Scheduling optimization method and system applied to AI edge computing server

Abstract

The present invention provides a scheduling optimization method and system for AI edge computing servers, which relate to the field of edge computing technology and include a pre-built edge service cluster. The method includes evenly dividing the resources of the task to be processed based on the first required resource amount of the task to be processed and the number of enabled edge servers in the edge server cluster to obtain subtasks and their required second required resource amounts; establishing a hierarchical mapping table of subtasks and edge servers based on dynamic weighting of selected matching indicators; based on the hierarchical mapping table, performing resource evaluation on the edge servers corresponding to the subtasks according to the second required resource amounts, and dynamically scheduling the subtasks according to the evaluation results. The dynamic scheduling includes allocating the subtasks to the edge servers that meet the resource requirements for execution, or migrating the subtasks when resources are insufficient. The present invention realizes more efficient and reasonable resource allocation and task scheduling, and improves resource utilization and task execution efficiency.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and in particular to a scheduling optimization method and system applied to an AI edge computing server. Background Art

[0002] With the rapid development of artificial intelligence (AI), edge computing, a distributed computing paradigm, is becoming a key technology for supporting large-scale, low-latency AI applications. Edge computing deploys computing resources at the edge of the network, enabling data processing close to the source, significantly reducing data transmission latency, improving response speeds, and alleviating the burden on core networks. However, resource allocation and task scheduling in edge computing environments remain a significant challenge, especially given the growing demand for computing resources in AI applications.

[0003] AI edge computing servers typically handle diverse computing tasks, including but not limited to deep learning model inference, image processing, and natural language processing. These tasks often require significant computing resources, memory resources, and network bandwidth to ensure completion within a reasonable timeframe. However, resources in edge computing environments are limited and unevenly distributed. Therefore, efficiently and rationally allocating these resources to meet the real-time and accuracy requirements of AI applications has become a pressing issue.

[0004] Traditional resource allocation and task scheduling methods are often based on static rules or algorithms, such as round-robin scheduling and priority scheduling. Although simple and easy to implement, these methods fail to fully utilize the dynamic and diverse nature of edge computing environments, resulting in low resource utilization and increased task execution delays.

[0005] Therefore, it is necessary to provide a scheduling optimization method and system for AI edge computing servers to solve the above technical problems. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a scheduling optimization method and system for AI edge computing servers. By comprehensively considering factors such as hardware compatibility, data proximity values, and real-time resource evaluation, more efficient and reasonable resource allocation and task scheduling are achieved, thereby improving resource utilization and task execution efficiency.

[0007] The present invention provides a scheduling optimization method for an AI edge computing server, including a pre-built edge service cluster, the method comprising the following steps:

[0008] Based on the first resource requirement of the task to be processed and the number of enabled edge servers in the edge server cluster, evenly divide the resources of the task to be processed to obtain subtasks and their second resource requirements;

[0009] Establishing a hierarchical mapping table between subtasks and edge servers based on dynamically weighted selected matching indicators, wherein the matching indicators are hardware compatibility and data proximity between the subtasks and edge servers, wherein each subtask in the hierarchical mapping table is mapped to at least two edge servers with a priority, and is used to guide the execution of the subtask on the mapped edge servers;

[0010] Based on the hierarchical mapping table, a resource evaluation is performed on the edge server corresponding to the subtask according to the second required resource amount, and the subtask is dynamically scheduled according to the evaluation result. The dynamic scheduling includes allocating the subtask to an edge server that meets the resource requirements for execution, or migrating the subtask when resources are insufficient.

[0011] Preferably, the even division of resources includes:

[0012] Obtaining a first required resource amount of the task to be processed and the number of currently enabled edge servers in the edge server cluster, wherein the first required resource amount is a total computing resource amount required to execute the task to be processed;

[0013] Based on the first required resource amount of the task to be processed and the number of edge servers currently enabled in the edge server cluster, the task to be processed is divided according to the constraint condition that the number of tasks matches the number of enabled edge servers to obtain multiple subtasks, and the second required resource amount required for each subtask is calculated, where the second required resource amount is the sub-computing resource amount required to execute the subtask.

[0014] Preferably, the dynamic weighting includes performing the following operations for each subtask:

[0015] Calculate the hardware compatibility score and data proximity score of the subtask and each edge server;

[0016] According to the values of the hardware compatibility score and the data proximity score, the corresponding hardware compatibility interval and data proximity interval in the predefined weight distribution mapping table are matched to obtain the corresponding weight value;

[0017] Calculating a comprehensive matching value based on the hardware compatibility score and the data proximity score and their respective weight values, and sorting the edge servers from high to low according to the comprehensive matching value;

[0018] Each subtask is mapped to at least two edge servers with the highest comprehensive matching values, and the execution priority is set according to the comprehensive matching values to generate a hierarchical mapping table.

[0019] Preferably, the weight distribution mapping table is a static weight distribution mapping table established based on a preset combination of hardware compatibility intervals and data proximity value intervals, wherein each set of hardware compatibility intervals and data proximity value intervals in the weight distribution mapping table corresponds to a set of qualified weight values.

[0020] Preferably, the calculation of the hardware compatibility score includes the following steps:

[0021] Matching hardware configuration parameters of the edge server based on the type of computing resources required by the subtask, where the hardware configuration parameters include at least one of a processor model, memory capacity, and accelerator type;

[0022] The matching metrics of the hardware configuration parameters are quantified into a hardware compatibility score through a preset compatibility scoring model, where the range of the hardware compatibility score is [0, 1].

[0023] Preferably, the calculation of the data proximity value score includes the following steps:

[0024] Obtain the data storage location and physical node location of the edge server required for processing the subtask;

[0025] Calculating the number of hops for data transmission based on the network topology relationship between the data storage location and the physical node of the edge server;

[0026] The hop count is mapped to a data proximity score through a preset proximity scoring model, where the hop count is negatively correlated with the data proximity score, and the range of the data proximity score is [0, 1].

[0027] Preferably, the resource assessment comprises the following steps:

[0028] Obtain in real time the current remaining available resources of the edge server with the highest priority mapped to the subtask, including computing resources, memory resources, and network bandwidth resources;

[0029] Based on the second required resource amount, determining whether the current remaining available resource amount of the edge server with the highest priority meets the subtask execution requirement, and calculating the resource matching difference;

[0030] If the resource matching difference exceeds the preset threshold, the edge server is marked as resource-scarce and the subtask migration in dynamic scheduling is triggered.

[0031] Preferably, the migration operation of subtasks in the dynamic scheduling includes the following steps:

[0032] According to the priority order set in the hierarchical mapping table, the current remaining available resources of other edge servers are evaluated in turn to determine whether the second required resource amount of the subtask is met;

[0033] If there is at least one edge server whose resources can meet the computing requirements of the subtask, the subtask is assigned to the edge server for execution;

[0034] If all the corresponding edge servers in the hierarchical mapping table cannot meet the resource requirements of the subtask, a backup edge server in the edge service cluster is started, and the subtask is assigned to the backup edge server for execution.

[0035] The present invention also provides a scheduling optimization system for an AI edge computing server, which is used to execute a scheduling optimization method for an AI edge computing server, including a pre-built edge service cluster. The system includes:

[0036] A splitting module is used to evenly split the resources of the task to be processed based on the first resource requirement of the task to be processed and the number of enabled edge servers in the edge server cluster, to obtain subtasks and their second resource requirements;

[0037] a mapping module, configured to dynamically weight and establish a hierarchical mapping table of subtasks and edge servers based on selected matching indicators, wherein the matching indicators are hardware compatibility and data proximity between the subtasks and edge servers, each subtask in the hierarchical mapping table is mapped to at least two edge servers with a certain priority, and to guide the execution of the subtask on the mapped edge servers;

[0038] A scheduling module is used to perform resource evaluation on the edge server corresponding to the subtask based on the hierarchical mapping table and the second required resource amount, and dynamically schedule the subtask according to the evaluation result, wherein the dynamic scheduling includes allocating the subtask to an edge server that meets the resource requirements for execution, or migrating the subtask when resources are insufficient.

[0039] Compared with related technologies, the scheduling optimization method and system for AI edge computing servers provided by the present invention have the following beneficial effects:

[0040] The present invention first divides the resources of the task evenly based on the required resources of the task to be processed and the number of enabled edge server clusters to obtain subtasks and their required resources. Then, according to the two matching indicators of hardware compatibility and data proximity, a hierarchical mapping table of subtasks and edge servers is dynamically weighted to establish. Next, based on the hierarchical mapping table and resource evaluation results, the subtasks are dynamically scheduled, including allocating them to edge servers that meet resource requirements or migrating them when resources are insufficient. The present invention comprehensively considers multiple aspects such as resource allocation, task execution, scheduling strategies and scalability, thereby achieving efficient management and optimal utilization of edge computing resources, and providing strong support for improving the overall performance and task processing capabilities of AI edge computing servers. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a scheduling optimization method for an AI edge computing server provided by the present invention;

[0042] Figure 2 This is a module structure diagram of a scheduling optimization system applied to an AI edge computing server provided by the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all of the structures. Furthermore, the embodiments of the present invention and the features of the embodiments may be combined with one another unless there is a conflict.

[0044] It should also be noted that, for ease of description, only portions relevant to the present invention are shown in the accompanying drawings, rather than all of the contents. Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the various operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. In addition, the order of the various operations can be rearranged. The process can be terminated when its operations are completed, but may also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0045] Example 1

[0046] The present invention provides a scheduling optimization method for AI edge computing servers, including a pre-built edge service cluster, referring to Figure 1 As shown, the method includes the following steps:

[0047] S1: Based on the first required resource amount of the task to be processed and the number of enabled edge servers in the edge server cluster, evenly divide the resources of the task to be processed to obtain subtasks and their second required resource amounts.

[0048] The even resource splitting specifically includes the following steps:

[0049] S11: Obtain a first required resource amount of the task to be processed and the number of currently enabled edge servers in the edge server cluster, wherein the first required resource amount is the total computing resource amount required to execute the task to be processed.

[0050] In an edge computing environment, resource allocation before task scheduling needs to be based on the total resource requirements of the task (the first required resource amount) and the number of currently available computing nodes (the number of edge servers). The core goal of step S11 is to accurately obtain these two key parameters to provide basic data support for subsequent task segmentation.

[0051] The first required resource is the total computing resources required for task execution, including:

[0052] Computing resources: number of CPU cores, GPU model and quantity (such as NVIDIA T4 / A100), FPGA computing power unit, etc.

[0053] Memory resources: RAM capacity (e.g., 32GB), video memory requirements (e.g., 16GB HBM2).

[0054] Storage resources: local storage space (such as a 200GB SSD) and distributed storage access bandwidth.

[0055] Network resources: bandwidth requirements (e.g., 1 Gbps), latency sensitivity (e.g., <50 ms).

[0056] When a task is submitted, resource requirements are displayed through, but not limited to, APIs or configuration files; a list of active edge servers in the edge server cluster is obtained in real time through, but not limited to, Kubernetes APIs, Apache Mesos, or custom cluster management tools (such as a ZooKeeper-based node registry).

[0057] S12: Based on the first required resource amount of the task to be processed and the number of edge servers currently enabled in the edge server cluster, the task to be processed is divided according to the constraint condition that the number of tasks matches the number of enabled edge servers to obtain multiple subtasks, and the second required resource amount required for each subtask is calculated, where the second required resource amount is the sub-computing resource amount required to execute the subtask.

[0058] After determining the initial resource requirements for the task and the number of currently enabled edge servers in the edge server cluster, the task needs to be split into multiple subtasks to ensure that each subtask is evenly distributed across the edge servers. This task splitting strategy aims to balance the load across servers while meeting task execution requirements.

[0059] Specifically, the second resource requirement for each subtask is calculated based on the ratio between the first resource requirement and the number of edge servers. For example, if the task to be processed requires 10 units of computing resources and there are currently 5 edge servers available, each server should theoretically take on 2 units of computing resources.

[0060] By evenly dividing the tasks to be processed into multiple subtasks, flexibility is provided for subsequent dynamic scheduling and optimization.

[0061] S2: Establish a hierarchical mapping table between subtasks and edge servers based on dynamic weighting of selected matching indicators, wherein the matching indicators are hardware compatibility and data proximity values between the subtasks and edge servers. Each subtask in the hierarchical mapping table is mapped to at least two edge servers with priorities, and is also used to guide the execution of the subtask in the mapped edge servers.

[0062] The dynamic weighting includes performing the following operations for each subtask:

[0063] S21: Calculate the hardware compatibility score and data proximity score of the subtask and each edge server.

[0064] In an edge computing environment, the core goal of task scheduling is to assign subtasks to the most suitable edge server for execution to achieve efficient task processing and resource utilization. However, due to the large differences in hardware configuration and network topology of edge servers, directly assigning tasks may lead to performance degradation or resource waste. Therefore, before establishing a mapping relationship between subtasks and edge servers, it is necessary to quantitatively evaluate the degree of match between each subtask and the candidate edge server. Step S21 provides a basic basis for subsequent dynamic weighting and comprehensive matching by calculating the hardware compatibility score and the data proximity score.

[0065] The calculation of the hardware compatibility score includes the following steps:

[0066] According to the type of computing resources required by the subtask, the hardware configuration parameters of the edge server are matched, and the hardware configuration parameters include at least one of a processor model, a memory capacity, and an accelerator type.

[0067] In this embodiment, the computing resource types required for the subtask are first extracted, such as the number of CPU cores, GPU model and quantity, memory capacity, accelerator type, etc. At the same time, corresponding parameters are obtained from the hardware configuration information of the edge server, such as the processor model (such as Intel Xeon E5 or AMD EPYC), memory size (such as 64GB DDR4), accelerator type (such as NVIDIA A100 GPU or Xilinx FPGA), etc. By comparing these parameters, the matching degree of each resource type is calculated. For example:

[0068] If the subtask requires an NVIDIA A100 GPU and the edge server is equipped with an A100 GPU, the score for this section is full;

[0069] If the edge server is equipped with a lower-performance GPU (such as T4), a discount score is assigned based on the performance gap;

[0070] If the edge server completely lacks the required hardware (e.g. no GPU), it will receive a score of zero for this section.

[0071] In addition, for other resource types (such as the number of CPU cores, memory capacity, etc.), similar comparison logic is used to calculate the matching score of each resource type.

[0072] The matching metrics of the hardware configuration parameters are quantified into a hardware compatibility score through a preset compatibility scoring model, where the range of the hardware compatibility score is [0, 1].

[0073] In this embodiment, after hardware configuration parameter matching is completed, the matching scores of all resource types are input into a preset compatibility scoring model for quantification. This model is typically a rule-based algorithm that integrates the matching scores of multiple resource types into a comprehensive hardware compatibility score in the range [0, 1], where 1 indicates a perfect match and 0 indicates a complete mismatch.

[0074] The calculation of the data proximity value score includes the following steps:

[0075] Get the data storage location and physical node location of the edge server required for processing the subtask.

[0076] In this step, you first need to determine the data storage location for the subtask and the physical node location of the candidate edge server. Specifically:

[0077] The data storage location can be obtained through the metadata information of the distributed file system (such as HDFS, Ceph) or object storage service (such as AWS S3, Alibaba Cloud OSS), recording the storage node or area where it is located.

[0078] The physical node location of the edge server is obtained through network management tools (such as SDN controllers and network topology databases), usually represented by identifiers such as IP addresses, computer room numbers, or switch levels.

[0079] By comparing the data storage location with the physical node location of the edge server, the relative relationship between the two in the network topology can be determined, providing a basis for subsequent hop count calculation.

[0080] The number of hops for data transmission is calculated according to the network topology relationship between the data storage location and the physical node of the edge server.

[0081] In this embodiment, after obtaining the physical location, the hop count between the data storage location and the edge server is calculated based on the network topology. This hop count refers to the number of routers or switches traversed between the data storage node and the edge server node, and is typically provided by a network topology map or a link-state protocol (such as OSPF).

[0082] The hop count is mapped to a data proximity score through a preset proximity scoring model, where the hop count is negatively correlated with the data proximity score, and the range of the data proximity score is [0, 1].

[0083] Then, the hop count is mapped to the data proximity score through the preset proximity scoring model. The specific process is as follows:

[0084] The proximity scoring model is designed with a negative correlation function. The smaller the number of hops, the higher the score; the larger the number of hops, the lower the score.

[0085] For example, if the number of hops is 1, the score is 1.0; if the number of hops is 5, the score is 0.2; and if the number of hops is 10 or more, the score is 0.0. This mapping rule can intuitively reflect the impact of data proximity on task performance.

[0086] The final data proximity score range is [0, 1], where 1 indicates that the data storage location is very close to the edge server physical location, and 0 indicates that the two are far apart.

[0087] S22: According to the values of the hardware compatibility score and the data proximity score, the corresponding hardware compatibility interval and data proximity interval in the predefined weight allocation mapping table are matched to obtain the corresponding weight value.

[0088] The weight distribution mapping table is a static weight distribution mapping table established based on a preset combination of hardware compatibility intervals and data proximity value intervals, wherein each set of hardware compatibility intervals and data proximity value intervals in the weight distribution mapping table corresponds to a set of qualified weight values.

[0089] Specifically, the weight distribution mapping table construction process is as follows:

[0090] First, divide the hardware compatibility score (range [0, 1]) and the data proximity score (range [0, 1]) into several discrete intervals. For example, each score can be divided into five intervals: [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), and [0.8, 1]. These intervals can be adjusted based on actual needs to better reflect the distribution characteristics of the scores.

[0091] For each combination of hardware compatibility intervals and data proximity intervals, a set of weight values is predefined. The setting of the weight values needs to be based on actual experience or experimental results.

[0092] The score ranges and weights defined above are combined into a two-dimensional weight distribution mapping table. For example, a hardware compatibility score range of [0.6, 0.8) and a data proximity score range of [0.4, 0.6) might correspond to weights of (0.7, 0.3), indicating that hardware compatibility accounts for 70% and data proximity accounts for 30%.

[0093] After the weight distribution mapping table is constructed, the execution process of step S22 is as follows:

[0094] For each combination of subtask and candidate edge server, its hardware compatibility score and data proximity score are calculated.

[0095] According to the calculated score value, the hardware compatibility interval and data proximity value interval to which it belongs are determined.

[0096] According to the score interval combination, the corresponding weight value is found in the weight allocation mapping table.

[0097] S23: Calculate a comprehensive matching value based on the hardware compatibility score and the data proximity score and their respective weight values, and sort the edge servers from high to low according to the comprehensive matching value.

[0098] In this embodiment, the weight values found are combined with the hardware compatibility score and the data proximity score to calculate the overall matching value. For example, assuming the hardware compatibility score is 0.75 and the data proximity score is 0.5, the corresponding weight values are (0.6, 0.4).

[0099] All candidate edge servers are sorted from high to low according to the comprehensive matching value, and the execution priority is set according to the sorting result.

[0100] S24: Map each subtask to at least two edge servers with the highest comprehensive matching values, set execution priorities according to the comprehensive matching values, and generate a hierarchical mapping table.

[0101] Specifically, each subtask is mapped to the edge server with the highest comprehensive matching value and set as the preferred execution node.

[0102] At the same time, the edge server with the second highest comprehensive matching value is selected as the backup execution node, and the priority is set according to the comprehensive matching value. This multi-level mapping mechanism ensures that even if the preferred server has insufficient resources, the system can still quickly switch to the backup server to avoid task interruption.

[0103] The resulting hierarchical mapping table records the mapping relationship between each subtask and multiple edge servers and their priorities, providing a basic basis for subsequent dynamic scheduling.

[0104] S3: Based on the hierarchical mapping table, perform resource evaluation on the edge server corresponding to the subtask according to the second required resource amount, and dynamically schedule the subtask based on the evaluation result. The dynamic scheduling includes allocating the subtask to an edge server that meets the resource requirements for execution, or migrating the subtask when resources are insufficient.

[0105] The resource assessment includes the following steps:

[0106] The current remaining available resources of the edge server with the highest priority mapped to the subtask are obtained in real time, including computing resources, memory resources, and network bandwidth resources.

[0107] In this step, the preferred execution node for each subtask (i.e., the edge server with the highest priority) must be determined from the hierarchical mapping table, and then the current remaining available resources of the server must be obtained in real time. The specific process is as follows:

[0108] Use monitoring tools or cluster management platforms (such as Kubernetes and Apache Mesos) to collect edge server resource status data in real time, including computing resources (number of CPU cores and GPUs), memory resources (RAM capacity and video memory size), storage resources (local storage space and distributed storage bandwidth), and network resources (bandwidth and latency).

[0109] Based on the subtask's requirements, extract relevant resource metrics. For example, if the subtask requires high GPU computing power, focus on the edge server's GPU model and remaining available number; if the task is latency-sensitive, evaluate network bandwidth and latency performance.

[0110] Based on the second required resource amount, it is determined whether the current remaining available resource amount of the edge server with the highest priority meets the subtask execution requirement, and the resource matching difference is calculated.

[0111] After completing resource information collection, the remaining available resources of the edge server need to be compared with the second required resource amount of the subtask to determine whether the execution requirements are met and calculate the resource matching difference. The specific process is as follows:

[0112] Compare the computing resources, memory resources, storage resources, and network resources required by each subtask to see if they are less than or equal to the remaining available resources of the edge server. If all resources are met, the server is considered capable of executing the subtask; otherwise, it is marked as insufficiently resourced.

[0113] For cases where demand is not met, the difference between each resource type (required quantity minus available quantity) is calculated and aggregated into a comprehensive resource matching difference. For example, if a subtask requires 4 CPU cores and 8GB of RAM, but the edge server only has 3 CPU cores and 6GB of RAM, the difference is calculated as: Difference = (4-3) + (8-6) = 1 + 2 = 3. A larger difference indicates a more significant resource gap.

[0114] If the resource matching difference exceeds the preset threshold, the edge server is marked as resource-scarce and the subtask migration in dynamic scheduling is triggered.

[0115] In this embodiment, when the resource matching difference exceeds the preset threshold, it is necessary to take measures such as migration or starting a backup server to ensure the smooth execution of the task. The specific process is as follows:

[0116] Based on the priority set in the hierarchical mapping table, the remaining available resources of other edge servers are evaluated sequentially to determine whether a suitable alternative server exists that meets the subtask's requirements. If a suitable alternative server is found, the subtask is migrated to that server. During the migration process, data transmission overhead and network latency must be considered, and alternative servers with high proximity scores should be selected as much as possible.

[0117] If all edge servers in the hierarchical mapping table cannot meet the resource requirements of the subtask, a backup edge server in the edge service cluster is started. The backup server is usually in a standby state. After startup, it needs to initialize the relevant configuration and load the necessary data. After completion, the subtask is assigned to the server for execution.

[0118] The migration operation of subtasks in the dynamic scheduling includes the following steps:

[0119] According to the priority order set in the hierarchical mapping table, the current remaining available resources of other edge servers are evaluated in turn to determine whether the second required resource amount of the subtask is met.

[0120] During dynamic scheduling, when the preferred execution node (i.e., the edge server with the highest priority) is marked as insufficient in resources, resource evaluation will be conducted on other candidate edge servers in turn according to the priority order pre-set in the hierarchical mapping table. The specific process is as follows:

[0121] All candidate edge servers associated with the current subtask are obtained from the ranking mapping table and sorted by priority.

[0122] Through cluster management tools or monitoring systems, the current remaining available resources of each candidate edge server can be obtained in real time, including computing resources, memory resources, storage resources, and network resources.

[0123] The second required resource amount of the subtask is compared with the remaining available resource amounts of the candidate edge servers one by one to determine whether there is at least one server that can fully meet the requirement of the subtask.

[0124] If there is at least one edge server whose resources can meet the computing requirements of the subtask, the subtask is assigned to the edge server for execution.

[0125] In this embodiment, once it is found that the resources of a candidate edge server can meet the second required resource amount of the subtask, the migration operation will be immediately performed to assign the subtask to the server for execution. The specific process is as follows:

[0126] Select the server with the highest priority from the candidate servers that meet the requirements as the new execution node.

[0127] Migrate the data required for the subtask from the original server to the target server. During this process, optimize the data transmission path to reduce network latency and bandwidth consumption.

[0128] Load the necessary operating environment on the target server and start the execution process of the subtask to ensure that the task can be resumed quickly.

[0129] If all the corresponding edge servers in the hierarchical mapping table cannot meet the resource requirements of the subtask, a backup edge server in the edge service cluster is started, and the subtask is assigned to the backup edge server for execution.

[0130] In this embodiment, when all candidate edge servers cannot meet the resource requirements of the subtask, the system will start the backup edge server in the edge service cluster to ensure the continuity of the task. The specific process is as follows:

[0131] Send a startup instruction to the backup edge server to switch it from standby state to working state.

[0132] After the standby server is started, it completes the necessary initialization operations, including loading the operating system, configuring network parameters, and preparing the operating environment.

[0133] Assign subtasks to the activated standby edge server and synchronize relevant data to ensure that the tasks can be executed normally in the new environment.

[0134] Example 2

[0135] The present invention also provides a scheduling optimization system for AI edge computing servers, which is used to execute a scheduling optimization method for AI edge computing servers, including a pre-built edge service cluster, reference Figure 2 As shown, the system includes:

[0136] The splitting module 100 is configured to evenly split the resources of the task to be processed based on the first required resource amount of the task to be processed and the number of enabled edge servers in the edge server cluster, to obtain subtasks and their second required resource amounts.

[0137] A mapping module 200 is used to establish a hierarchical mapping table between subtasks and edge servers based on dynamic weighting of selected matching indicators, wherein the matching indicators are hardware compatibility and data proximity values between the subtasks and the edge servers. Each subtask in the hierarchical mapping table is mapped to at least two edge servers with priority, and is also used to guide the execution of the subtask in the mapped edge servers.

[0138] The scheduling module 300 is used to perform resource evaluation on the edge server corresponding to the subtask based on the hierarchical mapping table and the second required resource amount, and dynamically schedule the subtask according to the evaluation result. The dynamic scheduling includes allocating the subtask to an edge server that meets the resource requirements for execution, or migrating the subtask when resources are insufficient.

[0139] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0140] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program. The program can be stored in a computer-readable storage medium, including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, magnetic disk storage, or magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0141] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

Claims

1. A scheduling optimization method for an AI edge computing server, comprising a pre-built edge service cluster, characterized in that: The method comprises the following steps: Based on the first resource requirement of the task to be processed and the number of enabled edge servers in the edge server cluster, evenly divide the resources of the task to be processed to obtain subtasks and their second resource requirements; Establishing a hierarchical mapping table between subtasks and edge servers based on dynamically weighted selected matching indicators, wherein the matching indicators are hardware compatibility and data proximity between the subtasks and edge servers, wherein each subtask in the hierarchical mapping table is mapped to at least two edge servers with a priority, and is used to guide the execution of the subtask on the mapped edge servers; Based on the hierarchical mapping table, performing resource evaluation on the edge server corresponding to the subtask according to the second required resource amount, and dynamically scheduling the subtask according to the evaluation result, wherein the dynamic scheduling includes allocating the subtask to an edge server that meets the resource requirements for execution, or migrating the subtask when resources are insufficient; The dynamic weighting includes performing the following operations for each subtask: Calculate the hardware compatibility score and data proximity score of the subtask and each edge server; According to the values of the hardware compatibility score and the data proximity score, the corresponding hardware compatibility interval and data proximity interval in the predefined weight distribution mapping table are matched to obtain the corresponding weight value; Calculating a comprehensive matching value based on the hardware compatibility score and the data proximity score and their respective weight values, and sorting the edge servers from high to low according to the comprehensive matching value; Map each subtask to at least two edge servers with the highest comprehensive matching values, set execution priorities based on the comprehensive matching values, and generate a hierarchical mapping table; The calculation of the hardware compatibility score includes the following steps: Matching hardware configuration parameters of the edge server based on the type of computing resources required by the subtask, where the hardware configuration parameters include at least one of a processor model, memory capacity, and accelerator type; Through the preset compatibility scoring model, the matching metric of the hardware configuration parameters is quantified into a hardware compatibility score, where the range of the hardware compatibility score is [0, 1]. The calculation of the data proximity value score includes the following steps: Obtain the data storage location and physical node location of the edge server required for processing the subtask; Calculating the number of hops for data transmission based on the network topology relationship between the data storage location and the physical node of the edge server; The hop count is mapped to a data proximity score through a preset proximity scoring model, where the hop count is negatively correlated with the data proximity score, and the range of the data proximity score is [0, 1].

2. A scheduling optimization method applied to an AI edge computing server according to claim 1, characterized in that: The uniform resource slicing includes: Obtaining a first required resource amount of the task to be processed and the number of currently enabled edge servers in the edge server cluster, wherein the first required resource amount is a total computing resource amount required to execute the task to be processed; Based on the first required resource amount of the task to be processed and the number of edge servers currently enabled in the edge server cluster, the task to be processed is divided according to the constraint condition that the number of tasks matches the number of enabled edge servers to obtain multiple subtasks, and the second required resource amount required for each subtask is calculated, where the second required resource amount is the sub-computing resource amount required to execute the subtask.

3. The scheduling optimization method applied to an AI edge computing server according to claim 2, characterized in that: The weight distribution mapping table is a static weight distribution mapping table established based on a preset combination of hardware compatibility intervals and data proximity value intervals, wherein each set of hardware compatibility intervals and data proximity value intervals in the weight distribution mapping table corresponds to a set of qualified weight values.

4. The scheduling optimization method for an AI edge computing server according to claim 3, wherein: The resource assessment includes the following steps: Obtain in real time the current remaining available resources of the edge server with the highest priority mapped to the subtask, including computing resources, memory resources, and network bandwidth resources; Based on the second required resource amount, determining whether the current remaining available resource amount of the edge server with the highest priority meets the subtask execution requirement, and calculating the resource matching difference; If the resource matching difference exceeds the preset threshold, the edge server is marked as resource-scarce and the subtask migration in dynamic scheduling is triggered.

5. The scheduling optimization method for AI edge computing server according to claim 4, characterized in that: The migration operation of subtasks in the dynamic scheduling includes the following steps: According to the priority order set in the hierarchical mapping table, the current remaining available resources of other edge servers are evaluated in turn to determine whether the second required resource amount of the subtask is met; If there is at least one edge server whose resources can meet the computing requirements of the subtask, the subtask is assigned to the edge server for execution; If all the corresponding edge servers in the hierarchical mapping table cannot meet the resource requirements of the subtask, a backup edge server in the edge service cluster is started, and the subtask is assigned to the backup edge server for execution.

6. A scheduling optimization system for an AI edge computing server, configured to execute a scheduling optimization method for an AI edge computing server as described in any one of claims 1 to 5, comprising a pre-built edge service cluster, characterized in that: The system comprises: A splitting module is used to evenly split the resources of the task to be processed based on the first resource requirement of the task to be processed and the number of enabled edge servers in the edge server cluster, to obtain subtasks and their second resource requirements; a mapping module, configured to dynamically weight and establish a hierarchical mapping table of subtasks and edge servers based on selected matching indicators, wherein the matching indicators are hardware compatibility and data proximity between the subtasks and edge servers, each subtask in the hierarchical mapping table is mapped to at least two edge servers with a certain priority, and to guide the execution of the subtask on the mapped edge servers; A scheduling module is used to perform resource evaluation on the edge server corresponding to the subtask based on the hierarchical mapping table and the second required resource amount, and dynamically schedule the subtask according to the evaluation result, wherein the dynamic scheduling includes allocating the subtask to an edge server that meets the resource requirements for execution, or migrating the subtask when resources are insufficient.

Images