent resource scheduling and optimization method according to claim 6, characterized in that: In step S4, after each resource adjustment, the system records performance indicators and updates the local model weights through the gradient descent method. The edge nodes regularly synchronize anonymized effect data to the cloud and participate in the global model iteration of federated learning. At the same time, the terminal automatically switches the scheduling strategy according to changes in user scenarios, forming a continuous closed loop of monitoring → prediction → execution → feedback.

8. A terminal device intelligent resource scheduling and optimization device, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 7.

Images