Computing power management method for distributed AI training tasks in mobile network

Abstract

The invention discloses a computing power management method for distributed AI training tasks in a mobile network, and relates to the technical field of mobile communication and artificial intelligence. Comprising the following steps: constructing a communication-sensing-computing integrated mobile private network, and when a mobile terminal accesses the communication-sensing-computing integrated mobile private network, establishing a session for an AI training task of the mobile terminal through a core network, an AI training task is split into distributed cooperative training subtasks for calculation according to computing power resources of a mobile terminal, a base station and a cloud node through a computing network integration service arrangement network element, when the mobile terminal triggers switching, a network switching signaling is identified through a core network, and computing power service switching of the corresponding AI training task is executed according to the network switching signaling.

Description

Technical Field

[0001] The present invention discloses a computing power management method for distributed AI training tasks in a mobile network, and relates to the technical fields of mobile communications and artificial intelligence. Background Art

[0002] With the rapid development of mobile communication technology and the widespread application of AI technology, the application of distributed AI training in mobile networks is increasing. However, when the mobile terminal switches in the mobile network, such as Xn switching or N2 switching, due to the dynamic changes in computing nodes and network quality, the continuity of distributed AI training tasks is often not effectively guaranteed, which easily leads to problems such as training interruption, repeated data transmission and waste of computing resources. Summary of the invention

[0003] In view of the problems of the prior art, the present invention provides a computing power management method for distributed AI training tasks in a mobile network. By enhancing the core network to build an integrated mobile private network with synergy and computing, computing power resources are reallocated and the progress of new and old computing power nodes are synchronized during network switching, ensuring the continuity of distributed AI training tasks during the switching process.

[0004] The specific scheme proposed by the present invention is:

[0005] The present invention also provides a computing power management method for distributed AI training tasks in a mobile network, comprising:

[0006] Build a mobile private network with integrated sensing and computing: Add computing-network fusion perception network elements, computing-network fusion control and scheduling network elements, and computing-network fusion service orchestration network elements on the core network side. Use the computing-network fusion perception network elements to perceive the status of network and computing resources, use the computing-network fusion control and scheduling network elements to control the scheduling strategy of network and computing resources, and use the computing-network fusion service orchestration network elements to orchestrate distributed AI training tasks.

[0007] When a mobile terminal accesses the integrated mobile private network, a session is established for the AI ​​training task of the mobile terminal through the core network. The network element of the computing network fusion service orchestration splits the AI ​​training task into distributed collaborative training subtasks for calculation according to the computing resources of the mobile terminal, base station and cloud node.

[0008] When the mobile terminal triggers switching, the core network identifies the network switching signaling and executes the computing power service switching of the corresponding AI training task according to the network switching signaling.

[0009] Furthermore, the computing power service switching of the computing power management method for distributed AI training tasks in a mobile network includes:

[0010] The core network identifies the network switching signaling Path Switch Request, and executes the XN switching of the computing power service of the mobile terminal according to the network switching signaling, causing the change of the computing power node.

[0011] The core network identifies the network switching signaling Handover Required, and executes N2 switching of the computing power service of the mobile terminal according to the network switching signaling, causing the computing power node to be changed.

[0012] Furthermore, the computing power management method for a distributed AI training task in a mobile network performs XN switching, causing a change in computing power nodes, including:

[0013] When the core network identifies the network switching signaling Path Switch Request, it switches the user SUPI, target base station Target-RAN-ID and switching type to Xn through the computing-network convergence service orchestration network element. The distributed AI training task identifier TASKID is found according to the user SUPI through the computing-network convergence service orchestration network element, and the computing power node and training strategy of the AI ​​training task are updated. When the Xn switch starts, it stops updating parameters with other computing power nodes, notifies the target base station computing power node to reserve computing power resources and preheat the model, and synchronizes the old computing power node parameters to the new computing power node through the Xn interface between the source base station and the target base station. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and performs collaborative training synchronization with the cloud node.

[0014] Furthermore, the computing power management method for distributed AI training tasks in a mobile network performs N2 switching, causing a change in computing power nodes, including:

[0015] When the core network recognizes the network switching signaling HandOver Requried, it switches the user SUPI, the target base station Target-RAN-ID and the switching type to N2 through the computing-network convergence service orchestration network element. The distributed AI training task identifier TASKID is found according to the user SUPI through the computing-network convergence service orchestration network element, and the computing power node and training strategy of the AI ​​training task are updated. In the N2 switching preparation phase, the target base station computing power node is notified to reserve computing power resources and preheat the model. When switching, it stops updating parameters with other computing power nodes, and synchronizes the old computing power node parameters to the new computing power node through the transmission tunnel between the source base station, the core network UPF and the target base station. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and performs collaborative training synchronization with the cloud node.

[0016] The present invention also provides a computing power management device for distributed AI training tasks in a mobile network, including a mobile private network management module and a computing power management module.

[0017] The mobile private network management module builds a mobile private network with integrated sensing and computing: on the core network side, computing network fusion perception network elements, computing network fusion control and scheduling network elements, and computing network fusion service orchestration network elements are added. The computing network fusion perception network elements are used to perceive the status of network and computing resources, the computing network fusion control and scheduling network elements are used to control the scheduling strategy of network and computing resources, and the computing network fusion service orchestration network elements are used to orchestrate distributed AI training tasks.

[0018] When a mobile terminal accesses the integrated mobile private network, the core network establishes a session for the AI ​​training task of the mobile terminal through the computing power management module. The computing power management module splits the AI ​​training task into distributed collaborative training subtasks for calculation based on the computing power resources of the mobile terminal, base station and cloud node through the computing network fusion service orchestration network element.

[0019] When the mobile terminal triggers switching, the computing power management module identifies the network switching signaling through the core network and executes the computing power service switching of the corresponding AI training task according to the network switching signaling.

[0020] Furthermore, the computing power management module of the computing power management device for distributed AI training tasks in a mobile network performs computing power service switching, including:

[0021] The core network identifies the network switching signaling Path Switch Request, and executes the XN switching of the computing power service of the mobile terminal according to the network switching signaling, causing the change of the computing power node.

[0022] The core network identifies the network switching signaling Handover Required, and executes N2 switching of the computing power service of the mobile terminal according to the network switching signaling, causing the computing power node to be changed.

[0023] Furthermore, the computing power management module of the computing power management device for distributed AI training tasks in a mobile network performs XN switching, causing a change in computing power nodes, including:

[0024] When the core network identifies the network switching signaling Path Switch Request, it switches the user SUPI, the target base station Target-RAN-ID, and the switching type to Xn through the computing-network convergence service orchestration network element. The computing-network convergence service orchestration network element finds the distributed AI training task identifier TASKID according to the user SUPI, and updates the computing node and training strategy of the AI ​​training task. When the Xn switch starts, it stops updating parameters with other computing nodes, notifies the target base station computing node to reserve computing resources and preheat the model, and synchronizes the old computing node parameters to the new computing node through the Xn interface between the source base station and the target base station. After the switch, the new computing node resumes computing based on the existing training experience of the old computing node and performs collaborative training synchronization with the cloud node.

[0025] Furthermore, the computing power management module of the computing power management device for distributed AI training tasks in a mobile network performs N2 switching, causing a change in computing power nodes, including:

[0026] When the core network recognizes the network switching signaling HandOver Requried, it switches the user SUPI, the target base station Target-RAN-ID and the switching type to N2 through the computing-network convergence service orchestration network element. The computing-network convergence service orchestration network element finds the distributed AI training task identifier TASKID according to the user SUPI, and updates the computing power node and training strategy of the AI ​​training task. In the N2 switching preparation phase, the target base station computing power node is notified to reserve computing power resources and preheat the model. When switching, it stops updating parameters with other computing power nodes, and synchronizes the old computing power node parameters to the new computing power node through the transmission tunnel between the source base station, the core network UPF and the target base station. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and performs collaborative training synchronization with the cloud node.

[0027] The benefits of the present invention are:

[0028] The present invention solves the problem of interruption of distributed AI training services due to terminal mobility in mobile communication networks, realizes dynamic switching of computing tasks, ensures the business continuity of AI training tasks, reduces repeated data transmission and calculation, and not only saves computing resources but also shortens training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a schematic diagram before the network architecture is switched.

[0030] Figure 2 This is a schematic diagram after the network architecture is switched.

[0031] Figure 3 This is a schematic diagram of the N2 switching process.

[0032] Figure 4 This is a schematic diagram of the XN switching process. DETAILED DESCRIPTION

[0033] The integration of synaesthesia and computing involves 6G networks, which deeply integrate the communication function of mobile networks and AI computing functions, providing available computing resources for mobile terminals anytime and anywhere.

[0034] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0035] The present invention is particularly suitable for scenarios where, when a mobile terminal triggers switching, computing nodes and network quality change, resulting in the inability to ensure the continuity of distributed AI training tasks.

[0036] Example 1

[0037] The present invention also provides a computing power management method for distributed AI training tasks in a mobile network, comprising:

[0038] Build an integrated mobile private network with communication, sensing and computing: Add computing-network fusion perception network elements, computing-network fusion control and scheduling network elements, and computing-network fusion service orchestration network elements on the core network side. The computing-network fusion perception network elements are used to perceive the status of network and computing resources. The computing-network fusion control and scheduling network elements are used to control the scheduling strategies of network and computing resources. The computing-network fusion service orchestration network elements are used to orchestrate distributed AI training tasks. These network elements interact with other programs through API interfaces based on HTTP2 service messages.

[0039] When a mobile terminal accesses the integrated mobile private network, a session is established for the AI ​​training task of the mobile terminal through the core network. The network element of the computing network fusion service orchestration splits the AI ​​training task into distributed collaborative training subtasks for calculation according to the computing resources of the mobile terminal, base station and cloud node.

[0040] When the mobile terminal triggers switching, the core network identifies the network switching signaling and executes the computing power service switching of the corresponding AI training task according to the network switching signaling.

[0041] The computing power service switching may include:

[0042] The core network identifies the network switching signaling Path Switch Request, and executes the XN switching of the computing power service of the mobile terminal according to the network switching signaling, causing the change of the computing power node.

[0043] The core network identifies the network switching signaling Handover Required, and executes N2 switching of the computing power service of the mobile terminal according to the network switching signaling, causing the computing power node to be changed.

[0044] Specifically, performing N2 switching causes a change in computing power nodes, which may specifically include:

[0045] The core network detects that the terminal is about to undergo N2 handover by identifying the network handover signaling Handover Required, and extracts information such as user SUPI and target base station Target-RAN-ID from the signaling.

[0046] The core network notifies the computing network convergence service orchestration network element that the mobile terminal is ready to switch, and switches the user SUPI, target base station Target-RAN-ID and switching type to N2 through the computing network convergence service orchestration network element.

[0047] The computing-network fusion service orchestration network element finds the computing task ID (TASKID) of the current terminal through the user SUPI, and requests a new resource scheduling strategy from the computing-network fusion control and scheduling network element.

[0048] The computing-network fusion service orchestration network element performs service orchestration according to the new resource scheduling strategy, notifies the computing power node on the target base station to start reserving the dispatched computing power indicators and pre-loads the operating environment, and then quickly tests the training model.

[0049] The core network receives the Handover Request Ack signaling, confirming that the mobile terminal can initiate N2 switching. After returning the Handover Command signaling, it notifies the edge cloud computing node to no longer update the gradient parameters and transmit training data to the source base station computing node. At the same time, it notifies the source base station computing node to stop training, and records the number of training rounds, gradients, weights, biases and other parameters, and packages the untrained data.

[0050] The source base station and the target base station synchronize parameters and untrained data through the fronthaul tunnel established with the core network UPF.

[0051] The core network receives the HandOver Notify signaling from the target base station, senses the handover completion and informs the computing network fusion service orchestration network element, which notifies the edge cloud computing power node to continue updating the gradient and transmitting training data to the target base station, and at the same time, notifies the source base station computing power node to clear the old task information. After the handover, the new computing power node resumes computing based on the training experience of the old computing power node and performs collaborative training synchronization with the cloud computing power node.

[0052] Specifically, performing XN switching causes changes in computing power nodes, including:

[0053] The core network detects that the mobile terminal is undergoing Xn switching by identifying the network switching signaling Path Switch Request, and extracts information such as SUPI and Target-RAN-ID from the signaling.

[0054] The core network notifies the computing network convergence service orchestration network element that the mobile terminal is switching, and switches the user SUPI, target base station Target-RAN-ID, and switching type to XN through the computing network convergence service orchestration network element.

[0055] The computing-network fusion service orchestration network element finds the computing task ID (TASKID) of the current terminal through the user SUPI, and requests a new resource scheduling strategy from the computing-network fusion control and scheduling network element.

[0056] The computing-network fusion service orchestration network element performs service orchestration according to the new resource scheduling strategy, immediately notifies the edge cloud computing power node to stop updating the gradient parameters and transmitting training data to the source base station computing power node, and notifies the computing power node on the target base station to start reserving the distributed computing power indicators and pre-load the operating environment, as well as quickly test-run the training model. At the same time, it also notifies the source base station computing power node to stop training, and records the number of training rounds, gradients, weights, biases and other parameters and packages the untrained data.

[0057] The source base station and the target base station directly synchronize parameters and untrained data through the Xn interface between the base stations.

[0058] After the core network completes the downlink tunnel update with the target base station, it sends a Path SwitchRequest Ack signaling to the target base station, indicating that the downlink link is connected. Then, the core network notifies the computing network convergence service orchestration network element that the switch is complete. The computing network convergence service orchestration network element instructs the edge cloud computing power node to start updating the gradient and transmitting data to the target base station, and also instructs the source base station computing power node to clear the old task information. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and synchronizes with the cloud computing power node for collaborative training.

[0059] Example 2

[0060] The present invention also provides a computing power management device for distributed AI training tasks in a mobile network, including a mobile private network management module and a computing power management module.

[0061] The mobile private network management module builds a mobile private network with integrated sensing and computing: on the core network side, computing network fusion perception network elements, computing network fusion control and scheduling network elements, and computing network fusion service orchestration network elements are added. The computing network fusion perception network elements are used to perceive the status of network and computing resources, the computing network fusion control and scheduling network elements are used to control the scheduling strategy of network and computing resources, and the computing network fusion service orchestration network elements are used to orchestrate distributed AI training tasks.

[0062] When a mobile terminal accesses the integrated mobile private network, the core network establishes a session for the AI ​​training task of the mobile terminal through the computing power management module. The computing power management module splits the AI ​​training task into distributed collaborative training subtasks for calculation based on the computing power resources of the mobile terminal, base station and cloud node through the computing network fusion service orchestration network element.

[0063] When the mobile terminal triggers switching, the computing power management module identifies the network switching signaling through the core network and executes the computing power service switching of the corresponding AI training task according to the network switching signaling.

[0064] As the information interaction and execution process between the modules in the above-mentioned device are based on the same concept as the embodiment of the method of the present invention, the specific contents can be found in the description of the embodiment of the method of the present invention and will not be repeated here.

[0065] Similarly, the device of the present invention solves the problem of interruption of distributed AI training services due to terminal mobility in mobile communication networks, realizes dynamic switching of computing tasks, ensures the business continuity of AI training tasks, reduces repeated data transmission and calculation, and not only saves computing resources but also shortens training time.

[0066] It should be noted that not all steps and modules in the above-mentioned processes and device structures are necessary, and some steps or modules can be ignored according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or some components in multiple independent devices may be implemented together.

[0067] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A computing power management method for distributed AI training tasks in a mobile network, characterized by include: Build a mobile private network with integrated sensing and computing: Add computing-network fusion perception network elements, computing-network fusion control and scheduling network elements, and computing-network fusion service orchestration network elements on the core network side. Use the computing-network fusion perception network elements to perceive the status of network and computing resources, use the computing-network fusion control and scheduling network elements to control the scheduling strategy of network and computing resources, and use the computing-network fusion service orchestration network elements to orchestrate distributed AI training tasks. When a mobile terminal accesses the integrated mobile private network, a session is established for the AI ​​training task of the mobile terminal through the core network. The network element of the computing network fusion service orchestration splits the AI ​​training task into distributed collaborative training subtasks for calculation according to the computing resources of the mobile terminal, base station and cloud node. When the mobile terminal triggers switching, the core network identifies the network switching signaling and executes the computing power service switching of the corresponding AI training task according to the network switching signaling.

2. According to claim 1, a computing power management method for distributed AI training tasks in a mobile network is characterized by: The computing power service switching includes: The core network identifies the network switching signaling Path Switch Request, and executes the XN switching of the computing power service of the mobile terminal according to the network switching signaling, causing the change of the computing power node. The core network identifies the network switching signaling Handover Required, and executes N2 switching of the computing power service of the mobile terminal according to the network switching signaling, causing the computing power node to be changed.

3. According to claim 2, a computing power management method for distributed AI training tasks in a mobile network is characterized by: Performing XN switching will cause changes in computing power nodes, including: When the core network identifies the network switching signaling Path Switch Request, it switches the user SUPI, target base station Target-RAN-ID and switching type to Xn through the computing-network convergence service orchestration network element. The distributed AI training task identifier TASKID is found according to the user SUPI through the computing-network convergence service orchestration network element, and the computing power node and training strategy of the AI ​​training task are updated. When the Xn switch starts, it stops updating parameters with other computing power nodes, notifies the target base station computing power node to reserve computing power resources and preheat the model, and synchronizes the old computing power node parameters to the new computing power node through the Xn interface between the source base station and the target base station. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and performs collaborative training synchronization with the cloud node.

4. According to claim 2, a computing power management method for distributed AI training tasks in a mobile network is characterized by: Switching to N2 causes changes in computing nodes, including: When the core network recognizes the network switching signaling HandOver Requried, it switches the user SUPI, the target base station Target-RAN-ID and the switching type to N2 through the computing-network convergence service orchestration network element. The distributed AI training task identifier TASKID is found according to the user SUPI through the computing-network convergence service orchestration network element, and the computing power node and training strategy of the AI ​​training task are updated. In the N2 switching preparation phase, the target base station computing power node is notified to reserve computing power resources and preheat the model. When switching, it stops updating parameters with other computing power nodes, and synchronizes the old computing power node parameters to the new computing power node through the transmission tunnel between the source base station, the core network UPF and the target base station. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and performs collaborative training synchronization with the cloud node.

5. A computing power management device for distributed AI training tasks in a mobile network, characterized by Including mobile private network management module and computing power management module, The mobile private network management module builds a mobile private network with integrated sensing and computing: on the core network side, computing network fusion perception network elements, computing network fusion control and scheduling network elements, and computing network fusion service orchestration network elements are added. The computing network fusion perception network elements are used to perceive the status of network and computing resources, the computing network fusion control and scheduling network elements are used to control the scheduling strategy of network and computing resources, and the computing network fusion service orchestration network elements are used to orchestrate distributed AI training tasks. When a mobile terminal accesses the integrated mobile private network, the core network establishes a session for the AI ​​training task of the mobile terminal through the computing power management module. The computing power management module splits the AI ​​training task into distributed collaborative training subtasks for calculation based on the computing power resources of the mobile terminal, base station and cloud node through the computing network fusion service orchestration network element. When the mobile terminal triggers switching, the computing power management module identifies the network switching signaling through the core network and executes the computing power service switching of the corresponding AI training task according to the network switching signaling.

6. The computing power management device for distributed AI training tasks in a mobile network according to claim 5, characterized in that The computing power management module performs computing power service switching, including: The core network identifies the network switching signaling Path Switch Request, and executes the XN switching of the computing power service of the mobile terminal according to the network switching signaling, causing the change of the computing power node. The core network identifies the network switching signaling Handover Required, and executes N2 switching of the computing power service of the mobile terminal according to the network switching signaling, causing the computing power node to be changed.

7. The computing power management device for distributed AI training tasks in a mobile network according to claim 6, characterized in that The computing power management module performs XN switching, causing changes in computing power nodes, including: When the core network identifies the network switching signaling Path Switch Request, it switches the user SUPI, the target base station Target-RAN-ID, and the switching type to Xn through the computing-network convergence service orchestration network element. The computing-network convergence service orchestration network element finds the distributed AI training task identifier TASKID according to the user SUPI, and updates the computing node and training strategy of the AI ​​training task. When the Xn switch starts, it stops updating parameters with other computing nodes, notifies the target base station computing node to reserve computing resources and preheat the model, and synchronizes the old computing node parameters to the new computing node through the Xn interface between the source base station and the target base station. After the switch, the new computing node resumes computing based on the existing training experience of the old computing node and performs collaborative training synchronization with the cloud node.

8. The method for managing computing power of distributed AI training tasks in a mobile network according to claim 6, characterized in that The computing power management module performs N2 switching, causing changes in computing power nodes, including: When the core network recognizes the network switching signaling HandOver Requried, it switches the user SUPI, the target base station Target-RAN-ID and the switching type to N2 through the computing-network convergence service orchestration network element. The computing-network convergence service orchestration network element finds the distributed AI training task identifier TASKID according to the user SUPI, and updates the computing power node and training strategy of the AI ​​training task. In the N2 switching preparation phase, the target base station computing power node is notified to reserve computing power resources and preheat the model. When switching, it stops updating parameters with other computing power nodes, and synchronizes the old computing power node parameters to the new computing power node through the transmission tunnel between the source base station, the core network UPF and the target base station. After the switch, the new computing power node resumes computing based on the existing training experience of the old computing power node and performs collaborative training synchronization with the cloud node.

Images