Dynamic unloading method and system for calculation task, electronic equipment and program product

Abstract

The invention provides a dynamic unloading method and system for a computing task, electronic equipment and a computer program product, and relates to the technical field of cloud network convergence, and the method comprises the steps that an edge side receives reported model task characteristics and equipment resource states of the computing task, and receives issued core side data; the core side data comprises network data and cost data; the edge side performs joint optimization on the model task characteristics, the equipment resource state, the network state and the cost data based on a dynamic unloading decision algorithm to generate a user equipment routing strategy; and a user equipment routing strategy is applied to the core side, so that the core side issues the user equipment routing strategy to the terminal side, and hierarchical unloading of the calculation task is executed according to the user equipment routing strategy. According to the invention, hierarchical unloading can be realized, and delay is reduced.

Description

Technical Field

[0001] This disclosure relates to the field of cloud-network convergence technology, and more specifically, to a method for dynamically offloading computing tasks, a system for dynamically offloading computing tasks, an electronic device, and a computer program product. Background Technology

[0002] With the development of 5G edge computing, terminal devices such as smart cameras and AR glasses often need to offload artificial intelligence inference tasks to the edge cloud via 5G networks to reduce latency and energy consumption.

[0003] In related technologies, network slices and paths are only statically allocated without considering the real-time computing needs of different layered tasks in the inference model. This can lead to high-precision models failing to meet latency requirements due to network congestion. Although user equipment routing strategies in related technologies can offload traffic, they are not dynamically linked to model offloading needs. For example, some computing tasks may still occupy high-priority slice resources even after being downgraded to local execution. Summary of the Invention

[0004] The purpose of this disclosure is to provide a method, system, electronic device, and computer program product for dynamically unloading computing tasks, thereby overcoming, at least to some extent, the problem of high latency caused by limitations and defects in related technologies.

[0005] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0006] According to one aspect of this disclosure, a method for dynamically unloading computing tasks is provided, comprising: The edge side receives the reported model task characteristics and device resource status of the computing task, and also receives the core side data sent down; the core side data includes network data and cost data. On the edge side, a dynamic offloading decision algorithm is used to jointly optimize the model task characteristics, device resource status, network status, and cost data to generate a user equipment routing strategy. The system requests a user equipment routing policy from the core side, so that the core side can distribute the user equipment routing policy to the terminal side and perform hierarchical offloading of computing tasks according to the user equipment routing policy.

[0007] In one exemplary embodiment of this disclosure, the edge side performs joint optimization on the model task characteristics, device resource status, network status, and cost data based on a dynamic offloading decision algorithm to generate a user equipment routing strategy, including: The model task features, device resource status, and network status are taken as the current state, and the action probability distribution output by the policy network is determined based on the original probability of executing the target action in the current state. Determine the action value function based on the target action and the current state; Based on the action value function, the comprehensive cost of executing the target action determined according to the cost data, and the temperature coefficient, cost parameters are determined; the target action is selecting a target slice or a target unloading path, and the generation method of the target action is determined according to the temperature coefficient. The user equipment routing strategy is determined based on the action probability distribution and the cost parameters.

[0008] In one exemplary embodiment of this disclosure, the user equipment routing policy includes at least one of slice identifier, offload path, multi-level offload ratio, quality of service, and quality of service identifier; the step of requesting a user equipment routing policy from the core side includes: The slice identifier and quality of service identifier in the offloading decision result are encapsulated into a core network signaling format to request a user equipment routing policy from the core side.

[0009] In one exemplary embodiment of this disclosure, the hierarchical offloading of computing tasks according to the user equipment routing policy includes: The offloading ratio of the computing task on the target side is determined based on the multi-level offloading ratio in the user equipment routing selection strategy. According to the stated unloading ratio and unloading path, the computing task is unloaded to the target side to execute the unloading of the computing task.

[0010] In an exemplary embodiment of this disclosure, the target side is the core side; the step of unloading the computing task to the target side according to the unloading ratio and unloading path includes: According to the aforementioned unloading ratio and unloading path, the computing task is unloaded to the slice corresponding to the slice identifier in the core side, and the slice quota is adjusted for the slice to execute the unloading of the computing task.

[0011] In one exemplary embodiment of this disclosure, adjusting the slice quota for the slice corresponding to the slice identifier in the target side includes: User equipment routing policy templates are generated based on standard protocols, and the slice quotas are adjusted according to core network analysis function data.

[0012] In one exemplary embodiment of this disclosure, the target side is the edge side, and the step of unloading the computing task to the target side according to the unloading ratio and the unloading path includes: If the target edge node among the multiple edge nodes included on the edge side meets the unloading requirements, the computing task is unloaded to the target edge node according to the unloading ratio of the edge side and the unloading path.

[0013] According to one aspect of this disclosure, a dynamic offloading system for computing tasks is provided, comprising: On the terminal side, it is used to collect model task characteristics and device resource status of computing tasks and report them to the edge side; On the edge side, the system receives the model task features, device resource status, network status, and cost data. Based on the dynamic offloading decision algorithm, it performs joint optimization on the model task features, device resource status, network status, and cost data to generate a user equipment routing selection strategy. On the core side, the system is used to distribute the user equipment routing policy to the terminal side and control the terminal side to perform hierarchical offloading of computing tasks according to the user equipment routing policy.

[0014] According to one aspect of this disclosure, an electronic device is provided, comprising: processor; and memory for storing the executable instructions of the processor; The processor is configured to execute the dynamic offloading method for any of the above-described computational tasks by executing the executable instructions.

[0015] According to one aspect of this disclosure, a computer program product is provided, which, when executed by a processor, implements the dynamic unloading method for the computing task described in any one of the preceding claims.

[0016] The technical solution provided in this disclosure, on the one hand, by fusing data at multiple levels including the terminal side, edge side, and core side, considers the needs of computing tasks and the real-time computing needs of different hierarchical tasks when determining the user equipment routing strategy. This improves the generation efficiency and comprehensiveness of the user equipment routing strategy, enabling the user equipment routing strategy to respond to changes in computing tasks and network status, improving real-time performance, and meeting latency requirements. On the other hand, by optimizing the offloading ratio of the terminal side, edge side, and core side in real time, hierarchical offloading is achieved, improving the utilization rate of computing resources. Through reasonable hierarchical offloading, the execution latency of computing tasks is reduced, and energy consumption is effectively reduced.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0019] Figure 1 The schematic diagram illustrates a flowchart of a dynamic unloading method for computing tasks according to an embodiment of the present disclosure.

[0020] Figure 2 The illustration shows a flowchart of the process for generating a user equipment routing policy in an embodiment of this disclosure.

[0021] Figure 3 The diagram illustrates the dynamic unloading interaction in an embodiment of this disclosure.

[0022] Figure 4 This diagram schematically illustrates a block diagram of a dynamic unloading system for computing tasks according to an embodiment of the present disclosure.

[0023] Figure 5 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0024] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0025] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0026] Traditional solutions statically allocate network slices and paths without considering the real-time computational needs of different layers of inference model tasks (such as model complexity and input data volume). This can lead to high-precision models failing to meet latency requirements due to network congestion. While existing URSP technology can offload traffic, it does not dynamically link with model offloading needs. For example, some computational tasks may still occupy high-priority slice resources even after being degraded to local execution. Existing ATSSS multipath transmission cannot optimize bandwidth allocation based on the input / output data characteristics of inference tasks; for example, video streams and feature maps are not routed differently.

[0027] To address the technical problems existing in related technologies, this disclosure provides a method for dynamically unloading computing tasks, which can be applied to scenarios such as smart factories, AGV visual navigation tasks, autonomous driving, and AR games that communicate via 5G networks and can perform multi-level task collaboration.

[0028] Next, the dynamic unloading method for computing tasks in the embodiments of this disclosure will be described in detail with reference to the accompanying drawings.

[0029] In step S110, the edge side receives the reported model task characteristics and device resource status of the computing task, and receives the core side data sent down; the core side data includes network data and cost data.

[0030] In this embodiment, a dynamic offloading method for computing tasks can be implemented at the terminal side, edge side, and core side. The computing task can be any type of inference task, such as an artificial intelligence inference task. The terminal can be a smart camera, AR glasses, or other terminal device. The terminal side can include a lightweight monitor. The lightweight monitor collects the model task characteristics and device resource status of the computing task, executes a local lightweight model computing task, and responds to the URSP (UE Route Selection Policy) offloading command issued by the edge side, for example, offloading 80% to edge node A. The model task characteristics include at least one of model complexity, input data volume, latency sensitivity, and privacy level. The device resource status can include device status and computing power status. The device status includes at least one of local computing power, memory, and remaining battery power. Local computing power can be, for example, remaining CPU / GPU computing power. The computing power status can include edge node CPU / GPU utilization and cloud CPU / GPU utilization.

[0031] Based on this, the lightweight monitor on the terminal side monitors the local status in real time, including the characteristics of the model task to be executed, device status, and wireless quality, and periodically reports the above local status data to the dynamic offloading controller on the edge side.

[0032] The edge side includes a dynamic offload controller. This controller receives local status data reported by the terminal side and core-side data reported by the core side, runs the URSP dynamic offload decision algorithm, generates the optimal URSP policy, and sends the generated URSP policy to the core network PCF. Local status data includes model task characteristics and device resource status. This may include local privacy processing capabilities and energy consumption data. The edge-side dynamic offload controller combines network data and cost data issued by the core network PCF to generate the URSP policy in real time. Network data can be network environment data, specifically used to represent the real-time status of network slices, such as slice throughput, latency, packet loss rate, URLLC / eMBB availability, RSRP (Reference Signal Received Power), and load, among one or more of these. The edge-side dynamic offload controller requests the real-time status of the network slice from the core side, and then uses the requested real-time status of the network slice as the network status. Cost data may include slice billing costs and signaling overhead costs.

[0033] In step S120, the edge side performs joint optimization on the model task characteristics, device resource status, network data, and cost data based on the dynamic offloading decision algorithm to generate a user equipment routing strategy.

[0034] In this embodiment, the edge-side dynamic offloading controller combines the real-time network slice status (URLLC / eMBB slice availability, RSRP, load) issued by the core network PCF to calculate the action value and overall cost, and generates a real-time URSP policy through the policy network. Specifically, the dynamic offloading controller receives terminal-side data (including local privacy processing capabilities and energy consumption data) and core-side data (including slice billing costs, signaling overhead costs, etc.), runs the URSP dynamic offloading decision algorithm in this embodiment, generates the optimal URSP policy, and generates a URSP policy request to be sent to the core side, which refers to the core network PCF. The URSP dynamic offloading decision algorithm is based on reinforcement learning to model URSP policy generation as a Markov decision process.

[0035] For example, the model task features reported by the terminal side, the device resource status, and the state space data represented by the core side data sent by the core side can be used as inputs. At the same time, cost data can also be used as inputs. These are fed into the dynamic offloading decision algorithm to obtain the user equipment routing selection strategy as the output. The user equipment routing selection strategy may include one or more of the following: slice identifier, QoS (Quality of Service), 5QI (5G QoS Identifier), offloading path, and multi-level offloading ratio.

[0036] The slice identifier can be used to indicate slice selection, referring to the slice ID of the 5G network, such as URLLC_Slice_ID or eMBB_Slice_ID. The offloading path refers to path control, specifically the execution location of the model computation task, including local execution (LocalOnly), edge offloading, and cloud offloading. The tiered offloading ratio can be the multi-path traffic distribution ratio, specifically the ratio of model task distribution using slices combined with ATSSS technology, such as 5G:Wi-Fi = 8:2.

[0037] User equipment routing strategies can be used to represent action space data, and can also be represented as offloading decision results. In this embodiment, offloading decision results are dynamically generated through joint optimization of state space data and action space data.

[0038] Figure 2 The flowchart illustrating the determination of the user equipment routing strategy is shown, and includes the following steps: Step S210: Take the model task features, device resource status, and network status as the current status, and determine the action probability distribution output by the policy network based on the original probability of selecting the target action in the current status; the target action is to select the target slice or the target unloading path. Step S220: Determine the action value function based on the target action and the current state; Step S230: Based on the action value function, determine the comprehensive cost of performing the target action and the temperature coefficient according to the cost data, and determine the cost parameters; Step S240: Determine the user equipment routing strategy based on the action value function and the cost parameters.

[0039] In this embodiment, the current state is first determined, which may include task characteristics, network / slice status, and computing power status. The current state can be reported in real-time by a lightweight monitor. The action refers to selecting a slice or selecting an unloading path. After determining the current state and the target action, the policy network can output an action probability distribution. This distribution represents the raw probability of selecting the target action in the current state; for example, the probability of selecting a URLLC slice is 0.7, and the probability of selecting an eMBB slice is 0.3. The policy network can be represented as follows: Policy Network Its purpose is to provide prior probabilities of the target action, avoiding getting trapped in local optima by relying solely on Q-Cost. In practical applications, this can be achieved through the PPO reinforcement learning algorithm, which pre-trains the policy network. This allows the system to initially perceive the matching relationship between the task and the network, and then dynamically adjust it based on real-time environmental data during use. Based on the action probability distribution, the action that maximizes the action probability distribution can be identified. This action can then be defined as the target action.

[0040] Once the target action is determined, an action value function can be derived based on the target action and the current state. The action value function quantifies the long-term benefits of selecting the target action in the current state, thereby achieving a deep match between business needs and network resources. The action value function includes metrics such as latency, congestion, and privacy protection levels.

[0041] The overall cost of executing the target action is used to balance system benefits with resource consumption. Its goal is to curb high-return but high-cost strategies, such as frequent slice switching that could lead to signaling storms or a surge in costs.

[0042] The overall cost of performing the target action is determined based on cost data provided by the core side. For example, it can be determined by weighted summation of the slice billing price, energy consumption reported by the terminal side, and privacy leakage risks. The slice billing price provided by the core side is higher for URLLC slices. Privacy leakage risks are lower, for example, due to lower scores for cloud-based execution.

[0043] Formula (2)

[0044] in, This refers to the unit price for slice billing provided by the core network PCF (e.g., URLLC slices are more expensive). Energy consumption reported by the terminal (e.g., offloading to the edge can save terminal power). This is used to indicate the risk of privacy breaches (e.g., lower scores when performing in the cloud). , , These are weighting coefficients, which can be preset according to specific business needs or obtained through training. The distributed computing on the terminal side, edge side, and core side in this embodiment of the disclosure achieves this through the collaboration between the various components. It enables cost perception across layers.

[0045] The temperature coefficient is used to control the exploration intensity. The target action is either selecting a target slice or a target unloading path. The temperature coefficient determines whether the target action is generated as the optimal action, a suboptimal action, or a randomly selected action. The temperature coefficient can be dynamically adjusted according to specific needs. (Low temperature) The approach tends to be more acute, with the method of determining the target action becoming more deterministic, choosing the current best action and ignoring suboptimal options. Conversely, the method of determining the target action is almost random, involving a wide range of potential options.

[0046] In some embodiments, cost parameters can be determined based on the action value function, the overall cost of performing the target action, and the temperature coefficient. Specifically, the cost parameters can be determined based on the difference between the action value function and the overall cost of performing the target action, and by calculating the ratio of the difference to the temperature coefficient.

[0047] Based on this, the URSP strategy can be determined according to the action probability distribution and cost parameters. The dynamic unloading decision algorithm can be expressed as formula (2): Formula (2) in, It is the generated URSP strategy, for example, the optimal URSP strategy, which includes 5G slice ID and tiered offload ratio. It indicates the current status, including task characteristics, network / slice status, and computing power status, and is reported in real time by the lightweight monitor. It is the action probability distribution output by the policy network. This represents the action-value function, used to quantify the value of an action in the current state. Select target action This will lead to long-term benefits, thereby achieving a deep match between business needs and network resources, including indicators such as latency, congestion, and privacy protection levels. Indicates the execution of an action The overall cost is used to balance system benefits and resource consumption. The goal is to curb high-return but high-cost strategies, such as frequent slice switching which may lead to signaling storms or a surge in costs. This refers to the temperature coefficient.

[0048] In this embodiment, reinforcement learning is used to model URSP policy generation as a Markov decision process. The optimal URSP offloading policy is dynamically generated through joint optimization of the state space (model task characteristics, network load, computing resources, cost assessment, etc.) and the action space (slice selection, offloading ratio, multi-path routing). Furthermore, a temperature coefficient is introduced to adaptively adjust the exploration intensity, automatically adjusting the policy conservatism under high load or high reliability scenarios.

[0049] In step S130, a user equipment routing policy is requested from the core side so that the core side can distribute the user equipment routing policy to the terminal side and perform hierarchical offloading of computing tasks according to the user equipment routing policy.

[0050] In this embodiment, after the edge side generates a user equipment routing policy, it can request a user equipment routing policy from the core network PCF. Specifically, the dynamic offload controller on the edge side performs URSP policy conversion and requests a URSP policy from the core network PCF. For example, the slice identifier and quality of service identifier in the user equipment routing policy are encapsulated into a core network signaling format to request a user equipment routing policy from the core side. The core network itself has a core network signaling format. After determining the user equipment routing policy, the original values ​​in the core network signaling format can be replaced according to the slice identifier and the quality of service identifier 5QI in the network, so that the user equipment routing policy is encapsulated into a core network signaling format based on the slice identifier and the quality of service identifier 5QI in the network. Furthermore, a user equipment routing policy can be requested from the core network PCF based on the core network signaling format.

[0051] After receiving the user equipment routing policy, the core side can review it. Upon approval, it sends a confirmation message to the edge side and distributes the routing policy to the terminal side for execution. For example, the core side can review the usage permissions and rationality of the user equipment routing policy. Slicing refers to dividing network resources into dedicated parts; resources in different slices cannot be shared. It can be determined whether the client has usage permissions for the slice corresponding to the slice identifier in the user equipment routing policy. If the client has usage permissions, it can be further determined whether the resources required by the terminal are available. If the resources are determined to be available, the user equipment routing policy can be considered approved.

[0052] After approval, a confirmation message can be sent to the edge side, and the user equipment routing policy can be distributed to the terminal side. This allows the terminal side to perform tiered offloading of computing tasks according to the user equipment routing policy. Tiered offloading includes terminal-side offloading, edge-side offloading, and core-side offloading. Offloading here can be understood as deploying computing tasks for execution on the terminal side, edge side, and core side.

[0053] The hierarchical offloading of computing tasks according to the user equipment routing policy includes: determining the offloading ratio of computing tasks on the target side based on computing resources and the multi-level offloading ratio in the user equipment routing policy; and offloading the computing tasks to the target side according to the offloading ratio and the offloading path. The offloading ratio refers to the proportion of computing tasks executed through the terminal side, edge side, and core side. Since the user equipment routing policy includes multi-level offloading ratios, slice identifiers, and offloading paths, based on this, when the computing resources on the target side meet the resource conditions, the offloading ratio of computing tasks on the target side can be determined according to the multi-level offloading ratio in the user equipment routing policy; and the computing tasks can be offloaded to the target side according to the offloading ratio and the offloading path. The multi-level offloading ratio may include a first offloading ratio on the terminal side, a second offloading ratio on the edge side, and a third offloading ratio on the core side. Therefore, the target side may include at least one of the terminal side, edge side, and core side. After determining the offloading ratio, the computing tasks can be offloaded to the target side according to the offloading path in the user equipment routing policy, for example, offloading 80% to the edge side and 20% to the core side.

[0054] If the target side is the core side, computing tasks can be offloaded to the slice corresponding to the slice identifier in the core side according to the offload ratio and offload path, and the slice quota can be adjusted to execute the offloading of computing tasks. When offloading computing tasks to the core side, the task offloading can be achieved according to the third offload ratio of the core side. At the same time, the core network can include multiple slices, and the slice identifier can be determined according to the user equipment routing selection policy, and then the computing tasks can be offloaded to the slice corresponding to the slice identifier. If it is detected that the resources of the slice corresponding to the slice identifier are insufficient, slice resources can be allocated to it, and the allocated slice resources can be determined according to the service resource requirements. If it is detected that the service is suitable for the slice corresponding to the slice identifier, the service is imported to the slice through end-to-end network elements. Specifically, the packet header of the service carries the slice identifier, and then the service is sent to the slice according to the slice identifier. Alternatively, eMBB access can be restricted during congestion.

[0055] Specifically, user equipment routing policy templates can be generated based on standard protocols, and slice quotas can be adjusted according to core network analysis data. The standard protocol can be 3GPP TS 23.503, which converts user equipment routing policies into standard URSP parameters, seamlessly connecting with 3GPP protocols. This avoids the reliance on manual configuration in traditional offloading algorithms, achieving automated adaptation of 5G native policies and ensuring deployment on existing 5G core networks.

[0056] If the target edge node among multiple edge nodes on the edge side meets the offloading requirements, the computing tasks will be offloaded to the target edge node according to the offloading ratio of the edge side. Specifically, if the edge side includes multiple edge nodes, and the target edge node among these multiple edge nodes can meet both the offloading ratio and latency requirements, the computing tasks can be offloaded to the target edge node on the edge side according to the offloading ratio of the edge side in the user equipment routing selection strategy. For example, if edge node A meets the offloading requirements, 80% of the computing tasks can be offloaded to edge node A.

[0057] Figure 3 The diagram illustrates the dynamic unloading interaction. (See reference...) Figure 3 As shown, the computing power mainly includes the terminal side, edge side, and core side. The computing power on the terminal side prioritizes privacy-sensitive tasks (such as data de-identification and preprocessing), while the computing power nodes on the edge side perform medium-complexity model inference (such as object detection) and support ATSSS multi-path transmission to ensure reliability. The computing power nodes on the core side are responsible for large model training and non-real-time analysis, and global resource scheduling is achieved through the core network PCF.

[0058] The terminal side collects model task characteristics and device resource status in real time. Model task characteristics can include model computation requirements, such as model complexity, latency sensitivity, input data volume, and privacy level. Device resource status can include the terminal device's local resources (remaining CPU / GPU computing power, memory, battery power, etc.) to execute local lightweight model computation tasks.

[0059] The edge-side dynamic offloading controller receives terminal-side data including local privacy processing capabilities and energy consumption, as well as core-side data including slice billing costs and signaling overhead costs. It runs the URSP dynamic offloading decision algorithm to generate the optimal URSP policy and sends a URSP policy request to the core network PCF. Specifically, the edge-side dynamic offloading controller combines the network slice status issued by the core network PCF, such as URLLC / eMBB availability, RSRP, and load, to calculate the action value and overall cost, and generates the URSP policy through the policy network. The URSP policy includes the decision slice ID, QoS, and tiered offloading ratio.

[0060] The edge-side dynamic offload controller performs URSP policy conversion, encapsulates the slice ID and service quality identifier 5QI in the URSP policy into the core network signaling format, and requests the URSP policy from the core network PCF.

[0061] The core network PCF reviews the URSP policy application submitted by the edge side. If the review is approved, it sends a confirmation reply to the edge side and distributes the URSP policy to the terminal side for execution. The PCF receives data such as computing load reported by the edge nodes on the edge side, combines it with network status (slice load, RTT), coordinates resources of multiple edge nodes, ensures the global service level agreement (SLA), generates user equipment routing policy URSP templates based on 3GPP TS23.503, adjusts slice quotas according to the 5G core network network analysis function (NWDAF) data, or restricts eMBB slice access when congestion occurs, and distributes it to the terminal side for execution.

[0062] In this embodiment, real-time data fusion at the terminal, edge, and core network levels enables millisecond-level dynamic policy generation, allowing offloading decisions to respond quickly to changes in task and network status. Using latency, energy consumption, and cost as comprehensive inputs, reinforcement learning is employed to model URSP policy generation as a Markov decision process. Through multi-objective joint optimization, the accuracy and comprehensiveness of user equipment routing strategies are improved. Multi-level offloading is achieved through user equipment routing strategies, enhancing computational resource utilization and reducing energy consumption.

[0063] By optimizing the multi-level offloading ratio at the terminal, edge, and core sides in real time, the utilization rate of computing resources is improved, while the load balancing effect of 5G network slicing is enhanced. Through reasonable hierarchical offloading, the execution latency of model tasks is significantly reduced, and the power consumption of terminal devices is effectively reduced, achieving end-to-end performance optimization. By converting URSP decision results into standard URSP parameters, seamless integration with 3GPP protocols is achieved, while avoiding the reliance on manual configuration in traditional offloading algorithms. This enables automated adaptation to native 5G strategies and achieves standardized compatibility.

[0064] In some embodiments of this disclosure, a dynamic offloading system for computing tasks is also provided, see reference. Figure 4 As shown, the dynamic offloading system 400 for the computing task includes: Terminal 401 is used to collect model task characteristics and device resource status of computing tasks and report them to the edge side. Edge side 402 receives the model task features, device resource status, network status and cost data, and performs joint optimization on the model task features, device resource status, network status and cost data based on the dynamic offloading decision algorithm to generate user equipment routing selection strategy; The core side 403 is used to distribute the user equipment routing policy to the terminal side and control the terminal side to perform hierarchical offloading of computing tasks according to the user equipment routing policy.

[0065] The system employs a three-tiered collaboration between the terminal, edge, and core sides. Terminal-side computing power prioritizes privacy-sensitive tasks (such as data anonymization and preprocessing), while edge-side computing nodes perform medium-complexity model inference (such as object detection), supporting ATSSS multi-path transmission to ensure reliability. Core-side computing nodes handle large-scale model training and non-real-time analysis, achieving global resource scheduling through the core network's PCF.

[0066] Reinforcement learning is employed to model URSP policy generation as a Markov decision process. URSP policies are dynamically generated through joint optimization of the state space (including model task characteristics, network load, computing resources, and cost assessment) and the action space (including slice selection, offloading ratio, and multi-path routing). Furthermore, a temperature coefficient is introduced to adaptively adjust the exploration intensity, automatically adjusting the policy conservatism under high load or high reliability scenarios. On the terminal side, real-time data collection is used to assess model task computational requirements (e.g., latency sensitivity, data volume, privacy level) and local terminal resources (e.g., CPU / GPU computing power remaining, memory, and battery power remaining). The edge-side dynamic offloading controller, combined with network status and cost data from the core network, generates URSP policies in real time.

[0067] In one exemplary embodiment of this disclosure, the edge side performs joint optimization on the model task characteristics, device resource status, network status, and cost data based on a dynamic offloading decision algorithm to generate a user equipment routing strategy, including: The model task features, device resource status, and network status are taken as the current state, and the action probability distribution output by the policy network is determined based on the original probability of executing the target action in the current state. Determine the action value function based on the target action and the current state; Based on the action value function, the comprehensive cost of executing the target action determined according to the cost data, and the temperature coefficient, cost parameters are determined; the target action is selecting a target slice or a target unloading path, and the generation method of the target action is determined according to the temperature coefficient. The user equipment routing strategy is determined based on the action probability distribution and the cost parameters.

[0068] In one exemplary embodiment of this disclosure, the user equipment routing policy includes at least one of slice identifier, offload path, multi-level offload ratio, quality of service, and quality of service identifier; the step of requesting a user equipment routing policy from the core side includes: The slice identifier and quality of service identifier in the offloading decision result are encapsulated into a core network signaling format to request a user equipment routing policy from the core side.

[0069] In one exemplary embodiment of this disclosure, the hierarchical offloading of computing tasks according to the user equipment routing policy includes: The offloading ratio of the computing task on the target side is determined based on the multi-level offloading ratio in the user equipment routing selection strategy. According to the stated unloading ratio and unloading path, the computing task is unloaded to the target side to execute the unloading of the computing task.

[0070] In an exemplary embodiment of this disclosure, the target side is the core side; the step of unloading the computing task to the target side according to the unloading ratio and unloading path includes: According to the aforementioned unloading ratio and unloading path, the computing task is unloaded to the slice corresponding to the slice identifier in the core side, and the slice quota is adjusted for the slice to execute the unloading of the computing task.

[0071] In one exemplary embodiment of this disclosure, adjusting the slice quota for the slice corresponding to the slice identifier in the target side includes: User equipment routing policy templates are generated based on standard protocols, and the slice quotas are adjusted according to core network analysis function data.

[0072] In one exemplary embodiment of this disclosure, the target side is the edge side, and the step of unloading the computing task to the target side according to the unloading ratio and the unloading path includes: If the target edge node among the multiple edge nodes included on the edge side meets the unloading requirements, the computing task is unloaded to the target edge node according to the unloading ratio of the edge side and the unloading path.

[0073] It should be noted that the specific details of each part of the above-mentioned dynamic unloading system for computing tasks have been described in detail in some implementations of the corresponding methods. For details that are not disclosed, please refer to the implementation content of the method section, and therefore will not be repeated here.

[0074] Exemplary embodiments of this disclosure also provide an electronic device. This electronic device may be the aforementioned terminal device or server. Generally, the electronic device may include a processor and a memory, the memory storing executable instructions of the processor, and the processor configured to perform the aforementioned dynamic offloading method for computational tasks by executing the executable instructions. Furthermore, the electronic device may also include a display for displaying an operating interface.

[0075] Those skilled in the art will understand that various aspects of this disclosure can be implemented as a system, method, or program product. Therefore, various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system."

[0076] The following reference Figure 5 To describe an electronic device 500 according to such an embodiment of the present disclosure. Figure 5 The electronic device 500 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0077] like Figure 5 As shown, the electronic device 500 is manifested in the form of a general-purpose computing device. The components of the electronic device 500 may include, but are not limited to: at least one processing unit 510, at least one storage unit 520, a bus 530 connecting different system components (including storage unit 520 and processing unit 510), and a display unit 540.

[0078] The storage unit stores program code that can be executed by the processing unit 510, causing the processing unit 510 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure. For example, the processing unit 510 can perform actions such as... Figure 1 The steps are shown.

[0079] Storage unit 520 may include readable media in the form of volatile storage units, such as random access memory (RAM) 5201 and / or cache memory 5202, and may further include read-only memory (ROM) 5203.

[0080] Storage unit 520 may also include a program / utility 5204 having a set (at least one) program module 5205, such program module 5205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0081] Bus 530 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0082] Electronic device 500 can also communicate with one or more external devices 600 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 500, and / or with any device that enables electronic device 500 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 550. Furthermore, electronic device 500 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 560. As shown, network adapter 560 communicates with other modules of electronic device 500 via bus 530. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 500, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0083] It should be noted that some embodiments of this disclosure also provide a computer program product, which includes a computer program that implements the above-described method when executed by a processor.

[0084] In one implementation, the computer program product can be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. The readable storage medium can be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, including but not limited to: random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory, hard disk drive (HDD), solid-state drive (SSD), etc. For example, the computer program product can be implemented as a non-volatile storage medium storing a computer program, such as read-only memory, NAND flash memory, etc.

[0085] In one implementation, the computer program product can be an intangible product containing a computer program. For example, the computer program product can be implemented as a virtual digital product, such as an executable file, installation package, or other digital file storing the computer program.

[0086] Computer program code can be written in one or more programming languages. Examples of programming languages ​​include C, Java, and C++. Program code can execute entirely on the user's computing device, partially on the user's computing device, or as a standalone software package. It can also execute partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, such as a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via an internet connection provided by a mobile network operator).

[0087] Computer programs can be carried or transmitted via signals such as electrical, magnetic, optical, electromagnetic, and infrared rays. Electronic devices can convert signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program runs on an electronic device, its code is used to cause the electronic device to execute (more specifically, to be executed by the processor of the electronic device) the method steps of various exemplary embodiments of this disclosure.

[0088] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0089] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0090] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for dynamically unloading computational tasks, characterized in that, include: The edge side receives the reported model task characteristics and device resource status of the computing task, and also receives the core side data sent down; the core side data includes network data and cost data. On the edge side, a dynamic offloading decision algorithm is used to jointly optimize the model task characteristics, device resource status, network status, and cost data to generate a user equipment routing strategy. The system requests a user equipment routing policy from the core side, so that the core side can distribute the user equipment routing policy to the terminal side and perform hierarchical offloading of computing tasks according to the user equipment routing policy.

2. The method for dynamically unloading computational tasks according to claim 1, characterized in that, The edge side performs joint optimization on the model task characteristics, device resource status, network status, and cost data based on a dynamic offloading decision algorithm to generate a user equipment routing strategy, including: The model task features, device resource status, and network status are taken as the current state, and the action probability distribution output by the policy network is determined based on the original probability of executing the target action in the current state. Determine the action value function based on the target action and the current state; Based on the action value function, the comprehensive cost of executing the target action determined according to the cost data, and the temperature coefficient, cost parameters are determined; the target action is selecting a target slice or a target unloading path, and the generation method of the target action is determined according to the temperature coefficient. The user equipment routing strategy is determined based on the action probability distribution and the cost parameters.

3. The method for dynamically unloading computational tasks according to claim 1, characterized in that, The user equipment routing strategy includes at least one of slice identifier, offload path, multi-level offload ratio, quality of service, and quality of service identifier. The request for user equipment routing selection policy from the core side includes: The slice identifier and quality of service identifier in the offloading decision result are encapsulated into a core network signaling format to request a user equipment routing policy from the core side.

4. The method for dynamically unloading computational tasks according to claim 3, characterized in that, The hierarchical offloading of computational tasks according to the user equipment routing selection policy includes: The offloading ratio of the computing task on the target side is determined based on the multi-level offloading ratio in the user equipment routing selection strategy. According to the stated unloading ratio and unloading path, the computing task is unloaded to the target side to execute the unloading of the computing task.

5. The method for dynamically unloading computational tasks according to claim 4, characterized in that, The target side is the core side; the process of unloading the computing task to the target side according to the unloading ratio and unloading path includes: According to the aforementioned unloading ratio and unloading path, the computing task is unloaded to the slice corresponding to the slice identifier in the core side, and the slice quota is adjusted for the slice to execute the unloading of the computing task.

6. The method for dynamically unloading computational tasks according to claim 5, characterized in that, The adjustment of slice quotas for slices corresponding to slice identifiers in the target side includes: User equipment routing policy templates are generated based on standard protocols, and the slice quotas are adjusted according to core network analysis function data.

7. The method for dynamically unloading computational tasks according to claim 4, characterized in that, The target side is the edge side. The process of unloading the computing task to the target side according to the unloading ratio and unloading path includes: If the target edge node among the multiple edge nodes included on the edge side meets the unloading requirements, the computing task is unloaded to the target edge node according to the unloading ratio of the edge side and the unloading path.

8. A dynamic unloading system for computational tasks, characterized in that, include: On the terminal side, it is used to collect model task characteristics and device resource status of computing tasks and report them to the edge side; On the edge side, the system receives the model task features, device resource status, network status, and cost data. Based on the dynamic offloading decision algorithm, it performs joint optimization on the model task features, device resource status, network status, and cost data to generate a user equipment routing selection strategy. On the core side, the system is used to distribute the user equipment routing policy to the terminal side and control the terminal side to perform hierarchical offloading of computing tasks according to the user equipment routing policy.

9. An electronic device, characterized in that, include: processor; and memory for storing the executable instructions of the processor; The processor is configured to execute the dynamic offloading method for the computing task according to any one of claims 1-7 by executing the executable instructions.

10. A computer program product, characterized in that, When the computer program is executed by the processor, it implements the dynamic unloading method for the computing task as described in any one of claims 1-7.

Images