Calculation unloading method and device, equipment and storage medium

Abstract

The invention provides a calculation unloading method and device, equipment and a storage medium. The calculation unloading method executed by the PEMC comprises the following steps: receiving an unloading request of a first PINE in a PIN; wherein the unloading request is used for requesting to unload a computing task of the first PINE; under the condition that a pre-acquired management strategy indicates that PIN calculation unloading is started, according to the management strategy, the unloading request and the current state of each PINE in the PIN, an unloading target used for executing the calculation task is determined in the PIN; and unloading the computing task of the first PINE to the unloading target. According to the embodiment of the invention, the calculation efficiency and quality in PIN can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of network technologies, and in particular, to a computing offloading method, apparatus, device, and storage medium. Background Art

[0002] A Personal IoT Network (PIN) is an intelligent service system based on personal intelligent devices and centered around users to link various scenarios. The PIN network can be used in scenarios such as medical care, education, home, and industry to realize the connection of people, things, and other information resources, and meet the high-quality and convenient living needs of individual users.

[0003] Devices in the PIN network may include: PINEs (PIN Elements with Management Capability, PEMC, also known as personal IoT management devices) with management capabilities, PINEs (PIN Element with Gateway Capability, PEGC, also known as personal IoT gateway devices) with gateway capabilities, and ordinary PINEs. These devices create a PIN network through signaling interaction with the 5GC (5G Core Network).

[0004] In related technologies, there are situations such as insufficient computing resources, diverse computing requirements, and difficult computing collaboration in the PIN network, resulting in low computing efficiency and quality in the PIN network. Summary of the Invention

[0005] Embodiments of the present disclosure provide a computing offloading method, apparatus, device, and storage medium.

[0006] In a first aspect, a computing offloading method is provided, which is executed by a PEMC; the method includes:

[0007] Receiving an offloading request from a first PINE in the PIN; wherein the offloading request is used to request offloading of the computing task of the first PINE;

[0008] When a management policy obtained in advance indicates to enable PIN computing offloading, determining, according to the management policy, the offloading request, and the current states of each PINE in the PIN, an offloading target in the PIN for executing the computing task;

[0009] Offloading the computing task of the first PINE to the offloading target.

[0010] In some embodiments, the method further includes:

[0011] Receive the management policy from a core network element or an application function network element.

[0012] In some embodiments, the method further includes:

[0013] Send an offloading status report for adjusting the management policy to the core network element or the application function network element; wherein, the offloading status report includes at least one of offloading efficiency, offloading quality, and offloading cost of the offloading operation for the computing task.

[0014] In some embodiments, the current state of each PINE in the PIN includes:

[0015] The currently reported computing resource status of multiple PINEs in the PIN, and the multiple PINEs include the first PINE and at least one second PINE;

[0016] The current network status of each PINE reported by the personal Internet of Things gateway device PEGC;

[0017] The artificial intelligence AI model supported by each PINE in the PIN.

[0018] In some embodiments, the offloading request carries the number of split layers and the computing performance requirements of the target AI model of the first PINE;

[0019] In the case where the pre-obtained management policy indicates to enable PIN computing offloading, according to the management policy, the offloading request, and the current state of each PINE in the PIN, determining an offloading target for executing the computing task in the PIN includes:

[0020] For each of the second PINEs, according to the management policy, match the current computing resource status and the current network status of the second PINE with the computing performance requirements of the first PINE to determine whether there is a candidate second PINE that meets the computing performance requirements;

[0021] When there is at least one candidate second PINE, match the number of split layers of the target AI model with the AI models supported by each candidate second PINE to determine a target second PINE with a successful match;

[0022] Determine the successfully matched target second PINE as the offloading target.

[0023] In some embodiments, the offloading of the computing task of the first PINE to the offloading target includes:

[0024] Transfer the computing tasks of some or all layers of the AI model of the first PINE to the offloading target.

[0025] In some embodiments, the method further includes:

[0026] Receive a recovery request sent by the first PINE; wherein, the recovery request is used to request the recovery of the computing tasks of the first PINE.

[0027] In response to the recovery request, restore the computing tasks unloaded by the first PINE to the first PINE.

[0028] In some embodiments, the recovery request carries the number of layers to be recovered of the AI model of the first PINE; the step of restoring the computing tasks unloaded by the first PINE to the first PINE in response to the recovery request includes:

[0029] In response to the recovery request, migrate the computing tasks of the number of layers to be recovered of the AI model requested to be recovered by the first PINE back to the first PINE from the offloading target.

[0030] In a second aspect, a computing offloading method is provided, which is executed by a first PINE; the method includes:

[0031] Send an offloading request to the PEMC; wherein, the offloading request is used to request the offloading of the computing tasks of the first PINE.

[0032] Receive a first response sent by the PEMC in response to the offloading request; wherein, the first response is used to indicate whether to offload the computing tasks requested to be offloaded by the first PINE.

[0033] In some embodiments, the method further includes:

[0034] Send a recovery request to the PEMC; wherein, the recovery request is used to request the recovery of the computing tasks unloaded by the first PINE.

[0035] Receive a second response sent by the PEMC in response to the recovery request; wherein, the second response is used to indicate that the computing tasks requested to be recovered by the first PINE are restored to the first PINE.

[0036] In some embodiments, the computing tasks of the first PINE include the computing tasks of some or all layers of the AI model of the first PINE.

[0037] In a third aspect, a computing offloading device is provided, which is applied to the PEMC; the device includes:

[0038] A receiving module, configured to receive an offloading request of a first PINE in a PIN; wherein, the offloading request is used to request offloading of the computing task of the first PINE.

[0039] A determining module, configured to, when a pre-acquired management policy indicates to enable PIN computing offloading, determine, according to the management policy, the offloading request, and the current states of the respective PINEs in the PIN, an offloading target in the PIN for executing the computing task.

[0040] An offloading module, configured to offload the computing task of the first PINE to the offloading target.

[0041] In a fourth aspect, there is provided a computing offloading device, applied to a first PINE; the device includes:

[0042] A sending module, configured to send an offloading request to a PEMC; wherein, the offloading request is used to request offloading of the computing task of the first PINE.

[0043] A receiving module, configured to receive a first response sent by the PEMC in response to the offloading request; wherein, the first response is used to indicate whether to offload the computing task requested to be offloaded by the first PINE.

[0044] In a fifth aspect, there is provided a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, the steps of the computing offloading method according to any one of the first aspect or the second aspect are implemented.

[0045] In a sixth aspect, there is provided a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the computing offloading method according to any one of the first aspect or the second aspect are implemented.

[0046] The embodiments of the present disclosure provide a computing offloading method, device, equipment, and storage medium. The offloading request of the first PINE in the PIN is received through the PEMC; the offloading request is used to request offloading of the computing task of the first PINE; when a pre-acquired management policy indicates to enable PIN computing offloading, an offloading target for executing the computing task is determined in the PIN according to the management policy, the offloading request, and the current states of the respective PINEs in the PIN, and the computing task of the first PINE is offloaded to the offloading target through the PEMC. In this way, through the computing offloading operation, the computing efficiency and quality in the PIN can be improved. Description of the Drawings

[0047] Figure 1 It is a schematic diagram of a PIN network architecture provided by an embodiment of the present disclosure.

[0048] Figure 2 Flow chart of a computing offloading method executed by PEMC provided by an embodiment of the present disclosure;

[0049] Figure 3 Flow chart of another computing offloading method executed by PEMC provided by an embodiment of the present disclosure;

[0050] Figure 4 Flow chart of yet another computing offloading method executed by PEMC provided by an embodiment of the present disclosure;

[0051] Figure 5 Flow chart of a computing offloading method executed by the first PINE provided by an embodiment of the present disclosure;

[0052] Figure 6 Flow chart of another computing offloading method executed by the first PINE provided by an embodiment of the present disclosure;

[0053] Figure 7 Interaction diagram of a PIN network autonomous AI computing offloading method provided by an embodiment of the present disclosure;

[0054] Figure 8 Interaction diagram of PEMC receiving a PIN autonomous policy provided by an embodiment of the present disclosure;

[0055] Figure 9 Interaction diagram of PEMC monitoring the status of PIN network AI computing resources provided by an embodiment of the present disclosure;

[0056] Figure 10 Interaction diagram of PINE A applying for AI computing offloading provided by an embodiment of the present disclosure;

[0057] Figure 11 Interaction diagram of PINE resuming autonomous computing provided by an embodiment of the present disclosure;

[0058] Figure 12 Structure diagram of a computing offloading device applied to PEMC provided by an embodiment of the present disclosure;

[0059] Figure 13 Structure diagram of a computing offloading device applied to the first PINE provided by an embodiment of the present disclosure;

[0060] Figure 14 Structure block diagram of a computer device provided by an embodiment of the present disclosure. Detailed implementation manners

[0061] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the embodiments of the present disclosure as detailed in the appended claims.

[0062] The terms used in the embodiments of the present disclosure are for the purpose of describing specific embodiments only and are not intended to limit the embodiments of the present disclosure. The singular forms "a", "the", and "said" used in the embodiments of the present disclosure and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0063] It should be understood that although the terms first, second, third, etc. may be used in the embodiments of the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the embodiments of the present disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "upon" or "in response to determining".

[0064] It can be understood that the present disclosure emphasizes the differences between the various embodiments, and the same or similar aspects can be referred to each other. For the sake of brevity, they will not be repeated one by one.

[0065] 3GPP (3rd Generation Partnership Project) defines the Personal Internet of Things Network (PIN). The battery life of Internet of Things devices within the PIN is typically several days or weeks. User plane traffic is usually forwarded in a restricted environment (such as around the body or indoors), that is, within the PIN. Notifications of events occurring within the PIN can be received on a smart terminal or other management device. PIN devices can communicate within the PIN using a variety of non-3GPP-based wireless technologies such as WLAN (Wireless Local Area Network), Z-Wave (wireless mesh network), Zigbee, and Bluetooth to allow users to interact and control. The PIN brings two benefits. First, since the user plane traffic generated by Internet of Things devices is small, traditional cellular operators do not need to specifically reserve resources in the network for Internet of Things devices, thus saving bandwidth resources. Second, the traffic generated by Internet of Things devices can stay local through a private network, facilitating user management and use.

[0066] Figure 1 Schematic diagram of a PIN network architecture provided for an embodiment of the present disclosure. As Figure 1 shown, the PIN network architecture includes PEMC, PEGC, and ordinary PINE (PIN Element, also referred to as "PIN device"). The ordinary PINE is connected to the PEGC through the P1 interface, and the PEGC is connected to the PEMC through the P2 interface; the ordinary PINE in the PIN network can be connected to the 5GC through the PEGC device and receive 5GC messages, or interact with the Application Function (AF) through the 5GC. The PEGC and PEMC are respectively connected to the AMF (Access and Mobility Management Function) network element in the 5GC through the N1 interface. Among them, the 5GC also includes: PCF (Policy Control Function) network element, AUSF (Authentication Server Function) network element, UDM (Unified Data Management) network element, etc.

[0067] PEGC: A PINE that can provide connections for ordinary PINE to enter and exit the 5G network or act as a relay for communication between ordinary PINEs.

[0068] PEMC: PINE with the ability to manage PINs, which includes a list of PINs, manages the behavior of each PIN, and controls the access rights of PINs. PEMC stores the following PIN information:

[0069] a) The unique identifier of the PIN within the PIN;

[0070] b) The connection types supported by the PIN;

[0071] c) The identification of the application on the PIN (such as the application ID);

[0072] d) Related metadata (such as event occurrence, event type, timestamp, etc.);

[0073] e) The security credentials of the PIN, etc.

[0074] In the related art, there are situations such as insufficient computing resources, diverse computing requirements, and difficult computing collaboration in the PIN network, resulting in low computing efficiency and quality in the PIN network.

[0075] Insufficient computing resources: Personal intelligent devices in the PIN network are often limited by factors such as battery capacity, processor performance, and storage space, and cannot carry out complex artificial intelligence (AI) computing tasks. This requires allocating AI computing tasks to the cloud or the edge, but this will bring problems such as network latency, data security, and communication costs.

[0076] Diverse computing requirements: Users in the PIN network may have different computing requirements, such as speech recognition, image processing, video analysis, recommendation systems, etc. This requires providing customized AI computing services according to different application scenarios and user preferences, but this will increase the difficulty of service management and maintenance.

[0077] Difficult computing collaboration: Personal intelligent devices in the PIN network may need to perform collaborative computing with other devices or services to achieve more efficient or higher-quality AI functions. This requires establishing an effective collaboration mechanism to solve problems such as device discovery, task allocation, data sharing, and result fusion, but this will also involve issues such as device compatibility, trust mechanism, and privacy protection.

[0078] Figure 2 The flowchart of a computing offloading method provided by an embodiment of the present disclosure is shown. This method is executed by a personal Internet of Things management device (PEMC), as Figure 2 shown, this method may include the steps:

[0079] 201: Receive an offloading request of a first PIN in the PIN; wherein, the offloading request is used to request offloading the computing task of the first PIN;

[0080] 202: When the pre-acquired management policy indicates to enable PIN computing offloading, determine, in the PIN, an offloading target for executing a computing task according to the management policy, the offloading request, and the current status of each PINE in the PIN.

[0081] 203: Offload the computing task of the first PINE to the offloading target.

[0082] Here, the PEMC can be a device for managing each PINE in the PIN. Exemplarily, the PEMC can be a UE (User Equipment), such as a mobile phone, a computer, a tablet computer, etc.; the PINE can be a device such as a sweeping robot, a wearable device, a speaker, a TV, etc.

[0083] For example, in the scenario where the PIN network is applied to the intelligent health care field, a 5G tablet computer can be used as the PEMC, and various types of intelligent monitoring, treatment, and rehabilitation devices such as intelligent wearable devices, intelligent blood oxygen meters, intelligent blood pressure monitors, intelligent body fat scales, and rehabilitation trainers can be used as ordinary PINEs to accept the unified management and control of the PEMC, and a 5G UE or a gateway device can be used as the PEGC to transmit the collected patient data to the hospital server through the 5G network.

[0084] Here, the first PINE can be any ordinary PINE in the PIN that needs to request computing offloading.

[0085] Here, the offloading request of the first PINE is used to request offloading of the computing task of the first PINE, and the computing task can be an AI computing task. For example, the computing task of the first PINE may include: the computing tasks of some or all layers of the AI model of the first PINE.

[0086] Here, the management policy pre-acquired by the PEMC (which can be referred to as the "PIN autonomy policy") is used to indicate whether to enable PIN computing offloading. Among them, whether to enable PIN AI autonomous computing offloading can determine whether the computing task of the PINE can be offloaded to appropriate computing resources to achieve distributed computing and resource optimization.

[0087] When the management policy indicates to enable PIN computing offloading, the management policy may include offloading parameters. The offloading parameters can be used to indicate at least one of an offloading target, an offloading algorithm, an offloading threshold, an offloading frequency, and an offloading priority.

[0088] The offloading target is a device or service selected by the PEMC for executing a computing task, such as a cloud server, an edge node, another PINE, a PEGC, etc. The selection of the offloading target depends on various factors, such as the computing power of the offloading target, network quality, AI model support, etc.

[0089] The offloading algorithm is what PEMC uses to decide whether to offload and the offloading strategy. According to different objectives and constraints (such as latency, energy consumption, accuracy, etc.), the offloading algorithm can adopt different methods, such as using machine learning or deep learning to optimize the effect and speed of offloading decisions.

[0090] The offloading threshold is a parameter or condition that PEMC uses to determine whether to offload. The offloading threshold can be set according to different metrics, such as battery power, resources, latency, etc. When a certain metric reaches or exceeds the offloading threshold, PEMC will determine whether to initiate the offloading operation, thus ensuring the execution of computing tasks and the management of resources.

[0091] The offloading frequency is the time interval or number of times that PEMC performs offloading operations. The offloading frequency can be adjusted according to different requirements and scenarios, such as real-time performance, stability, reliability, etc. The offloading frequency can affect the offloading effect and overhead.

[0092] The offloading priority is a parameter or rule that PEMC uses to sort or rank different computing tasks. The offloading priority can be set according to different criteria, such as urgency, importance, complexity, etc. When PEMC faces multiple AI computing tasks, it will select which tasks to offload and which tasks to keep for local computing according to the offloading priority.

[0093] Here, the current state of PINE may include the current computing resources of PINE, the current network state, and the supported AI models, etc.

[0094] The computing resource state of PINE is used to indicate information such as the available computing power, memory, storage, etc. of PINE.

[0095] The network state of PINE is used to indicate information such as the bandwidth, latency, packet loss rate, etc. of the network where PINE is located.

[0096] The AI model can be used for speech recognition, image recognition, video processing, etc.

[0097] In some examples, the current state of each PINE in the PIN includes:

[0098] The currently reported computing resource state of multiple PINEs in the PIN, where the multiple PINEs include a first PINE and at least one second PINE;

[0099] The currently reported network state of each PINE reported by PEGC;

[0100] The AI models supported by each PINE in the PIN.

[0101] Here, ordinary PINEs and PEGCs can actively report the status information of their own computing resources to the PEMC. This computing resource status information enables the PEMC to understand the computing resource situation of each ordinary PINE and PEGC, so as to make decisions on task allocation and offloading according to needs.

[0102] Here, ordinary PINEs and PEGCs can actively report the status information of their own networks to the PEMC. This network status information helps the PEMC evaluate the communication quality between PINEs, PEGCs and other networking components, so as to make decisions on task allocation and offloading.

[0103] Here, each PINE can report the AI model information it supports to the PEMC. The AI model information includes model type, version, computing requirements, etc. For example, the AI model information enables the PEMC to select appropriate other PINEs for task allocation and offloading according to the computing performance requirements and model requirements of the computing tasks of the first PINE.

[0104] In addition to active reporting, ordinary PINEs or PEGCs can also pass information to the PEMC in a passive reporting manner. In this case, the PEMC will obtain the resource information and network information of PINEs / PEGCs by pulling. The PEMC regularly sends requests to PINEs / PEGCs, asking them to provide computing resource status, network status and other relevant information. After receiving the requests, PINEs / PEGCs pass the corresponding information to the PEMC. The passive reporting method can enable the PEMC to flexibly obtain the information of PINEs / PEGCs according to needs, so as to timely understand the status and capabilities of each component in the network.

[0105] In step 201 above, the offloading request of the first PINE can be sent by the first PINE to the PEMC when the power is insufficient or the computing resources are insufficient.

[0106] Here, the insufficient power of the first PINE can include: the current power of the first PINE is less than the preset power and / or the ratio of the current power of the first PINE to the fully charged power is less than the preset ratio. The insufficient computing resources of the first PINE can include: the current computing resources of the first PINE cannot meet the computing resource requirements.

[0107] In some examples, the computing tasks requested to be offloaded by the first PINE can be: the computing tasks of some or all layers of the AI model of the first PINE.

[0108] Here, the offloading request of the first PINE can carry the number of split layers of the target AI model of the first PINE and the computing performance requirements.

[0109] The number of split layers of the target AI model can indicate which layers of the target AI model that the first PINE hopes to offload; the computing performance requirement represents the requirements of the first PINE for the computing power, network quality, latency, etc. of the offloading target.

[0110] Here, when the current power or computing resources of the first PINE are insufficient, it can generate information such as the number of split layers of the AI model and the computing performance requirement according to at least one of model quantization analysis, resource evaluation, latency and network requirements, and application scenarios, and carry the number of split layers of the AI model and the computing performance requirement in the offloading request sent to the PEMC.

[0111] In step 202 above, the PEMC can determine the offloading target for executing the computing task of the first PINE in the PIN according to the number of split layers of the target AI model and the computing performance requirement carried in the offloading request, and in combination with the current computing resource status, current network status of each PINE in the PIN, and the AI models supported by each PINE.

[0112] Here, the PEMC can select a suitable device or service as the offloading target to offload the computing task of the first PINE to meet the computing performance requirement of the first PINE.

[0113] The embodiments of the present disclosure provide a computing offloading method, which receives an offloading request from the first PINE in the PIN through the PEMC; the offloading request is used to request to offload the computing task of the first PINE; in the case where the pre-acquired management policy indicates to enable PIN computing offloading, according to the management policy, the offloading request, and the current status of each PINE in the PIN, determine the offloading target for executing the computing task in the PIN, and offload the computing task of the first PINE to the offloading target through the PEMC. In this way, through the computing offloading operation, the computing efficiency and quality in the PIN can be improved, and the computing cost and risk in the PIN can be reduced.

[0114] In some embodiments, the method may further include:

[0115] Receive a management policy from a core network element or an application function network element;

[0116] Here, the management policy may be sent by a core network element (such as a PCF) or an application function (AF) network element to the PEMC according to at least one of user requirements, network conditions, and resource availability.

[0117] In the embodiments of the present disclosure, the PEMC receives management policies from core network elements or application function network elements. In this way, when the management policy indicates to enable PIN computing offloading, the PEMC can coordinate the distribution and offloading of computing tasks to achieve optimized resource utilization. When the management policy indicates to disable PIN computing offloading, the PEMC will stop the offloading operation of the computing tasks and resume the local computing of the PINE, thus realizing the adaptive management of PIN computing offloading.

[0118] In some embodiments, the method may further include:

[0119] Sending an offloading status report for adjusting the management policy to a core network element or an application function network element; wherein, the offloading status report includes at least one of offloading efficiency, offloading quality, and offloading cost for the offloading operation of the computing task.

[0120] Here, the offloading efficiency is measured by the time or energy consumption saved by the offloading operation. An efficient offloading operation can reduce the local computing burden and improve the execution efficiency of the computing task.

[0121] The offloading quality is an indicator of the accuracy or credibility of the calculation result obtained by the offloading operation. A high-quality offloading operation can ensure that the offloaded calculation result is consistent with the local calculation result, thus maintaining the performance of the computing task.

[0122] The offloading cost refers to the resources or financial expenses involved in the offloading operation. A low-cost offloading operation can provide a more cost-effective management of computing tasks while ensuring the effective utilization of resources.

[0123] In the embodiments of the present disclosure, by reporting the offloading status report from the PEMC to the core network element or the application function network element, the core network element or the application function network element can dynamically adjust the management policy of the PIN according to the offloading status report, achieving the purpose of optimizing the user experience and network performance.

[0124] In some embodiments, the offloading request carries the number of split layers of the target AI model of the first PINE and the computing performance requirements. As Figure 3 shown, in step 202 above, when the pre-acquired management policy indicates to enable PIN computing offloading, according to the management policy, the offloading request, and the current status of each PINE in the PIN, determining the offloading target for executing the computing task in the PIN may include the steps:

[0125] 301: For each second PINE, according to the management policy, match the current computing resource status, the current network status of the second PINE with the computing performance requirements of the first PINE to determine whether there is a candidate second PINE that meets the computing performance requirements;

[0126] 302: When there is at least one candidate second PINE, match the number of split layers of the target AI model with the AI models supported by each candidate second PINE to determine the target second PINE with successful matching.

[0127] 303: Determine the successfully matched target second PINE as the offloading target.

[0128] Here, the second PINE may include ordinary PINEs other than the first PINE in the PIN and PEGC.

[0129] In step 301 above, for each second PINE, the PEMC can evaluate whether each second PINE has sufficient computing resources to undertake the offloading task of the first PINE according to the offloading parameters in the management policy, the computing resource status of each second PINE, and the computing performance requirements of the first PINE. The PEMC evaluates whether the network quality of each second PINE can meet the computing performance requirements of the PINE according to the network status of each second PINE and the computing performance requirements of the first PINE.

[0130] Here, the offloading parameters in the management policy can be used to indicate at least one of the offloading target, offloading algorithm, offloading threshold, offloading frequency, and offloading priority. The computing resource status can include, for example, computing power, memory, storage status, etc. The network status of the offloading target can include bandwidth, latency, packet loss rate, etc.

[0131] In this embodiment, by matching the computing resource status and network status of each second PINE with the computing performance requirements of the first PINE through the PEMC, it is possible to determine whether there is a candidate second PINE, and the computing resource status and network status of the candidate second PINE meet the computing performance requirements of the first PINE.

[0132] In step 302 above, the PEMC can match the number of split layers of the target AI model with the AI models supported by each candidate second PINE to determine whether the candidate second PINE supports the AI model required by the first PINE. The matching process here may include verifying whether the candidate second PINE has the AI model version and computing power required by the first PINE to ensure that the candidate second PINE can successfully offload the AI computing task of the first PINE.

[0133] In some examples, when a target second PINE is matched, the target second PINE can be used as the offloading target, so that the offloading target is used to calculate the computing task requested to be offloaded by the first PINE.

[0134] In some other examples, when multiple target second PINEs are matched, the multiple target second PINEs can be respectively used as offloading targets, so that the multiple offloading targets are used to collaboratively compute the computing tasks requested to be offloaded by the first PINE. The multiple target second PINEs may include multiple ordinary PINEs, or the multiple target second PINEs may include at least one ordinary PINE and a PEGC.

[0135] In still some other examples, when no candidate second PINE or target second PINE is matched, a response message for the offloading request is returned to the first PINE; wherein, the response message is used to indicate that there is currently no second PINE for offloading the computing tasks requested to be offloaded by the first PINE.

[0136] In some examples, the offloading target may include at least one of the following: other PINEs located close to the first PINE, the PEGC accessed by the first PINE, other PEGCs that the first PINE can access, the cloud, etc.

[0137] In the embodiments of the present disclosure, by determining an offloading target for offloading AI computing tasks according to management policies, the computing performance requirements of the first PINE, and the current computing resources status, current network status, and supported AI models of each second PINE in the PIN, it is beneficial to improve the AI computing efficiency and quality in the PIN network, and enhance network performance and user experience.

[0138] In some embodiments, in step 203 above, offloading the computing tasks of the first PINE to the offloading target may include:

[0139] Transferring the computing tasks of some or all layers of the AI model of the first PINE to the offloading target.

[0140] Here, the PEMC transfers the computing tasks of some or all layers of the AI model of the first PINE to the offloading target, so that the AI computing tasks of the first PINE are offloaded to a suitable device or service, realizing the splitting and migration of the AI model.

[0141] In the embodiments of the present disclosure, by transferring the computing tasks of some or all layers of the AI model of the first PINE to the offloading target, the splitting and migration of the computing tasks of the AI model of the PINE in the PIN can be realized, making the allocation and collaborative computing of the computing resources in the PIN more intelligent and efficient, and bringing a better computing experience to the personal Internet of Things network.

[0142] In some embodiments, as Figure 4 shown, the method may further include the steps:

[0143] 401: Receive the recovery request sent by the first PINE; wherein, the recovery request is used to request the recovery of the computing task of the first PINE.

[0144] 402: In response to the recovery request, restore the computing tasks that have been offloaded from the first PINE to the first PINE.

[0145] In step 401 above, after the power or computing resources of the first PINE are sufficient, it can send a recovery request to the PEMC to apply for the restoration of autonomous computing.

[0146] In step 402 above, after the PEMC receives the recovery request of the first PINE, it can restore the AI computing tasks that have been offloaded from the first PINE to the first PINE.

[0147] In the embodiments of the present disclosure, by the PEMC responding to the recovery request of the first PINE and restoring the computing tasks that have been offloaded from the first PINE to the first PINE, it is beneficial to improve the AI computing efficiency and quality in the PIN network, and enhance the network performance and user experience.

[0148] In some embodiments, the recovery request carries the number of layers to be restored of the AI model of the first PINE.

[0149] Here, the recovery request carries the number of layers to be restored. The number of layers to be restored indicates which layers of the AI model the first PINE hopes to restore, that is, migrate back from the offloading target to the first PINE.

[0150] Here, the number of layers to be restored can be set by the first PINE according to different tasks and scenarios. For example, the first PINE sets the number of layers to be restored according to at least one of the model complexity, data scale, and computing resources. Generally speaking, the more the number of layers to be restored, the smaller the amount of data to be restored, but the computing efficiency and quality may be improved.

[0151] In step 402 above, in response to the recovery request, restoring the computing tasks that have been offloaded from the first PINE to the first PINE may include:

[0152] In response to the recovery request, migrate the computing tasks of the number of layers to be restored of the AI model requested by the first PINE to be restored from the offloading target back to the first PINE.

[0153] In the embodiments of the present disclosure, the PEMC in response to the recovery request of the first PINE can migrate the computing tasks of some or all layers of the AI model requested by the first PINE to be restored back to the first PINE, thus realizing the migration of the computing tasks of the AI model of the PINE in the PIN, making the allocation and collaborative computing of the computing resources in the PIN more intelligent and efficient, and bringing a better computing experience to the personal Internet of Things network.

[0154] Based on the computing offloading method performed by the PEMC described in the foregoing embodiments, an embodiment of the present disclosure also provides a computing offloading method performed by a first PINE.

[0155] Figure 5 The flowchart of a computing offloading method provided by an embodiment of the present disclosure is shown. This method is performed by a first PINE; as Figure 5 shown, the method may include the steps:

[0156] 501: Send an offloading request to the PEMC; wherein, the offloading request is used to request offloading of the computing tasks of the first PINE;

[0157] 502: Receive a first response sent by the PEMC in response to the offloading request; wherein, the first response is used to indicate whether to offload the computing tasks requested to be offloaded by the first PINE.

[0158] The offloading request carries the number of split layers of the target AI model of the first PINE and the computing performance requirements.

[0159] The computing tasks of the first PINE include: the computing tasks of some or all layers of the target AI model of the first PINE.

[0160] The first PINE may generate information such as the number of split layers of the target AI model and the computing performance requirements according to at least one of factors such as model quantization analysis, resource assessment, latency and network requirements, and application scenarios.

[0161] The number of split layers indicates which layers of the AI model the first PINE hopes to offload. The computing performance requirements indicate the requirements of the first PINE for the computing power, network quality, latency, etc. of the offloading target.

[0162] In step 501 above, when the power or resources of the first PINE are insufficient, it may send an offloading request to the PEMC to request offloading of the computing tasks of the first PINE.

[0163] In some examples, the first PINE may determine which layers are suitable for offloading by analyzing the structure and computational amount of the model. Usually, some layers with relatively large forward computational overhead in the AI model will be candidate layers. According to the characteristics of the model, the first PINE may adopt a predetermined rule or heuristic algorithm to select the number of split layers to achieve better performance improvement.

[0164] In some examples, the first PINE needs to evaluate its own power, computing power, network quality and other resource conditions to determine which layers in the AI model are suitable for offloading. If the power of the first PINE is insufficient or the computing power is insufficient, the first PINE may tend to select a smaller number of split layers to reduce the computing burden. On the contrary, if the first PINE has sufficient computing resources, deeper split layers can be considered to obtain better offloading effects.

[0165] In some examples, the first PINE may determine the number of split layers according to the real-time requirements of the task and the network latency. If the task requires low latency, the first PINE may choose a smaller number of split layers to reduce the offloading and transmission time. At the same time, the first PINE can also consider the network condition and select the number of split layers and offloading targets suitable for the current network environment.

[0166] In some examples, different application scenarios may have different requirements for the number of split layers and performance requirements. For example, for tasks with high real-time requirements, a smaller number of split layers may be more preferred to reduce the computing and transmission latency.

[0167] In some examples, the first PINE determines the completed part and the uncompleted part in the computing task and decides which parts should be offloaded to other PINEs, which may include the following steps:

[0168] Task analysis: The first PINE can analyze the currently executed computing task to obtain the structure, computing volume and computing time consumption of each stage of the computing task, so as to determine the split points of the computing task and which parts can be offloaded.

[0169] Split point determination: The first PINE can determine appropriate split points according to the structure, computing volume and computing time consumption of the computing task. The split point refers to one or more positions in the computing task, where the computing task is divided into a completed part and an uncompleted part at these positions. Usually, the split point can be selected at a stage with a large computing volume or a long time consumption.

[0170] Status saving: After determining the split points, the first PINE saves the status information of the computing task. The status information of the computing task includes the currently completed part and the uncompleted part, such as model parameters, intermediate computing results, etc.

[0171] Request for offloading: When the power or resources of the first PINE are insufficient, it sends an offloading request to the PEMC. The request carries information such as the number of split layers and computing performance requirements. At this time, the first PINE also sends the saved task status information to the PEMC for subsequent offloading operations.

[0172] In step 502 above, after the first PINE receives the first response sent by the PEMC in response to the unloading request, it can determine whether the PEMC performs an unloading operation on the computing task requested by the first PINE to be unloaded based on this first response.

[0173] After receiving the unloading request from the first PINE, the PEMC will comprehensively consider information such as the network status, computing resource status of other PINEs in the PIN network, and the number of split layers and computing performance requirements carried in the unloading request to decide whether to unload at least part of the computing task of the first PINE to other PINEs, which may include determining the unloading target (i.e., the appropriate PINE node) so that at least part of the computing task requested by the first PINE to be unloaded can obtain appropriate computing resources and performance support.

[0174] When the PEMC determines the unloading target, the PEMC will coordinate the sending of at least part of the computing task of the first PINE to the target PINE node. After receiving at least part of the computing task of the first PINE, the target PINE will continue the calculation based on the task status and the unloaded part of the computing task of the first PINE to complete the entire computing task.

[0175] Each target PINE node will separately complete different parts of the computing task of the first PINE, and then send the calculation results to the first PINE for the first PINE to integrate them into a complete task result. This process may involve communication interaction and data exchange to ensure the consistency of the task results.

[0176] The computing unloading method provided by the embodiments of the present disclosure includes: sending an unloading request from the first PINE to the PEMC; the unloading request is used to request the unloading of the computing task of the first PINE; and receiving the first response sent by the PEMC in response to the unloading request; so that the first PINE can determine whether the PEMC performs an unloading operation on the computing task requested by the first PINE to be unloaded based on the first response. In this way, by the PEMC performing an unloading operation on the computing task requested by the first PINE to be unloaded, the computing efficiency and quality in the PIN can be improved, and the computing cost and risk in the PIN can be reduced.

[0177] In some embodiments, as Figure 6 shown, the method may further include the steps:

[0178] 601: Sending a recovery request to the PEMC; wherein, the recovery request is used to request the recovery of the computing task unloaded by the first PINE;

[0179] 602: Receiving the second response sent by the PEMC in response to the recovery request; wherein, the second response is used to indicate that the computing task requested by the first PINE to be recovered is recovered to the first PINE.

[0180] In step 601 above, after the first PINE has sufficient power or computing resources, it can send a recovery request to the PEMC to apply for the recovery of autonomous computing.

[0181] In step 602 above, the first PINE can receive the AI model layer transmitted by the offloading target through the PEMC, and integrate the AI model layer with the local AI model layer to recover autonomous computing. After the first PINE recovers autonomous computing, the first PINE will disconnect the communication connection with the offloading target and release resources.

[0182] In the embodiments of the present disclosure, by receiving, through the first PINE, the second response sent by the PEMC in response to the recovery request of the first PINE, the computing tasks of some or all layers of the AI model of the first PINE can be migrated back from the offloading target to the first PINE according to the second response. In this way, the migration of the computing tasks of the AI model in the PIN is realized, making the allocation of computing resources and collaborative computing in the PIN more intelligent and efficient, and bringing a better computing experience to the personal Internet of Things network.

[0183] The technical solution provided by the present disclosure will be further described in detail below with specific examples.

[0184] The embodiments of the present disclosure provide a PIN network autonomous AI computing offloading method. By introducing an autonomous computing offloading mechanism based on artificial intelligence in the 5G communication network, it makes up for the gap in the lack of an AI computing offloading mechanism in the current "PIN + AI" scenario, solves problems such as insufficient computing resources, diverse computing requirements, and difficult computing collaboration. The autonomous management of the PIN network makes the allocation of computing resources and collaborative computing more intelligent and efficient, bringing a better computing experience to the PIN network.

[0185] Figure 7 It is an interaction schematic diagram of a PIN network autonomous AI computing offloading method provided by the embodiments of the present disclosure. As Figure 7 shown, the computing offloading method may include:

[0186] 1: The PEMC receives the PIN autonomous policy.

[0187] In this process, the 5GC / AF sends the PIN autonomous policy (i.e., the management policy in the foregoing embodiments) to the PEMC, indicating whether to enable PIN AI autonomous computing offloading, as well as related parameters and conditions. After receiving the PIN autonomous policy, the PEMC performs corresponding configurations and operations according to the policy content of the PIN autonomous policy.

[0188] 2: The PEMC monitors the status of the PIN network AI computing resources.

[0189] During this process, PEMC can monitor the AI computing resource status of PEGC and PINE in the PIN network in real time, including computing resources, network status, supported AI models, etc. PEMC can dynamically adjust the allocation and offloading of AI computing tasks in the PIN network according to the monitoring results.

[0190] 3: PINE requests AI computing offloading.

[0191] During this process, when the power or resources of PINE are insufficient, it requests AI computing offloading from PEMC, and carries information such as the number of split layers and computing performance requirements in the request. After receiving the request from PINE, PEMC comprehensively judges the offloading target, including other PINEs adjacent to this PINE, the PEGC to which PINE is connected, other PEGCs that PINE can access, the cloud, etc. PEMC offloads the AI computing tasks of PINE to appropriate devices or services according to the offloading target. For example, PEMC offloads the AI computing tasks of PINE to appropriate devices or services through the AI model transfer method based on the PIN architecture.

[0192] 4: PINE resumes autonomous computing.

[0193] During this process, after the power or resources of PINE are sufficient, it requests to resume autonomous computing from PEMC, and carries information such as the number of resumed layers in the request. After receiving the request from PINE, PEMC restores the previously offloaded AI computing tasks of PINE to PINE. For example, PEMC restores the previously offloaded AI computing tasks of PINE to PINE through the AI model transfer method based on the PIN architecture.

[0194] Figure 8 It is an interaction schematic diagram for PEMC to receive the PIN autonomy policy provided by this embodiment of the disclosure. As Figure 8 shown, the implementation process for PEMC to receive the PIN autonomy policy may include:

[0195] Step S11: 5GC / AF sends the PIN autonomy policy to PEMC.

[0196] The 5G core network (5GC) or application function (AF) can send the PIN autonomy policy to PEMC according to factors such as user requirements, network conditions, resource availability, etc., for indicating whether to enable PIN AI autonomous computing offloading, as well as related parameters and conditions. Here, the parameters and conditions may include offloading target, offloading algorithm, offloading threshold, offloading frequency, offloading priority, etc.

[0197] Step S12: PEMC performs corresponding configurations and operations according to the PIN autonomy policy.

[0198] Here, upon receiving the offloading request from PINE, PEMC can determine the offloading target according to the PIN autonomy policy, and send the offloading policy and instructions to PINE based on the offloading target.

[0199] For example, the offloading policy and instructions sent by PEMC to PINE may include the following:

[0200] 1) Instructions to turn on or off PIN AI autonomous computing offloading.

[0201] According to the PIN autonomy policy issued by 5GC / AF, PEMC will send instructions to PINE indicating whether to turn on PIN AI autonomous computing offloading, which determines whether the computing tasks of PINE can be offloaded to appropriate computing resources to achieve distributed computing and resource optimization. After the AI computing offloading is turned on, PEMC will coordinate the distribution and offloading of computing tasks to optimize resource utilization.

[0202] 2) Offloading target: PEMC will send the relevant parameter information of the offloading target to PINE. The offloading target refers to the computing resource to which the computing task is to be offloaded, such as a cloud server, an edge node, other PINEs, PEGC, etc. PEMC transfers the corresponding target information to PINE according to the offloading target parameters specified in the policy.

[0203] In step S12, PEMC will send the parameter information of the offloading target to PINE. This parameter determines which computing resource the computing task will be offloaded to. The offloading target can include a cloud server, an edge node, or a device with stronger computing power. The selection of the offloading target will affect the offloading effect and performance.

[0204] 3) Offloading algorithm and parameters: PEMC will transfer the offloading algorithm and parameters to PINE according to the instructions and conditions in the policy. These algorithms and parameters are used to select appropriate computing resources for the offloading operation. For example, the offloading algorithm can select the optimal computing node according to different criteria, such as resource availability, network conditions, load, etc. The offloading threshold is the condition for triggering offloading. For example, when the load of a certain node exceeds a certain threshold, the offloading operation will be triggered.

[0205] If the autonomy policy indicates to turn on PIN AI autonomous computing offloading, then PEMC can, upon receiving the offloading request from PINE, select appropriate computing resources for offloading according to parameters such as the offloading target, offloading algorithm, offloading threshold, etc., such as a cloud server, an edge node, other PINEs, PEGC, etc. If the policy indicates to turn off PIN AI autonomous computing offloading, then PEMC stops the current offloading operation and resumes the local computing of PINE.

[0206] Step S13: The PEMC reports the offloading status and results regularly or on demand.

[0207] The PEMC can report the offloading status and results to the 5GC / AF regularly or on demand according to parameters such as offloading frequency and offloading priority, including but not limited to offloading efficiency, offloading quality, offloading cost, etc., so that the 5GC / AF can dynamically adjust the PIN autonomous policy based on the feedback information of the PEMC, which can optimize the user experience and network performance.

[0208] In the embodiments of the present disclosure, the PEMC realizes the adaptive management of AI autonomous computing offloading according to the PIN autonomous policy issued by the 5GC / AF.

[0209] Figure 9 It is an interactive schematic diagram for the PEMC provided by the embodiments of the present disclosure to monitor the AI computing resource status of the PIN network. The PEMC can monitor the AI computing resource status of the PIN network and dynamically adjust the allocation and offloading of AI computing tasks.

[0210] As Figure 9 shown, the implementation process of the PEMC monitoring the AI computing resource status of the PIN network may include:

[0211] Step S21: The PINE reports the PINE computing resources to the PEMC.

[0212] Step S22: The PEGC reports the PEGC network resources to the PEMC.

[0213] The PEMC monitors the AI computing resource status of the PINE and PEGC in the PIN network in real time, including computing resources, network status, supported AI models, etc.

[0214] Computing resource status: The PINE and PEGC will actively report the status of their computing resources, including information such as available computing power, memory, storage, etc. This information enables the PEMC to understand the computing resource situation of each PINE and PEGC, so as to make decisions on task allocation and offloading according to needs.

[0215] Network status: The PINE and PEGC will report the status of the network they are in, including indicators such as bandwidth, latency, packet loss rate, etc. This information helps the PEMC evaluate the communication quality between the PINE, PEGC and other networking components, so as to make decisions on task allocation and offloading.

[0216] Supported AI models: The PINE can report the information of the supported AI models, including model type, version, computing requirements, etc. This enables the PEMC to select the appropriate PINE for task allocation and offloading according to the requirements of the task and the model.

[0217] By collecting the information actively reported by PINE, PEMC can timely understand the computing resource status, network status, and supported AI models of each component in the PIN network, providing a basis for subsequent task allocation and offloading decisions.

[0218] PINE and PEGC synchronize computing resources to PEMC through active reporting, or PEMC obtains the resource information and network information of PINEs in the network through pulling. In addition to active reporting, PINE / PEGC can also pass information to PEMC through passive reporting. In this case, PEMC will obtain the resource information and network information of PINE / PEGC through pulling. PEMC regularly sends requests to PINE / PEGC, asking it to provide computing resource status, network status, and other relevant information. After receiving the request, PINE / PEGC passes the corresponding information to PEMC. The passive reporting method enables PEMC to flexibly obtain the information of PINE / PEGC as needed, so as to timely understand the status and capabilities of each component in the network.

[0219] Step S23: PEMC monitors and records the operating status of the PIN network.

[0220] PEMC monitors and records the AI computing resource status in the PIN network. According to factors such as the computing power, network quality, AI model support, and power of each PINE, PEMC dynamically adjusts the allocation and offloading of AI computing tasks. Currently, PEMC only monitors and records the resource situation of each PINE, and does not actively perform task allocation and offloading operations. When a PINE indicates through active reporting that it needs to offload a task to PEMC, PEMC then performs the corresponding computing offloading operation.

[0221] The above process is executed in a loop until PEMC receives a new PIN autonomy policy or a termination signal.

[0222] Figure 10 It is an interaction schematic diagram for PINE A in this embodiment of the disclosure to apply for AI computing offloading. PINE can apply for the process of AI computing offloading to split and migrate AI computing tasks.

[0223] As Figure 10 shown, the implementation process for PINE A to apply for AI computing offloading can include:

[0224] Step S31: PINE A applies to PEMC for computing offloading.

[0225] When PINE has insufficient power or resources, it applies to PEMC for AI computing offloading. Based on factors such as model quantization analysis, resource assessment, latency, network requirements, and application scenarios, it generates information such as the number of split layers and computing performance requirements. The number of split layers indicates which layers of the AI model PINE hopes to offload, and the computing performance requirements represent PINE's requirements for the computing power, network quality, latency, etc. of the offloading target.

[0226] Model quantization analysis: PINE can determine which layers are suitable for offloading by analyzing the structure and computational volume of the AI model. Usually, some layers with large forward computational overhead in the AI model will be candidate layers. According to the characteristics of the AI model, PINE can adopt certain rules or heuristic algorithms to select the number of split layers to achieve better performance improvement.

[0227] Resource assessment: PINE needs to assess its own resources such as power, computing power, and network quality. If PINE has insufficient power or computing power, it may tend to choose a smaller number of split layers to reduce the burden. On the contrary, if PINE has sufficient resources, it can consider deeper split layers to obtain better offloading effects.

[0228] Latency and network requirements: PINE may determine the number of split layers according to the real-time requirements of the task and network latency. If the task requires low latency, PINE may choose a smaller number of split layers to reduce the offloading and transmission time. At the same time, PINE can also consider the network condition and choose the number of split layers and offloading target suitable for the current network environment.

[0229] Application scenarios: Different application scenarios may have different requirements for the number of split layers and performance requirements. For example, for tasks with high real-time requirements, it may be more inclined to choose a smaller number of split layers to reduce computational and transmission latency.

[0230] To determine the completed and uncompleted parts of the task and decide which parts should be offloaded to other PINEs, the following steps and strategies are usually required:

[0231] 1) Task analysis: PINE first needs to analyze the currently executed task to understand the task's structure, computational volume, and computational time consumption at each stage. This helps to determine the split points of the task and which parts can be offloaded.

[0232] 2) Split point determination: PINE can determine the appropriate split points according to the task's structure and computational time consumption. The split points refer to one or more positions in the task where the task is divided into completed and uncompleted parts. Usually, the split points will be selected at stages with large computational volume or long time consumption.

[0233] 3) State saving: After determining the splitting point, PINE needs to save the current state of the task, including the completed part and the uncompleted part. This can include model parameters, intermediate calculation results, etc.

[0234] 4) Application for offloading: When the power or resources of PINE are insufficient, it applies to PEMC for offloading, carrying information such as the number of split layers and the computing performance requirements. At this time, PINE also sends the saved task status information to PEMC together for subsequent offloading operations.

[0235] 5) Offloading strategy: After receiving the application from PINE, PEMC will comprehensively consider information such as the status of other PINEs in the network, the computing resource status, and the number of split layers to decide whether to offload part of the task to other PINEs. This may include identifying suitable PINE nodes so that the offloaded calculation can obtain appropriate computing resources and performance support.

[0236] 6) Offloading execution: Once PEMC determines the offloading target, it will coordinate to send part of the task to the target PINE node. After receiving the part of the task, the target PINE will continue the calculation according to the task status and the offloaded part to complete the entire task.

[0237] 7) Result integration: Finally, each PINE node will complete different parts of the task respectively, and then integrate the results to obtain the complete task result. This may involve communication and data exchange to ensure the consistency of the task results.

[0238] In summary, determining the completed part and the uncompleted part of the task, and deciding which parts to offload to other PINEs require analyzing the task structure, computing time consumption, and resource status, and also involve the coordination and decision-making of PEMC to achieve effective offloading and processing of the task.

[0239] Step S32: PEMC generates an offloading plan.

[0240] After receiving the application from PINE, PEMC comprehensively judges the offloading target, including other PINEs adjacent to this PINE, the PEGC to which the PINE is connected, other PEGCs that the PINE can access, the cloud, etc. PEMC follows the following steps according to factors such as the computing resources, network status, and supported AI models of the offloading target:

[0241] Offloading judgment process: Collect information of the offloading target: PEMC first needs to collect relevant information of the offloading target, including the computing resource status, network status, and supported AI models of each PINE. This information can be obtained through the PINE application in step S31 or by monitoring and recording the status of PINE.

[0242] Evaluate the computing resources of the offloading target: PEMC evaluates whether the offloading target has sufficient computing resources to undertake the offloading tasks of PINE based on the computing resource status of the offloading target, such as computing power, memory, storage, etc. This requires establishing a communication connection with the offloading target to obtain the resource status and make comparisons.

[0243] Evaluate the network status of the offloading target: PEMC evaluates the network status of the offloading target, including metrics such as bandwidth, latency, packet loss rate, etc. These metrics can be obtained by establishing a communication connection with the offloading target and conducting network measurements. PEMC needs to ensure that the network quality of the offloading target can meet the computing performance requirements of PINE.

[0244] Match the offloading target and the computing performance requirements: PEMC matches the computing resources and network status of the offloading target with the computing performance requirements of PINE. This involves comparing whether the computing resources of the offloading target are powerful enough to meet the requirements of PINE and ensuring that the network status can provide sufficient bandwidth and low latency.

[0245] Consider the AI model support of the offloading target: PEMC also needs to consider whether the offloading target supports the AI models required by PINE. This includes verifying whether the offloading target has the required AI model version and computing requirements to ensure the successful offloading of PINE's AI computing tasks.

[0246] Select a suitable device or service for offloading: Based on the results of the comprehensive evaluation, PEMC selects a suitable device or service as the offloading target. This may involve selecting other nearby PINEs, PEGCs accessed by PINE, other PEGCs accessible to PINE, the cloud, etc.

[0247] Select a suitable device or service for offloading AI computing tasks to meet the computing performance requirements of PINE.

[0248] The feasibility is described as follows:

[0249] Determine specific evaluation metrics and thresholds: For computing resources and network status, PEMC can set specific evaluation metrics and thresholds for more precise judgment. For example, minimum computing power and maximum latency requirements can be set.

[0250] Consider the actual deployment situation: Based on the specific deployment environment and available resources, PEMC needs to consider the actually available offloading targets. This can include factors such as the locations of deployed PINEs and PEGCs, the network topology, and the availability of cloud resources.

[0251] Adopt adaptive algorithms and mechanisms: To cope with dynamic network and computing environments, PEMC can adopt adaptive algorithms and mechanisms to monitor and adjust the offloading target selection strategy in real time. This can help adapt to network load, resource changes, and fault situations.

[0252] Exemplarily, assume that in an edge computing network, there are multiple PINEs and several PEGCs. Due to insufficient power, PINE A applies to PEMC for AI computing offloading, carrying the number of split layers and computing performance requirements. PINE A hopes to offload the last two layers of the AI model, and the computing performance requirements require the offloading target to have at least 4GB of memory, support TensorFlow (a deep neural network model) version 2.0, and the network latency to be less than 20 milliseconds.

[0253] In the process of generating the offloading scheme in step S32, after receiving the application from PINE A, PEMC starts to comprehensively judge the offloading target. PEMC collects the computing resource status, network status, and supported AI model information of all PINEs and PEGCs. Then, PEMC evaluates the computing resources and network status of each offloading target, such as establishing a communication connection with the target and obtaining the resource status. At the same time, PEMC considers whether the offloading target supports TensorFlow version 2.0, and whether it meets the requirements of at least 4GB of memory and low latency.

[0254] In the result of the comprehensive evaluation, PEMC finds that PINE B near PINE A has sufficient computing resources and a low-latency network, and at the same time supports TensorFlow version 2.0. Therefore, PEMC selects PINE B as the offloading target. PEMC establishes a communication connection with PINE B, and through the "AI model transfer method", offloads the last two layers of AI computing tasks of PINE A to PINE B.

[0255] Step S33: PEMC issues a computing offloading task to PINE B.

[0256] Here, PINE B is the offloading target to which the computing task of PINE A selected by PEMC needs to be offloaded.

[0257] PEMC can offload the AI computing tasks of PINE to appropriate devices or services through the AI model transfer method according to the offloading target.

[0258] The AI model transfer method is a technology that transfers some or all layers of an AI model to other devices or services, and can achieve the splitting and migration of the AI model. PEMC establishes a communication connection with the offloading target, transmits data and results. After this process ends, it returns to step S31.

[0259] During the communication process of the AI model transfer method, the following three fields need to be transferred: the AI model layer, data, and results.

[0260] The AI model layer may involve the number of split layers or the number of restored layers.

[0261] The number of split layers is a parameter indicating which layers of the AI model are to be transferred. The number of split layers can be set according to different tasks and scenarios, such as model complexity, data scale, computing resources, etc. Generally speaking, the more the number of split layers, the smaller the amount of data transferred, but the computing efficiency and quality may decrease.

[0262] The number of restored layers is a parameter indicating which layers of the AI model are to be restored. The number of restored layers can be set according to different tasks and scenarios, such as model complexity, data scale, computing resources, etc. Generally speaking, the more the number of restored layers, the smaller the amount of data restored, but the computing efficiency and quality may increase.

[0263] Data is a field referring to the input or output data used for training or testing the AI model. The data contains information such as the type, format, and size of the data. Data is the auxiliary content for transfer and restoration, and it determines the input and output of the AI computing task. During the communication process, PEMC and the offloading target need to transmit data information to each other to implement the execution and verification of the AI computing task.

[0264] Results is a field referring to the output results of the AI computing task. The results contain information such as the type, format, and size of the results. Results are the feedback content for transfer and restoration, and it determines the effect and quality of the AI computing task. During the communication process, PEMC and the offloading target need to transmit result information to each other to implement the evaluation and optimization of the AI computing task.

[0265] Figure 11 This is the interactive schematic diagram of PINE restoring autonomous computing provided by the embodiments of the present disclosure. PINE can restore autonomous computing and perform the restoration and migration of AI computing tasks.

[0266] Such as Figure 11 shown, the implementation process of PINE restoring autonomous computing may include:

[0267] Step S41: PINE A applies to restore autonomous computing.

[0268] PINE has the ability of autonomous computing. For example, after the power or resources are sufficient, it can apply to PEMC to restore autonomous computing, carrying information such as the number of restored layers. The number of restored layers indicates which layers of the AI model PINE hopes to restore, that is, migrate back from the offloading target to PINE.

[0269] Step S42: PEMC stops the computing offloading task.

[0270] After the PEMC receives the application from the PINE, it restores the AI computing tasks previously offloaded by the PINE to the PINE through the AI model transfer method.

[0271] Step S43: The PINE B transfers part of the model parameters and calculation results to the PINE A through the PEMC.

[0272] Here, the PINE B is the offloading target to which the computing task of the PINE A selected by the PEMC is offloaded.

[0273] After the PINE receives the AI model layer information transferred by the PEMC, it integrates it with the local AI model layer and resumes self-computation. The PINE disconnects the communication connection with the offloading target and releases resources. After this process ends, it returns to step S41.

[0274] In summary, the technical solutions provided by the embodiments of the present disclosure have at least the following technical advantages:

[0275] 1) Through the offloading operation, the AI computing efficiency and quality in the PIN network can be improved, and the AI computing cost and risk in the PIN network can be reduced;

[0276] 2) Through the AI model transfer method, the splitting and migration of the AI model can be realized to adapt to different offloading targets and scenarios.

[0277] 3) By dynamically adjusting parameters such as the offloading target, offloading algorithm, offloading threshold, offloading frequency, and offloading priority, the adaptive management of AI computing offloading can be realized.

[0278] In the embodiments of the present disclosure, some or all of the steps and their optional implementation manners can be arbitrarily combined with some or all of the steps in other embodiments, and can also be arbitrarily combined with the optional implementation manners in other embodiments.

[0279] Figure 12 The structure diagram of a computing offloading device provided by the embodiments of the present disclosure is shown. The computing offloading device is applied to the PEMC, as Figure 12 shown, the computing offloading device 100 includes:

[0280] A receiving module 110, configured to receive an offloading request from the first personal Internet of Things device PINE in the personal Internet of Things PIN; wherein, the offloading request is used to request the offloading of the computing task of the first PINE;

[0281] A determining module 120, configured to determine, in the PIN, an offloading target for executing the computing task according to the management policy, the offloading request, and the current status of each PINE in the PIN when the management policy obtained in advance indicates to enable PIN computing offloading;

[0282] An offloading module 130, configured to offload the computing tasks of the first PINE to the offloading target.

[0283] In some embodiments, the receiving module 110 is further configured to:

[0284] Receive the management policy from a core network element or an application function network element.

[0285] In some embodiments, the apparatus further includes:

[0286] A sending module, configured to send an offloading status report for adjusting the management policy to the core network element or the application function network element; wherein, the offloading status report includes at least one of offloading efficiency, offloading quality, and offloading cost of the offloading operation for the computing task.

[0287] In some embodiments, the current states of the PINEs in the PIN include:

[0288] The currently reported computing resource states of multiple PINEs in the PIN, and the multiple PINEs include the first PINE and at least one second PINE;

[0289] The currently reported network states of the respective PINEs reported by the personal Internet of Things gateway device PEGC;

[0290] The artificial intelligence AI models supported by the respective PINEs in the PIN.

[0291] In some embodiments, the offloading request carries the number of split layers and the computing performance requirements of the target artificial intelligence AI model of the first PINE;

[0292] The determining module 120 is configured to:

[0293] For each of the second PINEs, according to the management policy, match the currently reported computing resource state and the currently reported network state of the second PINE with the computing performance requirements of the first PINE, and determine whether there is a candidate second PINE that meets the computing performance requirements;

[0294] When there is at least one candidate second PINE, match the number of split layers of the target AI model with the AI models supported by the respective candidate second PINEs, and determine the target second PINE with a successful match;

[0295] Determine the target second PINE with a successful match as the offloading target.

[0296] In some embodiments, the offloading module 130 is configured to:

[0297] Transfer the computing tasks of some or all layers of the AI model of the first PINE to the offloading target.

[0298] In some embodiments, the receiving module 110 is further configured to:

[0299] Receive a recovery request sent by the first PINE; wherein, the recovery request is used to request the recovery of the computing tasks of the first PINE.

[0300] The device further includes:

[0301] A recovery module, configured to, in response to the recovery request, restore the computing tasks unloaded from the first PINE to the first PINE.

[0302] In some embodiments, the recovery request carries the number of layers to be recovered of the AI model of the first PINE; the recovery module is configured to:

[0303] In response to the recovery request, migrate the computing tasks of the number of layers to be recovered of the AI model requested to be recovered by the first PINE from the offloading target back to the first PINE.

[0304] It should be noted that: when the computing offloading device provided in the above embodiment is used to implement the computing offloading method that should be executed by the PEMC, only the above division of each program module is used for illustration. In actual applications, the above processing can be allocated to different program modules according to needs, that is, the internal structure of the computing offloading device is divided into different program modules to complete all or part of the above-described processing. In addition, the computing offloading device provided in the above embodiment and the corresponding method embodiment belong to the same concept. For the specific implementation process, please refer to the method embodiment, which will not be elaborated here.

[0305] Figure 13 Shows the structural diagram of a computing offloading device provided by an embodiment of the present disclosure. The computing offloading device is applied to the first personal Internet of Things device (PINE), such as Figure 13 As shown, the computing offloading device 200 includes:

[0306] A sending module 210, configured to send an offloading request to the PEMC; wherein, the offloading request is used to request the offloading of the computing tasks of the first PINE.

[0307] A receiving module 220, configured to receive a first response sent by the PEMC in response to the offloading request; wherein, the first response is used to indicate whether to offload the computing tasks requested to be offloaded by the first PINE.

[0308] In some embodiments, the sending module 210 is further configured to:

[0309] send a recovery request to the PEMC; wherein, the recovery request is used to request the recovery of the computing tasks unloaded by the first PINE;

[0310] The receiving module 220 is further configured to:

[0311] receive a second response sent by the PEMC in response to the recovery request; wherein, the second response is used to indicate that the computing tasks requested to be recovered by the first PINE are recovered to the first PINE.

[0312] In some embodiments, the computing tasks of the first PINE include: the computing tasks of some or all layers of the AI model of the first PINE.

[0313] It should be noted that: when the above-described computing offloading device applied to the first PINE implements the computing offloading method that should be executed by the first PINE, only the above-mentioned division of each program module is used for illustration. In practical applications, the above-mentioned processing can be allocated to different program modules according to needs, that is, the internal structure of the computing offloading device is divided into different program modules to complete all or part of the above-described processing. In addition, the above-described computing offloading device and the corresponding method embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0314] Figure 14 is a structural block diagram of a computer device provided by an embodiment of the present disclosure; as Figure 14 shown, the computer device 1400 includes: a processor 1401 and a memory 1402 for storing a computer program that can run on the processor; wherein, the processor 1401 is used to implement the steps in the above-described computing offloading method when running the computer program.

[0315] In practical applications, the computer device 1400 may further include: at least one network interface 1403. Each component in the computer device 1400 is coupled together through a bus system 1404. It can be understood that the bus system 1404 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 1404 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in Figure 14 all kinds of buses are labeled as the bus system 1404. Among them, the number of processors 1401 can be at least one. The network interface 1403 is used for wired or wireless communication between the computer device 1400 and other devices.

[0316] The memory 1402 in the embodiments of the present disclosure is used to store various types of data to support the operation of the computer device 1400.

[0317] The methods disclosed in the embodiments of the present disclosure above can be applied to the processor 1401 or implemented by the processor 1401. The processor 1401 may be an integrated circuit chip with signal processing capabilities. In the implementation process, the steps of the above methods can be completed by the integrated logic circuit in the hardware of the processor 1401 or instructions in the form of software. The above-mentioned processor 1401 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 1401 can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present disclosure. The general-purpose processor may be a microprocessor or any conventional processor, etc. Combining the steps of the methods disclosed in the embodiments of the present disclosure, it can be directly embodied as being executed and completed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, and this storage medium is located in the memory 1402. The processor 1401 reads the information in the memory 1402 and combines its hardware to complete the steps of the foregoing computing offloading method.

[0318] In an exemplary embodiment, the computer device 1400 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontroller units (MCUs), microprocessors, or other electronic components, and is used to execute the foregoing computing offloading method.

[0319] The embodiments of the present disclosure also provide a computer-readable storage medium, on which a computer program is stored; when the computer program is executed by a processor, the steps of the foregoing computing offloading method are implemented.

[0320] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the couplings, direct couplings, or communication connections between the components shown or discussed can be through some interfaces. The indirect couplings or communication connections of devices or units can be electrical, mechanical, or other forms.

[0321] The units described above as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0322] In addition, each functional unit in the embodiments of the present disclosure can be all integrated in a processing unit, or each unit can be separately used as a unit, or two or more units can be integrated in a unit; the above-mentioned integrated units can be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0323] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other various media that can store program codes.

[0324] Alternatively, if the above-mentioned integrated units of the present disclosure are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present disclosure essentially or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present disclosure. And the foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks or optical disks and other various media that can store program codes.

[0325] As described above, it is only the specific implementation manner of the present disclosure. However, the protection scope of the present disclosure is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present disclosure can easily think of changes or substitutions, which should all be covered by the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be subject to the protection scope of the claimed rights.

Claims

1. A computing offloading method, characterized in that, Executed by a personal Internet of Things management device PEMC; the method includes: Receiving an uninstallation request from the first personal Internet of Things device PINE in the personal Internet of Things PIN; wherein, the uninstallation request is used to request the uninstallation of the computing task of the first PINE. When the pre-acquired management policy indicates to enable PIN computing offloading, according to the management policy, the uninstallation request, and the current status of each PINE in the PIN, determining an uninstallation target for executing the computing task in the PIN. Offloading the computing task of the first PINE to the uninstallation target.

2. The method according to claim 1, wherein The method further includes: Receiving the management policy from a core network element or an application function network element.

3. The method according to claim 2, characterized in that, The method further includes: Sending an uninstallation status report for adjusting the management policy to the core network element or the application function network element; wherein, the uninstallation status report includes at least one of the uninstallation efficiency, uninstallation quality, and uninstallation cost of the uninstallation operation for the computing task.

4. The method according to claim 1, characterized in that The current status of each PINE in the PIN includes: The currently reported computing resource status of each of the multiple PINEs in the PIN, and the multiple PINEs include the first PINE and at least one second PINE. The current network status of each PINE reported by the personal Internet of Things gateway device PEGC. The artificial intelligence AI models supported by each PINE in the PIN.

5. The method according to claim 4, wherein The uninstallation request carries the number of split layers and the computing performance requirements of the target artificial intelligence AI model of the first PINE. When the pre-acquired management policy indicates to enable PIN computing offloading, according to the management policy, the uninstallation request, and the current status of each PINE in the PIN, determining an uninstallation target for executing the computing task in the PIN includes: For each of the second PINEs, according to the management policy, matching the current computing resource status and current network status of the second PINE with the computing performance requirements of the first PINE to determine whether there is a candidate second PINE that meets the computing performance requirements. When there is at least one candidate second PINE, matching the number of split layers of the target AI model with the AI models supported by each candidate second PINE to determine a target second PINE with a successful match. Determining the target second PINE with a successful match as the uninstallation target.

6. The method according to claim 1, characterized in that, The offloading the computing task of the first PINE to the uninstallation target includes: Transferring the computing tasks of some or all layers of the AI model of the first PINE to the uninstallation target.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Receiving a recovery request sent by the first PINE; wherein, the recovery request is used to request the recovery of the computing task of the first PINE. In response to the recovery request, restoring the uninstalled computing task of the first PINE to the first PINE.

8. The method according to claim 7, wherein The recovery request carries the number of layers to be recovered of the AI model of the first PINE; in response to the recovery request, restoring the computing tasks unloaded from the first PINE to the first PINE includes: In response to the recovery request, migrating the computing tasks of the number of layers to be recovered of the AI model requested to be recovered by the first PINE back from the offloading target to the first PINE.

9. A computing offloading method, characterized in that, Executed by the first Personal Internet of Things Device (PINE); the method includes: Sending an offloading request to the Personal Internet of Things Management Device (PEMC); wherein, the offloading request is used to request offloading of the computing tasks of the first PINE. Receiving a first response sent by the PEMC in response to the offloading request; wherein, the first response is used to indicate whether to offload the computing tasks requested to be offloaded by the first PINE.

10. The method according to claim 9, wherein The method further includes: Sending a recovery request to the PEMC; wherein, the recovery request is used to request recovery of the computing tasks unloaded from the first PINE. Receiving a second response sent by the PEMC in response to the recovery request; wherein, the second response is used to indicate restoring the computing tasks requested to be recovered by the first PINE to the first PINE.

11. The method according to claim 9 or 10, characterized in that, The computing tasks of the first PINE include: the computing tasks of some or all layers of the AI model of the first PINE.

12. A computing offloading device, characterized in that, Applied to the Personal Internet of Things Management Device (PEMC); the apparatus includes: A receiving module, configured to receive an offloading request from the first Personal Internet of Things Device (PINE) in the Personal Internet of Things (PIN); wherein, the offloading request is used to request offloading of the computing tasks of the first PINE. A determining module, configured to, when a pre-acquired management policy indicates to enable PIN computing offloading, determine, according to the management policy, the offloading request, and the current states of each PINE in the PIN, an offloading target in the PIN for executing the computing tasks. An offloading module, configured to offload the computing tasks of the first PINE to the offloading target.

13. A computing offloading device, characterized in that, Applied to the first Personal Internet of Things Device (PINE); the apparatus includes: A sending module, configured to send an offloading request to the Personal Internet of Things Management Device (PEMC); wherein, the offloading request is used to request offloading of the computing tasks of the first PINE. A receiving module, configured to receive a first response sent by the PEMC in response to the offloading request; wherein, the first response is used to indicate whether to offload the computing tasks requested to be offloaded by the first PINE.

14. A computer device, characterized in that, Includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the steps of the computing offloading method according to any one of claims 1 to 8 or claims 9 to 11 are implemented.

15. A computer-readable storage medium, characterized in that, On which a computer program is stored, and when the computer program is executed by a processor, the steps of the computing offloading method according to any one of claims 1 to 8 or claims 9 to 11 are implemented.

Images