Resource migration method and device for AI task, equipment, medium and product

Abstract

The invention discloses an AI task resource migration method, device and equipment, a medium and a product, and the method comprises the steps: detecting the resource demand data and time delay requirements of an AI task sent by a terminal according to the pre-obtained resource state data of a base station node; and when different resource data in the resource demand data are distributed in different first base station nodes, or the terminal time delay between the base station node simultaneously having all resource data in the resource demand data and the terminal does not meet the time delay requirement, sending the resource demand data to the terminal. And judging to execute resource migration, determining a resource migration strategy according to the resource state data, the resource demand data and the time delay requirement, and performing resource migration control. According to the scheme, resource migration control between the base stations can be realized, and the resource utilization rate of the base stations is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a resource migration method and device for AI task, equipment, medium and product. BACKGROUND

[0002] At present, artificial intelligence has become the core driving force of the new round of industrial upgrading. The automation, digitization and intelligentization of the industry need ubiquitous intelligence. With the increasing scale of AI models in various fields, network side, user side and enterprise side have higher demands for efficient and fast AI services. It is difficult to support real-time and guaranteed AI services by simply increasing the computing power of user terminals and cloud AI services. Therefore, as a bridge between users and cloud AI services, the network not only needs to meet the needs of user communication, but also is expected to meet the needs of real-time AI services of users. By using the extensive and near-terminal deployment of network equipment, real-time AI service is realized.

[0003] At present, the prior art uses the spare computing power resources and storage resources of wireless network equipment (such as base stations) to provide real-time AI services. When the access base station resources of the user terminal are insufficient, the user AI task is unloaded and migrated to a base station with sufficient resources for service, and the user AI task is completed. However, the computing power of a single base station, the model and data set resources that can be stored are limited, and the resources of each base station cannot be effectively utilized, resulting in low resource utilization rate of the base station when executing AI tasks. SUMMARY

[0004] In order to solve the above problems, the present application provides a resource migration method, device, equipment, medium and product for AI task, which can realize resource migration control between base stations and improve the utilization rate of base station resources.

[0005] The present application provides a resource migration method for AI task, which specifically comprises:

[0006] According to the pre-acquired resource state data of the base station node, the resource requirement data and the time delay requirement of the AI task sent by the terminal are detected;

[0007] When different resource data in the resource requirement data are distributed in different first base station nodes, or the terminal time delay between the base station node that has all the resource data in the resource requirement data and the terminal does not meet the time delay requirement, it is determined to execute resource migration;

[0008] According to the resource state data, the resource requirement data and the time delay requirement, a resource migration strategy is determined, and resource migration control is performed.

[0009] Preferably, the different resource data in the resource requirement data include data set resources and model resources.

[0010] determining a resource migration strategy according to the resource state data, the resource requirement data and the latency requirement, and performing resource migration control, comprises:

[0011] triggering resource migration control when there is a terminal latency satisfying the latency requirement in the first base station node, and there is a base station node whose storage resource satisfies the storage requirement of the dataset resource and the model resource;

[0012] determining a resource migration strategy between different base station nodes according to the resource state data, the resource requirement data and the latency requirement when the AI task is executed;

[0013] controlling different base station nodes to complete resource migration according to the resource migration strategy.

[0014] Preferably, the determining a resource migration strategy according to the resource state data, the resource requirement data and the latency requirement, and performing resource migration control, comprises:

[0015] when there is no base station node whose terminal latency satisfies the latency requirement in the first base station node, or there is no base station node whose storage resource satisfies the storage requirement of the dataset resource and the model resource, determining whether all resource state data contained in all base station nodes contains dataset resource and model resource required by AI tasks to be executed in all base station nodes;

[0016] if not, outputting a feedback that the task cannot be completed;

[0017] if yes, triggering resource and task migration control;

[0018] determining a task migration strategy and a resource migration strategy of each AI task according to the resource state data, the resource requirement data of all AI tasks to be executed and the latency requirement;

[0019] migrating each AI task to a corresponding base station node according to the task migration strategy, and completing resource migration between different base station nodes for the AI task.

[0020] Preferably, the detecting the resource requirement data and the latency requirement of the received AI task comprises:

[0021] detecting a correspondence between dataset resource and model resource in the resource requirement data of the AI task and resource state data of each base station node;

[0022] when there is a base station node that owns the dataset resource, and there is a base station node that owns the model resource, and there is no base station node that simultaneously owns the dataset resource and the model resource, determining that different resource data in the resource requirement data are distributed in different base station nodes.

[0023] Preferably, the resource migration strategy between different base station nodes is determined according to the resource state data, the resource demand data and the latency requirement, comprising:

[0024] According to the resource state data, the resource demand data and the latency requirement, the source node of resource migration, the migration resource type, the resource migration mode and the target node of resource migration in the base station node are determined as the resource migration strategy based on the latency-minimizing greedy strategy.

[0025] Preferably, the resource migration strategy between different base station nodes is determined according to the resource state data, the resource demand data and the latency requirement, comprising:

[0026] sending an instruction containing the migration resource type, the resource migration mode and the target node of resource migration to the source node of migration resource, so as to make it migrate the resource data of the migration resource type to the target node through the resource migration mode;

[0027] sending feedback information containing the target node to the terminal, so as to make the terminal unload the AI task to the target node for running.

[0028] Preferably, the resource migration strategy between different base station nodes is determined according to the resource state data, the resource demand data and the latency requirement, comprising:

[0029] traversing the resource state data of all base station nodes, determining a first base station node set which stores resources satisfying the storage demand of the data set resource and the model resource, a second base station node set which owns the data set resource, and a third base station node set which owns the model resource;

[0030] determining the target node of resource migration from the first base station node set, and respectively determining the source node of corresponding migration resource type from the second base station node set and the third base station node set, to determine the resource migration strategy.

[0031] Preferably, the resource migration strategy between different base station nodes is determined according to the resource state data, the resource demand data and the latency requirement, comprising:

[0032] calculating the terminal latency of each base station node in the first base station node set, and determining the base station node with the shortest terminal latency as the target node of resource migration;

[0033] calculating the node latency between each base station node in the second base station node set and the target node, and determining the base station node with the shortest node latency as the source node of data set resource migration;

[0034] calculating a node delay between each base station node in the third set of base station nodes and the target node, and determining a base station node with the shortest node delay as a source node of model resource migration.

[0035] Preferably, the terminal delay is a sum of an air interface delay, a wired transmission delay and a calculation delay between the base station node and a user terminal sending the AI task.

[0036] The node delay is a ratio of a data amount size and a wired rate between two base station nodes.

[0037] The air interface delay is a ratio of the data amount size of the AI task and an air interface rate of the base station node, the wired transmission delay is a ratio of the data amount size of the AI task and a wired rate of the base station node, and the calculation delay is a ratio of the data amount size of the AI task and a calculation resource of the base station node.

[0038] Preferably, determining the resource migration strategy comprises:

[0039] When the node delays between the source nodes of different migration resource types and the target node, and the terminal delay of the target node all meet the delay requirement, outputting a preset resource migration mode, the target node, a migration resource type and a corresponding source node of the migration resource type as the resource migration strategy.

[0040] Preferably, according to the resource state data, resource requirement data of all to-be-executed AI tasks and the delay requirement, determining a task migration strategy and a resource migration strategy of each AI task comprises:

[0041] According to the delay requirement of all to-be-executed AI tasks, sorting the AI tasks from high to low according to the delay requirement, generating a task set, and initializing resource state data of all base station nodes.

[0042] According to the sorting of the task set, sequentially selecting an AI task therefrom as a migration AI task.

[0043] For each migration AI task, determining a resource migration strategy between different base station nodes according to the current resource state data of the base station nodes, resource requirement data of the migration AI task and the delay requirement, and updating resource state data of all base station nodes after resource migration of the task.

[0044] Outputting the target node of different migration AI tasks as a task migration strategy thereof, and correspondingly outputting a resource migration strategy of the migration AI task.

[0045] Preferably, for each migrated AI task, the resource migration strategy between different base station nodes is determined according to the current resource state data of the base station node, the resource requirement data of the migrated AI task and the delay requirement, and the resource state data of all base station nodes after the resource migration of the task is updated, including:

[0046] For each migrated AI task, the resource state data of all base station nodes is traversed to determine a first base station node set that stores data set resources and model resources that meet the storage requirements of the migrated AI task, a second base station node set that owns the data set resources of the migrated AI task, and a third base station node set that owns the model resources of the migrated AI task;

[0047] It is judged whether the first base station node set, or the second base station node set or the third base station node set is empty;

[0048] If yes, it is recorded that the migrated AI task cannot be completed;

[0049] If no, the target node of the resource migration of the migrated AI task is determined from the first base station node set, and the source node of the corresponding migration resource type is determined from the second base station node set and the third base station node set, respectively. It is judged whether the node delay between the source node and the target node of different migration resource types of the migrated AI task and the terminal delay of the target node meet the delay requirement of the migrated AI task, if not, it is recorded that the migrated AI task cannot be completed; if yes, the resource migration strategy of the migrated AI task is recorded, and the resource state data of all base station nodes after the resource migration of the task is updated.

[0050] Preferably, the resource state data acquisition process specifically includes:

[0051] Receiving the resource state data actively uploaded by each base station node after the change of the resource state; and / or,

[0052] Receiving the resource state data reported by other network function network elements;

[0053] The resource state data includes the communication resources, the computing resources, the model resources, the data set resources and the storage resources available to the base station node.

[0054] The present application provides an AI task resource migration device, which comprises:

[0055] An information acquisition module is configured to detect the resource requirement data and the delay requirement of the AI task sent by a terminal according to the resource state data of the base station node acquired in advance;

[0056] a determining module configured to determine to perform resource migration when different resource data in the resource requirement data are distributed in different first base station nodes, or a terminal delay between a base station node having all the resource data in the resource requirement data and the terminal does not satisfy the delay requirement;

[0057] a migration module configured to determine a resource migration strategy according to the resource state data, the resource requirement data and the delay requirement, and perform resource migration control.

[0058] Preferably, the migration module is specifically configured to:

[0059] trigger resource migration control when a terminal delay in the first base station node satisfies the delay requirement, and a base station node having storage resources satisfying the storage requirement of the dataset resource and the model resource exists;

[0060] determine a resource migration strategy between different base station nodes according to the resource state data, the resource requirement data and the delay requirement;

[0061] control different base station nodes to complete resource migration according to the resource migration strategy.

[0062] Preferably, the migration module is specifically configured to:

[0063] when there is no base station node in the first base station node satisfying the delay requirement, or there is no base station node having storage resources satisfying the storage requirement of the dataset resource and the model resource, determine whether all resource state data contained in all base station nodes contain dataset resources and model resources required by AI tasks to be executed in all base station nodes;

[0064] if not, output a feedback that the task cannot be completed;

[0065] if yes, trigger resource and task migration control;

[0066] determine a task migration strategy and a resource migration strategy of each AI task according to the resource state data, resource requirement data of all AI tasks to be executed and the delay requirement;

[0067] migrate each AI task to a corresponding base station node according to the task migration strategy, and complete resource migration between different base station nodes for the AI task.

[0068] Preferably, the information collection module is specifically configured to:

[0069] detect a corresponding relationship between dataset resources and model resources in the resource requirement data of the AI task and resource state data of each base station node;

[0070] when there is a base station node owning the dataset resource, and there is a base station node owning the model resource, and there is no base station node owning both the dataset resource and the model resource, determining that different resource data in the resource requirement data is distributed in different base station nodes.

[0071] Preferably, the migration module is specifically used to include:

[0072] According to the resource state data, the resource requirement data, and the latency requirement, determining, as the resource migration strategy, a source node of resource migration, a migration resource type, a resource migration mode, and a target node of resource migration in the base station nodes based on a latency-minimizing greedy strategy.

[0073] Preferably, the migration module is specifically used to include:

[0074] sending an instruction containing the migration resource type, the resource migration mode, and the target node of resource migration to the source node of the migration resource, so as to make the source node of the migration resource migrate resource data of the migration resource type to the target node through the resource migration mode;

[0075] sending feedback information containing the target node to the terminal, so as to make the terminal unload the AI task to the target node for running.

[0076] Preferably, the migration module is specifically used to include:

[0077] traversing resource state data of all base station nodes, determining a first base station node set storing resources satisfying storage requirements of the dataset resource and the model resource, a second base station node set owning the dataset resource, and a third base station node set owning the model resource;

[0078] determining a target node of resource migration from the first base station node set, and respectively determining a source node of a corresponding migration resource type from the second base station node set and the third base station node set, and determining the resource migration strategy.

[0079] Preferably, the migration module is specifically used to include:

[0080] calculating terminal latency of each base station node in the first base station node set, and determining a base station node with the shortest terminal latency as a target node of resource migration;

[0081] calculating node latency between each base station node in the second base station node set and the target node, and determining a base station node with the shortest node latency as a source node of dataset resource migration;

[0082] The node delay between each base station node in the third base station node set and the target node is calculated, and the base station node with the shortest node delay is determined as the source node of the model resource migration.

[0083] Preferably, the terminal delay is the sum of the air interface delay, the wired transmission delay and the calculation delay between the base station node and the user terminal sending the AI task.

[0084] The node delay is the ratio of the data amount to the wired rate between two base station nodes.

[0085] The air interface delay is the ratio of the data amount of the AI task to the air interface rate of the base station node, the wired transmission delay is the ratio of the data amount of the AI task to the wired rate of the base station node, and the calculation delay is the ratio of the data amount of the AI task to the calculation resource of the base station node.

[0086] Preferably, the migration module specifically includes:

[0087] When the node delay between the source node of different migration resource types and the target node, and the terminal delay of the target node all meet the delay requirement, output the preset resource migration mode, the target node, the migration resource type and the corresponding source node as the resource migration strategy.

[0088] Preferably, the migration module specifically includes:

[0089] According to the delay requirement of all AI tasks to be executed, the AI tasks are sorted from high to low according to the delay requirement, a task set is generated, and the resource state data of all base station nodes is initialized.

[0090] According to the sorting of the task set, AI tasks are selected from the task set in sequence as migration AI tasks.

[0091] For each migration AI task, the resource migration strategy between different base station nodes is determined according to the current resource state data of the base station node, the resource requirement data of the migration AI task and the delay requirement, and the resource state data of all base station nodes after the task resource migration is updated.

[0092] The target node of different migration AI tasks is output as the task migration strategy, and the resource migration strategy of the migration AI task is correspondingly output.

[0093] Preferably, the migration module specifically includes:

[0094] For each migration AI task, the resource state data of all base station nodes is traversed to determine a first base station node set satisfying the storage requirements of the dataset resources and the model resources of the migration AI task, a second base station node set owning the dataset resources of the migration AI task, and a third base station node set owning the model resources of the migration AI task;

[0095] It is determined whether the first base station node set, or the second base station node set or the third base station node set is empty;

[0096] If yes, it is recorded that the migration AI task cannot be completed;

[0097] If no, a target node of resource migration of the migration AI task is determined from the first base station node set, and a source node corresponding to the migration resource type is determined from the second base station node set and the third base station node set, respectively; it is determined whether the node delay between the source node and the target node of different migration resource types of the migration AI task and the terminal delay of the target node all satisfy the delay requirement of the migration AI task, if no, it is recorded that the migration AI task cannot be completed; if yes, a resource migration strategy of the migration AI task is recorded, and the resource state data of all base station nodes after the task resource migration is updated.

[0098] Preferably, the process of acquiring the resource state data by the information collection module specifically includes:

[0099] receiving the resource state data actively uploaded by each base station node after the resource state changes; and / or,

[0100] receiving the resource state data reported by other network function network elements;

[0101] The resource state data includes the communication resources, the computing resources, the model resources, the dataset resources and the storage resources available to the base station node.

[0102] The embodiment of the application further provides a terminal device, including a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, when the processor executes the computer program, the AI task resource migration method of any one of the above embodiments is realized.

[0103] The embodiment of the application further provides a computer readable storage medium, including a stored computer program, wherein when the computer program runs, the device where the computer readable storage medium is located executes the AI task resource migration method of any one of the above embodiments.

[0104] The embodiment of the present application also provides a computer program product comprising computer programs / instructions which, when executed by a processor, implement the steps of any of the above-described methods.

[0105] The present application provides a resource migration method and device for AI tasks, equipment, medium and product, resource state data of base station nodes is acquired in advance, resource requirement data and time delay requirement of AI tasks sent by a terminal are detected, when different resource data in the resource requirement data are distributed in different first base station nodes, or terminal time delay between a base station node with all resource data in the resource requirement data and the terminal does not meet the time delay requirement, it is determined to perform resource migration, resource migration strategy is determined according to the resource state data, the resource requirement data and the time delay requirement, and resource migration control is performed. The present application can realize resource migration control between base stations, and improve base station resource utilization. BRIEF DESCRIPTION OF DRAWINGS

[0106] Figure 1 is a process schematic diagram of AI task migration in the prior art;

[0107] Figure 2 is a process schematic diagram of the AI task resource migration method provided by the embodiment of the present application;

[0108] Figure 3 is another process schematic diagram of the AI task resource migration method provided by the embodiment of the present application;

[0109] Figure 4 is still another process schematic diagram of the AI task resource migration method provided by the embodiment of the present application

[0110] Figure 5 is a structural schematic diagram of the AI task resource migration device provided by the embodiment of the present application;

[0111] Figure 6 is a structural schematic diagram of a terminal device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0112] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0113] In the AI task execution, the existing scheme usually has task offloading and task migration. When the access base station resource of the user terminal is insufficient, the user AI task can be offloaded to the resource sufficient base station for service; or the low latency demand AI task in the access base station can be migrated to the remote base station for completion, so as to empty the computing resource to complete the user AI task.

[0114] Referring to Figure 1 is a flowchart of the AI task migration in the prior art. The user UE needs to perform AI service 2, which is migrated through the UPF to the second base station from the nearest first base station for execution. The second base station uses the model 2 therein to execute the AI service 2 based on the remaining 20% of the computing resource. The first base station determines the AI service 1 to be executed and the AI service 2 based on the AI service registration and discovery, wherein the AI service 1 is migrated by other terminals, and the first base station executes the AI service 2 based on the data 1 and the model 2 based on the remaining 80% of the computing resource.

[0115] The task migration scheme in the prior art cannot reasonably utilize the resources of each base station by only migrating the AI task. When the AI task has high real-time requirements and the required resources are in the remote base station node, the task offloading method can only offload the AI task to the remote base station node for service, which makes it difficult to meet the time delay. The task migration method can only migrate the AI task to the base station node that has the model and data resources required by the AI task, which easily leads to all AI tasks of the same type being completed in the same base station node, which is not conducive to efficient resource utilization. Moreover, when the model and data required by the AI task are distributed in different nodes, only task offloading and task migration cannot achieve the service provision of the AI task. When the AI task has high real-time requirements and the computing resources and storage resources of the near-end base station node are occupied by low-priority AI tasks and AI data sets and models, the high-priority task will be offloaded to the remote base station node for service. The existing task migration strategy only considers the migration of AI task data, and still cannot solve the problem of the occupation of storage resources by the existing AI data sets and models in the base station. Even if the computing resources of the near-end base station are emptied by AI task migration, only the AI tasks corresponding to the models and data resources owned by the base station can be completed, and the AI tasks that are not supported by the current resources of the base station cannot be completed.

[0116] Based on the above defects of the prior art, the AI task resource migration method provided in the present application is proposed, referring to Figure 2 is a flowchart of the AI task resource migration method provided in the embodiment of the present application. The method comprises the following steps:

[0117] In step S1, the resource requirement data and the time delay requirement of the AI task sent by the terminal are detected according to the pre-acquired resource state data of the base station node.

[0118] Step S2, when different resource data in the resource requirement data are distributed in different first base station nodes, or the terminal delay between the base station node simultaneously owning all resource data in the resource requirement data and the terminal does not meet the delay requirement, it is determined to perform resource migration;

[0119] Step S3, determining a resource migration strategy according to the resource state data, the resource requirement data and the delay requirement, and performing resource migration control.

[0120] In specific implementation, it is necessary to pre-acquire the resource state data of each base station node at present, wherein the resource state data includes state data and resource data related to AI task execution, which can represent the adaptation ability and computing ability of the base station node to different AI tasks.

[0121] When the user terminal sends an AI task, the task requirement corresponding to the AI task needs to be sent, the resource and end-to-end delay requirement information after the AI task is disassembled, including resource requirement data, i.e. the resource data needed for AI task execution, and the end-to-end delay requirement that needs to be met for AI task execution, which is taken as the delay requirement of the AI task.

[0122] When receiving the AI task requirement sent by the user terminal, when performing AI task execution allocation, the resource migration method for executing the above-mentioned AI task needs to be performed, which has the following steps:

[0123] The resource requirement data and delay requirement of the AI task are disassembled and detected, and the corresponding resource requirement data is searched in the resource state data of the base station node.

[0124] It is judged whether the resource data required by the AI task requested by the current user is distributed in different first base station nodes, which includes two layers of judgment, i.e. whether the resource state data of all base station nodes can meet the resource requirement of the AI task, and the purpose of this judgment is to judge whether the base station node can execute the AI task. Only when different resource data in the resource requirement data are in the resource state data of all base station nodes, all base station nodes can meet the execution requirement of the AI task. In addition, a single base station node cannot meet all the resource data required by the AI task, and this judgment excludes the case that a certain device node can execute the AI task alone.

[0125] And it is judged whether the terminal delay between the base station node simultaneously owning all resource data in the resource requirement data and the terminal meets the delay requirement; i.e. a certain device node meets all the resource conditions of the AI task, but does not meet the delay requirement, which excludes the case that a certain device node can execute the AI task alone.

[0126] Based on the above detection, the situation that a certain device node can independently perform an AI task is excluded, and therefore resource migration needs to be performed.

[0127] The resource migration strategy is determined according to the resource state data, the resource requirement data and the time delay requirement, and a corresponding optimization algorithm is called to output a corresponding resource migration strategy.

[0128] It should be noted that when the application outputs the resource migration strategy, different base station nodes corresponding to the resource data can be queried according to the resource requirement data of different AI tasks, the target base station node is determined, the missing resource data that the target base station node does not have in the resource requirement data is obtained, the base station node corresponding to the corresponding missing resource data is queried as the source base station node of migration, and the migration is completed through a preset migration mode.

[0129] According to the resource migration strategy, the resources on different base station nodes are migrated to realize the running processing of the received AI task.

[0130] When the application detects that the AI task needs to be migrated, the migration strategy is automatically generated, and the resource migration control is performed, so as to support the flexible migration of the AI task and the required resources in the wireless side network, and improve the number of AI tasks and the resource utilization rate that can be supported by the network.

[0131] In another embodiment provided by the application, the different resource data in the resource requirement data of the AI task includes dataset resources and model resources.

[0132] The resource and time delay requirement of the AI task includes the data amount size of completing the AI task, the required communication resource (air interface rate and wired rate), the calculation resource (the time delay required by the calculation unit data), the model resource (model type and data amount size), the data resource (data type and data amount size) and the storage resource, and the end-to-end time delay.

[0133] When the resource migration trigger is determined, the following determination is performed:

[0134] Step 1: Determine whether the specific dataset resource and model resource required by the AI task are distributed in different first base station nodes, if yes, trigger the resource migration control, otherwise go to step 2;

[0135] Step 2, determine whether the base station node capable of providing services for the AI task can meet the end-to-end time delay requirement, if yes, trigger the resource migration control, otherwise it is indicated that there is a single base station node that meets the resource data requirement and the time delay requirement, and the AI task does not need to be performed.

[0136] It should be noted that the base station node capable of providing services for the AI task is specifically a base station node simultaneously having all resource data in the resource requirement data.

[0137] After triggering the resource migration control, the following steps are specifically executed:

[0138] Step three, determining whether the terminal delay between a base station node and the terminal in the first base station node meets the AI task end-to-end delay requirement, if yes, entering step four, otherwise, it is indicated that the delay between all base station nodes and the terminal does not meet the delay requirement, and the existing resources cannot execute the AI task.

[0139] Step four, determining whether the storage resource of the base station node in the first base station node can meet the storage requirement of the data set resource and the model resource required by the AI task, if yes, it is indicated that the storage data of the current base station node can meet the execution requirement of the AI task, and only resource migration between different base station nodes is required.

[0140] In specific implementation, referring to Figure 3 is another flowchart of the AI task resource migration method provided by the embodiment of the application.

[0141] In the AI task resource migration method, the resource and task migration control module is executed, wherein the information acquisition module is used to acquire the AI task requirement of the user terminal and perform resource state perception on the base station a, the base station b and the base station C, acquire the resource state data, and different base stations also upload the data resource required by the AI task to make a decision on the AI task resource migration.

[0142] After the information acquisition module acquires the data, the data is sent to the decision module, through decision, the decision module determines whether to trigger the resource migration flow according to steps one to four, that is, whether to perform separate resource migration, and triggers the resource migration flow.

[0143] According to the resource migration strategy, different base station nodes complete resource migration.

[0144] After determining that resource migration is required, it is determined that the resource state data of the current base station node can meet the execution requirement of the AI task, the separate resource migration strategy can meet the execution requirement of the AI task, the optimal resource migration strategy output is realized, the resource utilization rate of the base station node is improved, and the interference on other AI task execution is reduced.

[0145] In another embodiment provided by the application, when the judgment of resource migration triggering is performed, the following judgment is performed:

[0146] Step one: judge whether the specific data set resources and model resources required by the AI task simultaneously are distributed in different first base station nodes, if yes, trigger resource migration control, otherwise enter step two;

[0147] Step two, judge whether the base station node capable of providing service for the AI task can meet the end-to-end delay requirement, if yes, trigger resource migration control, otherwise it is indicated that there is a separate base station node meeting the resource data requirement and the delay requirement, and the AI task does not need to be executed by resource migration.

[0148] It should be noted that the base station node capable of providing service for the AI task is specifically a base station node simultaneously having all resource data in the resource requirement data.

[0149] After triggering the resource migration control, the following steps are specifically executed:

[0150] Step three, judge whether the terminal delay between a base station node in the first base station node and the terminal meets the AI task end-to-end delay requirement, if not, enter step four, if yes, it is indicated that the delay between all first base station nodes and the terminal does not meet the delay requirement, and the existing resources cannot execute the current AI task, and step five is executed.

[0151] Step four, judge whether the storage resource of a base station node in the first base station node can meet the storage requirement of the data set resource and the model resource required by the AI task, if yes, it is indicated that the resource state data of the current base station node can meet the execution requirement of the AI task, and only resource migration between different base station nodes is needed. Otherwise, it is indicated that the storage resource of the existing base station node cannot support the resource migration of the resource required by the AI task, and the current AI task cannot be executed, and the AI task cannot be completed by simply resource migration, and step five is executed.

[0152] Participate Figure 4 It is another flowchart of the AI task resource migration method provided by the embodiment of the application.

[0153] The user terminal requests a high delay requirement task, the first base station can support the bottom oath requirement task and the model / resource data, and the second base station supports the model / data resource required by the high delay requirement task.

[0154] When the decision module judges that the resource migration process condition is not met, that is, there is no base station node in the first base station node whose terminal delay meets the delay requirement, or there is no base station node whose storage resource meets the storage requirement of the data set resource and the model resource, after the separate resource migration cannot complete the AI task, step five is executed, that is, the existing resources of the base station can meet the resources required by the AI task, if yes, the resource and task migration process is triggered.

[0155] Step five, judge whether all resource state data contained by all base station nodes contain data set resources and model resources required by AI tasks to be executed in all base station nodes; at this time, the resource state of the current base station node cannot complete the current AI task, therefore, it is necessary to judge whether all resources of all base station nodes can support data set resources and model resources required by all AI tasks (including user requests and AI tasks being executed) to be executed by the current base station node in combination with all resource state data contained by all base station nodes;

[0156] Judge whether the data set types and model types owned by all current base stations contain the data set types and model types required by all current AI tasks. That is, in the above process of traversing the current known base station nodes, when the following conditions are met, it is judged whether the data set types and model types owned by all current base stations contain the data set types and model types required by all current AI tasks, then the task and resource migration process is triggered, otherwise the judgment process of the task and resource migration process is ended, and the user is fed back that the task cannot be completed:

[0157] For any AI task (including user requests and AI tasks being executed), there is a base station node that owns the data set type required by the AI task;

[0158] For any AI task (including user requests and AI tasks being executed), there is a base station node that owns the model type required by the AI task;

[0159] According to the resource state data, the resource requirement data of all AI tasks to be executed and the time delay requirement, the resource migration strategy and the AI task migration strategy are determined, and the task migration strategy of each AI task and the resource migration strategy thereof are determined.

[0160] According to the task migration strategy, each AI task is migrated to the corresponding base station node, and resource migration between different base station nodes is completed for the AI task.

[0161] By detecting that all resources of all base station nodes can meet the resource requirements of all AI tasks, when the current remaining resources cannot support AI task execution, all resources are comprehensively considered, the number of supported AI tasks can be improved, and the resource utilization rate is improved.

[0162] In another embodiment provided by the application, it is judged whether the specific data set and model resources required by the AI task are distributed in different base station nodes. That is, the current known base station nodes are traversed, when the following three conditions are met, it is judged that the specific data set and model resources required by the AI task are distributed in different base station nodes, then it is judged that different resource data in the resource requirement data are distributed in different base station nodes, otherwise it is indicated that the resource requirement data are distributed in the same base station node, or there is demand resource that is not distributed in the base station node, and the task cannot be executed.

[0163] Any base station node does not own the specific data set and model resource of the AI task requirement at the same time;

[0164] There is a base station node that owns the specific data set resource of the AI task requirement;

[0165] There is a base station node that owns the model resource of the AI task requirement.

[0166] The different resource data in the resource requirement data is distributed in different base station nodes, which can determine whether resource migration needs to be performed, and whether the current resource data can complete the AI task.

[0167] In another embodiment provided by the application, when the resource migration strategy is executed alone, according to the resource state data, the resource requirement data and the time delay requirement, a resource migration target decision process is called based on the greedy strategy of minimum time delay to determine the node that needs to be migrated, the resource to be migrated, the resource migration mode and the target node of resource migration.

[0168] The target node, the source node and the resource type to be migrated by the source node are determined, the determined target node meets the time delay requirement, and the target node after migration has corresponding resource data. Based on this resource migration strategy, resource migration can be efficiently completed, and the AI task is executed in the target node.

[0169] In another embodiment provided by the application, when different base station nodes complete resource migration according to the resource migration strategy, referring to Figure 3 After determining the base station a as the target node of resource migration, the source node of resource migration base station b is informed, the target node of resource migration, the resource to be migrated and the resource migration mode are informed; the model resource migration source node c is informed of the resource migration target node a, the resource to be migrated and the resource migration mode.

[0170] That is, an instruction containing the migration resource type, the resource migration mode and the target node of resource migration is sent to the source node of the migrated resource, so that the source node of the migrated resource migrates the resource data of the migration resource type to the target node through the resource migration mode;

[0171] The feedback information containing the target node is sent to the terminal, so that the user terminal sends the AI task to the target base station node executing the AI task through the access base station, and the resource migration source node migrates the corresponding resource to the target node according to the information sent by the migration module, to complete the execution of the AI task.

[0172] In another embodiment provided by the application, the resource migration strategy between different base station nodes is determined according to the resource state data, the resource requirement data and the time delay requirement, and the following steps are specifically executed:

[0173] Traverse the base station nodes to find a base station node set A that stores resources and computing resources that meet the AI task demand, a base station node set B that can meet the AI task data set resource demand, and a base station node set C that can meet the AI task model resource demand. If set A is empty, feedback that the AI task cannot be completed.

[0174] When the remaining computing resources of the traversed base station node are greater than the required computing resources of the AI task, and the remaining storage resources are greater than the sum of the storage resources occupied by the data of the AI task, the required data set and the model, the base station node is included in set A.

[0175] When the data set type possessed by the traversed base station node contains the data set type required by the AI task, the base station node is included in set B.

[0176] When the model type possessed by the traversed base station node contains the model type required by the AI task, the base station node is included in set C.

[0177] From the first base station node set, determine the target node of resource migration according to a preset strategy, and respectively determine the source node of the corresponding migration resource type from the second base station node set and the third base station node set.

[0178] The target node is determined by searching for the minimum delay, which can reduce the end-to-end delay of the AI task.

[0179] In another embodiment provided by the application, when determining the target node of resource migration and the source node of different migration resource types, a minimum delay strategy is adopted, specifically:

[0180] Traverse the base station nodes in set A, calculate the end-to-end delay between the user and the node, and select the base station node with the minimum end-to-end delay as the target node of resource migration.

[0181] Traverse the base station nodes in set B, calculate the delay of transmitting the data set resource required by the AI task from the base station node to the target node, and select the base station node with the minimum delay as the source node of data set resource migration.

[0182] Traverse the base station nodes in set C, calculate the delay of transmitting the model resource required by the AI task from the base station node to the target node, and select the base station node with the minimum delay as the source node of model resource migration.

[0183] Determine the source node of resource migration that meets the resource demand and has the minimum delay with the target node, so as to ensure that the AI task has the minimum delay when executed.

[0184] In another embodiment provided by the application, the terminal delay is the sum of the air interface delay between the base station node and the user terminal sending the AI task, the wired transmission delay and the calculation delay.

[0185] The air interface delay between the base station node and the user is the ratio of the AI task data size to the air interface rate, the wired transmission delay between the base station node and the user is the ratio of the AI task data size to the wired rate, and the calculation delay between the base station node and the user is the ratio of the AI task data size to the calculation resource.

[0186] The node delay between the base station node and the base station node is the ratio of the transmission data size to the wired rate between the two base station nodes.

[0187] In another embodiment provided by the application, in the process of determining the resource migration strategy, the following steps are further performed:

[0188] It is judged whether the node delay between the source node of different migration resource types and the target node and the terminal delay of the target node meet the delay requirement.

[0189] That is, it is judged whether the maximum delay of the power-off delay between the user terminal and the target node, the data set transmission delay between the source node of the model resource migration and the target node, and the model transmission delay between the source node of the data set resource migration and the target node meets the delay requirement of the AI task.

[0190] If yes, the strategy is output, otherwise it is fed back that the AI task cannot be completed.

[0191] After judging the resource migration strategy, the delay requirement can be met, and the scheme after resource migration can meet the delay requirement of the AI task, avoiding meaningless resource migration due to delay.

[0192] In another embodiment provided by the application, when the AI task executed on the base station node needs to be migrated, and the resource migration of the AI task being executed is performed, according to the resource state data, the resource requirement data of all to-be-executed AI tasks and the delay requirement, the task migration strategy and the resource migration strategy of each AI task are determined, including:

[0193] An AI task set T is established by all AI tasks, and the tasks in the AI task set are sorted according to the end-to-end delay requirement of the AI tasks in the AI task set. The lower the end-to-end delay, the higher the priority of the AI task with higher delay requirement. The base station node set A is established by traversing the base station node, and the storage resource and the calculation resource of each base station are initialized.

[0194] According to the ordering of the task set, the AI tasks are sequentially selected as the migration AI tasks, for each migration AI task, the resource migration strategy between different base station nodes is determined according to the current resource state data of the base station node, the resource requirement data of the migration AI task and the time delay requirement, and the AI tasks with lower end-to-end time delay requirements are preferentially met according to the ordered AI task set T. It is clear that the AI task t is the task with the highest real-time requirement in the current AI task set T, and the base station node set A is the set of target base stations to be selected, Bi is the set of data set source base stations according to the data set resource type i required by the task t, and Ck is the set of data set source base stations according to the model resource type k required by the task t.

[0195] It should be noted that task migration is required for all tasks being executed, and since the resource migration is performed according to the order of received tasks during the execution of the previous task, the task resource migration efficiency may be poor, and when the AI task received subsequently cannot be executed, the AI task being executed is re-planned according to the priority of the time delay requirement, the resource migration strategy of each AI task is determined, and after the resource migration strategy of each task is determined, the resource state data of all base station nodes after the task resource migration is updated, and the resource migration strategy of the task with lower priority is determined based on the task.

[0196] The specific resource migration strategy determination scheme is the same as the single resource migration strategy determination scheme in the foregoing embodiment.

[0197] The target nodes of different migration AI tasks are output as the task migration strategies of the migration AI tasks, and the resource migration strategies of the migration AI tasks are correspondingly output.

[0198] When the resource data of the current base station node cannot meet the execution requirement of the current AI task, it may be because the task with low time delay requirement being executed occupies the low time delay base station node, so by re-migrating all tasks, the resource allocation rationality is ensured, the number of supported AI tasks is improved, and the resource utilization rate is improved.

[0199] In another embodiment of the present application, when each migration AI task is migrated, the resource state data of all base station nodes is traversed to determine a first base station node set that stores data set resources and model resources that meet the storage requirements of the migration AI task, a second base station node set that owns the data set resources of the migration AI task, and a third base station node set that owns the model resources of the migration AI task; the target node of the resource migration of the migration AI task is determined from the first base station node set, and the source node corresponding to the migration resource type is determined from the second base station node set and the third base station node set, respectively, and in specific implementation:

[0200] Traverse the base station node, find the base station node set A that the storage resource and the computing resource meet the AI task demand, the base station node set B that can meet the AI task data set resource demand, the base station node set C that can meet the AI task model resource demand.If set A is empty, then feedback the AI task cannot be completed.

[0201] When the remaining computing resource of the traversed base station node is greater than the required computing resource of the AI task, and the remaining storage resource is greater than the sum of the storage resource occupied by the AI task data, the required data set and the model, then the base station node is included in set A.

[0202] When the data set type possessed by the traversed base station node contains the required data set type of the AI task, then the base station node is included in set B.

[0203] When the model type possessed by the traversed base station node contains the required model type of the AI task, then the base station node is included in set C.

[0204] From the first base station node set, determine the target node of resource migration according to a preset strategy, and respectively determine the source node of the corresponding migration resource type from the second base station node set and the third base station node set,

[0205] Traverse the base station node, find the base station node set A that the storage resource and the computing resource meet the AI task demand, the base station node set B that can meet the AI task data set resource demand, the base station node set C that can meet the AI task model resource demand.If set A is empty, then feedback the AI task cannot be completed.

[0206] When the remaining computing resource of the traversed base station node is greater than the required computing resource of the AI task, and the remaining storage resource is greater than the sum of the storage resource occupied by the AI task data, the required data set and the model, then the base station node is included in set A.

[0207] When the data set type possessed by the traversed base station node contains the required data set type of the AI task, then the base station node is included in set B.

[0208] When the model type possessed by the traversed base station node contains the required model type of the AI task, then the base station node is included in set C.

[0209] Determine whether the first base station node set, or the second base station node set or the third base station node set is empty;

[0210] If there is an empty node set, it means that the task cannot be met.

[0211] Referring to Figure 4At this time, it is determined that a suitable node is not found, and feedback information indicates that the requested AI task cannot be met.

[0212] If there is no node set that is empty, a target node for resource migration is determined from the first base station node set according to a preset strategy, and source nodes corresponding to the migration resource types are determined from the second base station node set and the third base station node set, respectively.

[0213] The base station nodes in set A are traversed, the end-to-end delay between the user and the node is calculated, and the base station node with the smallest end-to-end delay is selected as the target node for resource migration.

[0214] The base station nodes in set B are traversed, the delay of the dataset resource required by the AI task from the base station node to the target node is calculated, and the base station node with the smallest delay is selected as the source node for dataset resource migration.

[0215] The base station nodes in set C are traversed, the delay of the model resource required by the AI task from the base station node to the target node is calculated, and the base station node with the smallest delay is selected as the source node for model resource migration.

[0216] After searching for the corresponding target node and source node, it is further determined whether the resource migration strategy of the AI task meets the delay requirement, whether the node delays between the source nodes and the target nodes of different migration resource types of the migration AI task meet the delay requirement of the migration AI task, and whether the terminal delay of the target node meets the delay requirement of the migration AI task.

[0217] If not, it is recorded that the migration AI task cannot be completed; if yes, the resource migration strategy of the migration AI task is recorded, and the resource state data of all base station nodes after the task resource migration is updated.

[0218] In the determination of the source node that meets the resource requirement and has the smallest delay with the target node, the smallest delay is ensured when the AI task is executed.

[0219] According to the task migration strategy, each AI task is migrated to the corresponding base station node, and resource migration between different base station nodes is completed for the AI task.

[0220] When performing task migration and resource migration, referring to Figure 4 When the first base station is the target node for executing the AI task with high delay requirement, the second base station is informed of the resource migration target node required by the AI task with high delay requirement, the resources that need to be migrated, and the resource migration method, i.e., the task originally executed in the second base station needs to be unloaded to the first base station.

[0221] The first base station is informed of the low-latency demand AI task and the resource migration target node, the resources that need to be migrated and the resource migration mode, that is, the low-latency task originally executed in the second base station needs to be offloaded to the second base station.

[0222] The target node executing the AI task is informed of the task migration strategy of migrating the low-latency demand AI task to the second base station and the task migration strategy of migrating the high-latency demand AI task to the first base station.

[0223] The first base station migrates the low-latency demand AI task and the model / data resources to the second base station, and the user terminal completes the high-latency demand AI task offloading and migrates the resources to the second base station to execute the low-latency demand AI task.

[0224] It should be noted that the present embodiment only illustrates the process of the present scheme embodiment with two AI tasks, and the working principle is similar when multiple AI tasks are executed, which is not described here.

[0225] The corresponding task is migrated to the target node, the user terminal sends the AI task to the target base station node executing the AI task through the access base station, and the resource migration corresponding to different AI tasks is completed, the task migration and resource migration are realized, and the resource utilization is improved.

[0226] In another embodiment provided by the present application, the resource state data acquisition process specifically includes:

[0227] Receiving resource state data actively uploaded by each base station node after the resource state changes; and / or,

[0228] Receiving resource state data reported by other network function network elements;

[0229] The information source can be divided into two ways. The first is a direct collection way, when the resource state of each base station node changes, the user terminal proposes an AI task demand, and the information is reported; the second is an indirect collection way, indirectly reporting information through other network functions.

[0230] Obtaining the resource and end-to-end latency demand information of the AI task after being decomposed from the network task arrangement management function.

[0231] The resource state data includes the communication resources, computing resources, model resources, dataset resources and storage resources available to the base station node.

[0232] The resource state data includes resource states of each base station node, resources of the AI task, and end-to-end latency requirements. The resource state includes communication resources (air interface rate and wired rate), computing resources (latency required for computing unit data), model resources (model type and data size), data resources (data type and data size), and storage resources available to the base station node, and also includes communication resources (air interface rate and wired rate), computing resources (latency required for computing unit data), model resources (model type and data size), data resources (data type and data size), and storage resources occupied by each AI task of the base station node.

[0233] The resources and latency requirements of the AI task include data size, required communication resources (air interface rate and wired rate), computing resources (latency required for computing unit data), model resources (model type and data size), data resources (data type and data size), and storage resources, and end-to-end latency.

[0234] Referring to Figure 5 is a structural schematic diagram of an AI task resource migration device provided by an embodiment of the present application. The device includes:

[0235] An information collection module is configured to detect resource requirement data and latency requirements of an AI task sent by a terminal according to pre-acquired resource state data of a base station node.

[0236] A decision module is configured to determine to perform resource migration when different resource data in the resource requirement data are distributed in different first base station nodes, or when terminal latency between a base station node that simultaneously has all resource data in the resource requirement data and the terminal does not meet the latency requirements.

[0237] A migration module is configured to determine a resource migration strategy according to the resource state data, the resource requirement data, and the latency requirements, and perform resource migration control.

[0238] The AI task resource migration device provided by the embodiment can perform all steps and functions of the AI task resource migration method provided by any of the above embodiments, and the specific functions of the device are not described here.

[0239] Referring to Figure 6 is a structural schematic diagram of a terminal device provided by an embodiment of the present application. The terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an AI task resource migration program. The processor implements the steps in each of the above AI task resource migration method embodiments when executing the computer program, such as Figure 1The steps S1-S3 are shown. Alternatively, the processor implements the functions of the modules in each of the above device embodiments when executing the computer program.

[0240] For example, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the resource migration device of the AI task. For example, the computer program can be divided into several modules, and the specific functions of each module have been described in detail in the AI task resource migration method provided in any of the above embodiments. Here, the specific functions of the device will not be repeated.

[0241] The terminal device can be a desktop computer, a notebook computer, a palm computer, a cloud server and the like. The terminal device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the schematic diagram is only an example of the terminal device, and does not limit the resource migration device of the AI task, and can include more or fewer components than the diagram, or combine certain components, or different components, for example, the terminal device can also include an input / output device, a network access device, a bus and the like.

[0242] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the resource migration device of the AI task, and connects each part of the resource migration device of the AI task through various interfaces and lines.

[0243] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the resource migration device of the AI task by running or executing the computer program and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required for a function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0244] The modules integrated in the resource migration device of the AI task can be stored in a computer-readable storage medium if they are realized in the form of software function units and sold or used as independent products. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be realized. The computer program includes computer program code, which can be in the form of source code, object code, an executable file, or some intermediate form, etc. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0245] The embodiment of the present application also provides a computer program product, which includes computer programs / instructions, and the computer programs / instructions realize the steps of the method when executed by a processor.

[0246] The computer program product provided by the embodiment can execute all steps and functions of the AI task resource migration method provided by any of the above-mentioned embodiments, and the specific functions of the product are not described here.

[0247] It should be noted that, for those skilled in the art, several improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements are also considered to be within the scope of protection of the present application.

Claims

1. A resource migration method of an AI task, characterized in that, The method comprises: According to the pre-acquired resource state data of the base station node, the resource requirement data and the time delay requirement of the AI task sent by the terminal are detected; When the different resource data in the resource requirement data are distributed in different first base station nodes, or the terminal time delay between the base station node that simultaneously possesses all the resource data in the resource requirement data and the terminal does not satisfy the time delay requirement, it is determined that resource migration is performed; According to the resource state data, the resource requirement data and the time delay requirement, the resource migration strategy is determined, and resource migration control is performed.

2. The AI task resource migration method of claim 1, wherein, The different resource data in the resource requirement data include dataset resources and model resources; The resource migration control according to the resource state data, the resource requirement data and the time delay requirement comprises: When there is a base station node in the first base station node whose terminal time delay satisfies the time delay requirement and whose storage resource satisfies the storage requirement of the dataset resources and the model resources, resource migration control is triggered; According to the resource state data, the resource requirement data and the time delay requirement, the resource migration strategy between different base station nodes during the execution of the AI task is determined; According to the resource migration strategy, the resource migration of different base station nodes is controlled. 3.The AI task resource migration method of claim 1, wherein, The resource migration control according to the resource state data, the resource requirement data and the time delay requirement comprises: When there is no base station node in the first base station node whose terminal time delay satisfies the time delay requirement, or there is no base station node whose storage resource satisfies the storage requirement of the dataset resources and the model resources in the resource requirement data, it is determined whether all the resource state data contained in all the base station nodes contain the dataset resources and the model resources required by the AI tasks to be executed in all the base station nodes; If not, a feedback that the task cannot be completed is outputted; If yes, resource and task migration control is triggered; According to the resource state data, the resource requirement data of all the AI tasks to be executed and the time delay requirement, the task migration strategy of each AI task and its resource migration strategy are determined; According to the task migration strategy, each AI task is migrated to the corresponding base station node, and the resource migration between different base station nodes for the AI task is completed. 4.The AI task resource migration method of claim 1, wherein, The detection of the resource requirement data and the time delay requirement of the received AI task comprises: The correspondence between the dataset resources and the model resources in the resource requirement data of the AI task and the resource state data of each base station node is detected; When there is a base station node that possesses the dataset resources, and there is a base station node that possesses the model resources, and there is no base station node that simultaneously possesses the dataset resources and the model resources, it is determined that the different resource data in the resource requirement data are distributed in different base station nodes. 5.The AI task resource migration method of claim 2, wherein, The determination of the resource migration strategy between different base station nodes according to the resource state data, the resource requirement data and the time delay requirement comprises: According to the resource state data, the resource demand data, and the latency requirement, a source node, a resource migration type, a resource migration mode, and a target node of resource migration in a base station node are determined based on a latency-minimizing greedy strategy as the resource migration strategy. 6.The AI task resource migration method of claim 2, wherein, The resource migration strategy is used to control different base station nodes to complete resource migration, including: sending an instruction containing the resource migration type, the resource migration mode, and the target node of resource migration to the source node of the migration resource, so that the source node migrates resource data of the migration resource type to the target node through the resource migration mode; sending feedback information containing the target node to the terminal, so that the terminal offloads the AI task to the target node for running. 7.The AI task resource migration method of claim 2, wherein, According to the resource state data, the resource demand data, and the latency requirement, a resource migration strategy between different base station nodes is determined, including: traversing the resource state data of all base station nodes to determine a first base station node set that stores resources satisfying the storage demand of the data set resource and the model resource, a second base station node set that owns the data set resource, and a third base station node set that owns the model resource; determining a target node of resource migration from the first base station node set, and determining a source node of a corresponding migration resource type from the second base station node set and the third base station node set, respectively, to determine the resource migration strategy. 8.The AI task resource migration method of claim 7, wherein, The determination of the target node of resource migration from the first base station node set, and the determination of the source node of the corresponding migration resource type from the second base station node set and the third base station node set, respectively, includes: calculating the terminal latency of each base station node in the first base station node set to determine a base station node with the shortest terminal latency as the target node of resource migration; calculating the node latency between each base station node in the second base station node set and the target node to determine a base station node with the shortest node latency as the source node of the data set resource migration; calculating the node latency between each base station node in the third base station node set and the target node to determine a base station node with the shortest node latency as the source node of the model resource migration.

9. The AI task resource migration method of claim 8, wherein, The terminal latency is the sum of the air interface latency, the wired transmission latency, and the calculation latency between the base station node and the user terminal sending the AI task; The node latency is the ratio of the transmission data size to the wired rate between two base station nodes; The air interface latency is the ratio of the data size of the AI task to the air interface rate of the base station node, the wired transmission latency is the ratio of the data size of the AI task to the wired rate of the base station node, and the calculation latency is the ratio of the data size of the AI task to the calculation resource of the base station node. 10.The AI task resource migration method of claim 7, wherein, The determination of the resource migration strategy includes: when the node latency between the source node of different migration resource types and the target node, and the terminal latency of the target node all satisfy the latency requirement, outputting a preset resource migration mode, the target node, a migration resource type, and a corresponding source node as the resource migration strategy. 11.The AI task resource migration method of claim 3, wherein, According to the resource state data, resource requirement data of all AI tasks to be executed, and the delay requirement, a task migration strategy and a resource migration strategy of each AI task are determined, including: According to the delay requirement of all AI tasks to be executed, the AI tasks are sorted from high to low according to the delay requirement, a task set is generated, and the resource state data of all base station nodes is initialized; According to the sorting of the task set, an AI task is selected as a migration AI task from the task set in sequence; For each migration AI task, a resource migration strategy between different base station nodes is determined according to the current resource state data of the base station node, the resource requirement data of the migration AI task, and the delay requirement, and the resource state data of all base station nodes after the resource migration of the task is updated; The target node of different migration AI tasks is output as the task migration strategy, and the resource migration strategy of the migration AI task is correspondingly output.

12. The AI task resource migration method of claim 11, wherein, For each migration AI task, a resource migration strategy between different base station nodes is determined according to the current resource state data of the base station node, the resource requirement data of the migration AI task, and the delay requirement, and the resource state data of all base station nodes after the resource migration of the task is updated, including: For each migration AI task, the resource state data of all base station nodes is traversed to determine a first base station node set that stores data set resources and model resources that meet the storage requirements of the migration AI task, a second base station node set that owns the data set resources of the migration AI task, and a third base station node set that owns the model resources of the migration AI task; Determine whether the first base station node set, or the second base station node set, or the third base station node set is empty; If yes, record that the migration AI task cannot be completed; If not, determine the target node of the resource migration of the migration AI task from the first base station node set, and determine the source node of the corresponding migration resource type from the second base station node set and the third base station node set respectively; judge whether the node delay between the source node and the target node of different migration resource types of the migration AI task and the terminal delay of the target node meet the delay requirement of the migration AI task, if not, record that the migration AI task cannot be completed; if yes, record the resource migration strategy of the migration AI task, and update the resource state data of all base station nodes after the resource migration of the task.

13. The AI task resource migration method of claim 1, wherein, The resource state data acquisition process specifically includes: Receiving resource state data actively uploaded by each base station node after the resource state changes; and / or, Receiving resource state data reported by other network function network elements; The resource state data includes available communication resources, computing resources, model resources, data set resources, and storage resources of the base station node.

14. An AI task resource migration apparatus, characterized by comprising: The device includes: An information collection module configured to detect resource requirement data and delay requirements of AI tasks sent by a terminal according to pre-acquired resource state data of base station nodes; a determining module configured to determine to perform resource migration when different resource data in the resource requirement data is distributed in different first base station nodes, or a terminal delay between a base station node having all resource data in the resource requirement data and the terminal does not satisfy the delay requirement; a migrating module configured to determine a resource migration strategy according to the resource state data, the resource requirement data and the delay requirement, and perform resource migration control.

15. A terminal device, comprising: A computer readable storage medium includes a computer program stored therein, wherein the computer program, when executed, controls a device in which the computer readable storage medium is located to perform the resource migration method for an AI task according to any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The computer program / instruction, when executed by a processor, implements the steps of the method according to any one of claims 1 to 13.

17. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instruction, when executed by a processor, implements the steps of the method according to any one of claims 1 to 13.

Images