Search

CN-121996411-A - Model scheduling method based on edge equipment, edge equipment and storage medium

CN121996411ACN 121996411 ACN121996411 ACN 121996411ACN-121996411-A

Abstract

The application discloses a model scheduling method based on edge equipment, the edge equipment and a storage medium, and relates to the technical field of data processing, wherein the method comprises the steps of responding to a task to be processed in the edge equipment, and determining task complexity of the task to be processed; detecting a computing power index of the edge equipment, determining a computing power state representing computing power utilization rate in the edge equipment according to the computing power index, wherein the computing power index comprises CPU utilization rate, GPU utilization rate, memory occupancy rate and residual electric quantity of the edge equipment, selecting a target model from at least two models with different computing power requirements according to task complexity and/or computing power state, and scheduling the target model to execute a task to be processed. The application realizes the dynamic scheduling of the edge equipment aiming at the model.

Inventors

  • AN KANG
  • CHEN QIANG
  • WANG MINGHUI

Assignees

  • 歌尔股份有限公司

Dates

Publication Date
20260508
Application Date
20251229

Claims (10)

  1. 1. The model scheduling method based on the edge equipment is characterized by comprising the following steps of: Responding to a task to be processed in the edge equipment, and determining task complexity of the task to be processed; Detecting a computing power index of the edge equipment, and determining a computing power state representing computing power utilization rate in the edge equipment according to the computing power index, wherein the computing power index comprises CPU utilization rate, GPU utilization rate, memory occupancy rate and residual electric quantity of the edge equipment; And selecting a target model from at least two models with different computational power demands according to the task complexity and/or the computational power state, and scheduling the target model to execute the task to be processed.
  2. 2. The edge device-based model scheduling method of claim 1, wherein the models include a lightweight model, a standard model, and an enhanced model, the enhanced model having a computational power demand greater than that of the standard model, the standard model having a computational power demand greater than that of the lightweight model.
  3. 3. The edge device-based model scheduling method of claim 2, wherein the step of selecting a target model among models of at least two different computational power demands in accordance with the task complexity and the computational power states comprises: Comparing the calculated force state value of the calculated force state with a preset first calculated force threshold value; if the calculation force state value is smaller than the first calculation force threshold value, selecting the light-weight model or the standard model as a target model when the task complexity is low; when the task complexity is middle complexity, selecting the standard version model as a target model; And when the task complexity is high, selecting the enhancement model as a target model.
  4. 4. The edge device-based model scheduling method of claim 3, further comprising, after the step of comparing the calculated force state value of the calculated force state with a preset first calculated force threshold value: if the calculated force state value is larger than or equal to the first calculated force threshold value, detecting whether the calculated force state value is smaller than or equal to a preset second calculated force threshold value; if the calculation force state value is smaller than or equal to the second calculation force threshold value, selecting the standard version model as a target model when the task complexity is low complexity or medium complexity; And when the task complexity is high, selecting the lightweight model as a target model.
  5. 5. The edge device-based model scheduling method of claim 4, wherein after the step of detecting whether the calculated force state value is less than or equal to a preset second calculated force threshold value, further comprising: and if the calculated force state value is larger than the second calculated force threshold value, selecting the lightweight model as a target model.
  6. 6. The edge device-based model scheduling method of any one of claims 1 to 4, wherein the step of determining task complexity of the task to be processed comprises: Determining that the task complexity of the task to be processed is low in response to the task to be processed representing single task processing; responding to the task to be processed to represent a plurality of task processes of the same type, and determining the task complexity of the task to be processed as medium complexity; And determining that the task complexity of the task to be processed is high in response to the task to be processed characterizing a plurality of different types of task processing.
  7. 7. The edge device-based model scheduling method of any one of claims 1 to 4, wherein the step of determining a computational effort state in the edge device that characterizes computational effort usage in accordance with the computational effort metrics comprises: Respectively carrying out standardization processing on the CPU utilization rate, the GPU utilization rate, the memory occupancy rate and the residual electric quantity to respectively obtain a first standard value corresponding to the CPU utilization rate, a second standard value corresponding to the GPU utilization rate, a third standard value corresponding to the memory occupancy rate and a fourth standard value corresponding to the residual electric quantity; And carrying out weighted summation on the first standard value, the second standard value, the third standard value and the fourth standard value to obtain a calculated force state value representing the calculated force state of the edge equipment.
  8. 8. The edge device-based model scheduling method of any one of claims 1 to 4, wherein after the step of determining a computational effort state in the edge device that characterizes computational effort usage in accordance with the computational effort metrics, further comprising: determining the computational power demand of a lightweight model in at least two models with different computational power demands as a first computational power demand; When the task complexity of the task to be processed is middle complexity or high complexity and the calculation force state value of the calculation force state is smaller than a preset first calculation force threshold value, performing task decomposition on the task to be processed according to the first calculation force demand to obtain a first subtask matched with the first calculation force demand and a second subtask except the first subtask; And scheduling a lightweight model to execute the first subtask in the models with at least two different computational demands, and scheduling a standard version model to execute the second subtask.
  9. 9. An edge device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the edge device based model scheduling method of any one of claims 1 to 8.
  10. 10. A storage medium, characterized in that the storage medium is a computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the steps of the edge device based model scheduling method according to any one of claims 1 to 8.

Description

Model scheduling method based on edge equipment, edge equipment and storage medium Technical Field The present application relates to the field of data processing technologies, and in particular, to a model scheduling method based on edge devices, an edge device, and a storage medium. Background With the rapid development of internet of things technology, edge devices (such as wearable devices, vehicle-mounted terminals, etc.) have become important carriers of AI technologies, such as audio AI technologies (e.g., automatic speech recognition ASR, voice activity detection VAD, voiceprint recognition, etc.). The inherent characteristics of limited computing power and large resource fluctuation of the edge equipment generally exist, the CPU/GPU performance, the memory capacity and the battery endurance are far lower than those of a cloud server, and the conventional audio AI model is designed in a unified specification and lacks of adaptability to the edge equipment. And when the edge equipment performs task processing by using a fixed scheduling model (such as an audio AI model), flexible scheduling cannot be performed according to a specific actual scene, so that a model with complete functions cannot be normally operated by the equipment, and further a task execution failure phenomenon is caused. Therefore, how to implement dynamic scheduling of edge devices for models is an urgent problem to be solved at present. The foregoing is provided merely for the purpose of facilitating understanding of the technical solutions of the present application and is not intended to represent an admission that the foregoing is prior art. Disclosure of Invention The application mainly aims to provide a model scheduling method based on edge equipment, the edge equipment and a storage medium, and aims to solve the technical problem of how to realize dynamic scheduling of the edge equipment for a model. In order to achieve the above object, the present application provides a model scheduling method based on edge equipment, the model scheduling method based on edge equipment includes the following steps: responding to a task to be processed in the edge equipment, and determining task complexity of the task to be processed; detecting a computing power index of the edge equipment, and determining a computing power state representing computing power utilization rate in the edge equipment according to the computing power index, wherein the computing power index comprises CPU utilization rate, GPU utilization rate, memory occupancy rate and residual electric quantity of the edge equipment; And selecting a target model from at least two models with different computational power demands according to the task complexity and/or the computational power state, and scheduling the target model to execute the task to be processed. Optionally, the model comprises a lightweight model, a standard model, and an enhanced model, the computational power requirement of the enhanced model is greater than the computational power requirement of the standard model, and the computational power requirement of the standard model is greater than the computational power requirement of the lightweight model. Optionally, the step of selecting the target model among the models of at least two different computational demands in dependence on the task complexity and the computational power status comprises: Comparing the calculated force state value of the calculated force state with a preset first calculated force threshold value; If the calculated force state value is smaller than the first calculated force threshold value, selecting a light-weight model or a standard model as a target model when the task complexity is low; when the task complexity is middle complexity, selecting a standard version model as a target model; And when the task complexity is high, selecting the enhancement model as a target model. Optionally, after the step of comparing the calculated force state value of the calculated force state with the preset first calculated force threshold value, the method further comprises: If the calculated force state value is larger than or equal to the first calculated force threshold value, detecting whether the calculated force state value is smaller than or equal to a preset second calculated force threshold value; if the calculated force state value is smaller than or equal to the second calculated force threshold value, selecting a standard version model as a target model when the task complexity is low complexity or medium complexity; and when the task complexity is high, selecting the lightweight model as the target model. Optionally, after the step of detecting whether the calculated force state value is less than or equal to the preset second calculated force threshold value, the method further includes: and if the calculated force state value is larger than the second calculated force threshold value, selecting the lightweight model as the ta