A Localized Low-Power Computing Aggregation and Scheduling System and Method Based on Heterogeneous SoC

Abstract

This invention relates to the field of data processing technology, and more particularly to a localized low-power computing power aggregation and scheduling system and method based on heterogeneous SoCs. The system includes: a local networking module, a modeling module, a scheduling and allocation module, an execution module, a result aggregation module, and an optimization scheduling module. This invention constructs a technology chain from dynamic perception to intelligent decision-making to closed-loop control, deeply coupling the multi-dimensional parameters of computing nodes with the computational load, data dependencies, and other demand parameters of tasks. These parameters are input into a scheduling model with total system energy consumption as the optimization objective. In a battery-powered local dynamic network, the system can adaptively select the node combination and task decomposition method with the lowest energy cost, effectively solving the problem that reliable and efficient persistent computing power aggregation cannot be achieved in such scenarios because the optimization objective and resource perception model do not conform to the real constraints of low-power dynamic networks.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a localized low-power computing power aggregation and scheduling system and method based on heterogeneous SoC. Background Technology

[0002] With the widespread adoption of smart IoT terminals and edge devices, computing tasks are rapidly shifting from the cloud to the network edge and user proximity, in pursuit of low latency, high privacy, and strong real-time service experiences. However, in localized, self-organized scenarios comprised of battery-powered portable devices such as portable Wi-Fi devices, mobile gateways, and embedded industrial control devices, these heterogeneous devices exhibit varying computing power, limited battery capacity, and dynamically changing network connections with highly unstable topologies. How to securely and reliably aggregate dispersed and heterogeneous computing power and efficiently complete increasingly complex computational inference tasks under such harsh conditions of severely limited resources and frequent state fluctuations has become a core challenge driving the true implementation of edge intelligence.

[0003] Chinese Patent Application Publication No. CN114449671A discloses a method for collaborative scheduling of heterogeneous resources at the network edge. The method includes: S1. Dividing user requests into three sub-queues: direct delivery, computation processing, and cloud acquisition; establishing a user request queuing model at the network edge; calculating the average latency of user requests based on the queuing model; using a caching strategy as the optimization object; using minimizing the average latency of user requests as the objective function; and using the resource limitations corresponding to caching and computation as constraints, constructing three resource collaborative scheduling optimization models; S2. Solving the above collaborative scheduling optimization models to obtain the minimized average latency of user requests and the corresponding resource quantity under each resource bottleneck; S3. Allocating communication, computation, and caching resources to the network edge according to the resource quantity corresponding to each resource bottleneck.

[0004] Therefore, the existing technology has the following problems: the method has a single optimization objective, and its scheduling model takes minimizing the average latency of user requests as the only optimization objective, without taking system energy consumption into consideration. It cannot be applied to battery-powered local device scenarios that are extremely sensitive to power consumption; the resource modeling of this method is static and coarse. Its coordinated scheduling is macroscopic communication, computing, and cache resources, and it lacks perception and quantitative modeling of multi-dimensional fine-grained states such as real-time computing power, dynamic power consumption, and remaining power of heterogeneous computing nodes. Summary of the Invention

[0005] To address this, the present invention provides a localized low-power computing power aggregation and scheduling system and method based on heterogeneous SoCs. This system overcomes the problem in existing technologies where reliable and efficient persistent computing power aggregation cannot be achieved in such scenarios because the optimization objectives and resource awareness models do not conform to the real constraints of low-power dynamic networks. This is achieved by implementing fine-grained task sharding and scheduling with energy consumption as the core constraint in dynamic local networking.

[0006] To achieve the above objectives, in one aspect, the present invention provides a localized low-power computing power aggregation and scheduling system based on heterogeneous SoC, comprising:

[0007] The local networking module is used to establish a local communication network between the master node and each computing node, wherein the computing nodes include heterogeneous SoC or MCU terminals with different computing power capabilities.

[0008] The modeling module is used to discover and authenticate the computing nodes connected to the local communication network, and to collect the computing power, real-time power consumption status, power information and network communication status of each computing node in real time to generate a node capability profile.

[0009] The scheduling and allocation module is used to decompose the task to be executed into several task fragments according to the node capability profile and the task requirements of the task to be executed, perform optimization calculations with system energy consumption as the core constraint, generate task scheduling decisions, and allocate the task fragments to the selected computing nodes.

[0010] An execution module is used to execute the assigned tasks on each of the computing nodes and generate execution results;

[0011] The result aggregation module is used to summarize and merge all the execution results to generate the final task result;

[0012] The optimization scheduling module is used to monitor the running status of each computing node in the local communication network, and when it detects that a computing node has joined, gone offline or has experienced a performance abnormality, it triggers rescheduling of the pending tasks based on the execution status of each pending task.

[0013] Furthermore, the local networking module includes:

[0014] A networking protocol unit is used to run a protocol stack conforming to the IEEE 802.11ah standard to build a wireless peer-to-peer network between the master node and each of the computing nodes;

[0015] The link management unit is used to maintain the network connection status with each of the computing nodes and to update the network topology when a computing node joins or leaves.

[0016] Furthermore, the modeling module includes:

[0017] A discovery and detection timing unit is used to broadcast discovery beacons at preset intervals and respond to access requests from the computing nodes;

[0018] The multi-dimensional sensing unit is used to actively detect the real-time information of the computing node after it is connected. The real-time information includes CPU / NPU computing power frequency, current available memory, remaining battery capacity, instantaneous power consumption of the chip, network round-trip latency with the master node, and packet loss rate.

[0019] A profile modeling unit is used to fuse the real-time information with the static hardware configuration information of the computing node to generate a profile of the node's capabilities.

[0020] Furthermore, the scheduling and allocation module includes:

[0021] The task parsing and sharding unit receives tasks to be executed, parses them into several task shards with dependencies according to the task type, and calculates the computational load and data throughput requirements of each task shard.

[0022] The energy consumption cost determination unit is used to determine the total execution cost of each task segment and the corresponding computing node pair based on the node capability profile and the computational and data throughput requirements of the task segment, according to a preset cost calculation function.

[0023] The scheduling and allocation unit is used to minimize the total execution cost as the core optimization objective and meet the overall task completion time limit as the constraint. It solves the task allocation scheme under the constraint condition, generates scheduling decisions, and allocates the task fragments to the selected computing nodes according to the scheduling decisions.

[0024] Furthermore, the task parsing and fragmentation unit includes:

[0025] The task type determination subunit is used to identify whether the task to be executed is a parallelizable general computing task or a lightweight artificial intelligence inference task.

[0026] A general computing sharding subunit is used to divide the task to be executed into several task shards according to its process parallelism when the task to be executed is identified as the general computing task, and to estimate the number of processor operations required for each task shard as the computational load, and the size of its input data and output data as the data throughput requirement.

[0027] The AI ​​model sharding subunit is used to divide the task into several task shards based on the structural dependencies of its model computation graph when the task to be executed is identified as the lightweight artificial intelligence inference task. The computational volume is estimated based on the number of floating-point operations, and the data throughput requirement is estimated based on the size of the inter-layer feature map.

[0028] Furthermore, the energy consumption cost determination unit includes:

[0029] The cost parameter acquisition unit is used to calculate the wireless communication power consumption, the active power consumption of the computing node, the overhead of waking up from the dormant state, the estimated execution time of the task fragment, and the network transmission time based on the node capability profile.

[0030] The cost calculation subunit is used to input the wireless communication power consumption, the active power consumption of the computing node, the overhead of waking up from the sleep state, the estimated execution time, and the network transmission time into the preset cost calculation function to calculate the total execution cost.

[0031] Furthermore, the execution module includes:

[0032] A task container unit, which is used to build an isolated runtime environment on the compute node for receiving, loading and executing the code and data of the task fragments;

[0033] The batch processing unit is used to maintain a task queue locally on the computing node, merge all task fragments that belong to the same or different tasks to be executed within a preset batch processing time into an execution batch, and inject each task fragment in the execution batch into the isolated runtime environment in sequence to trigger its loading and execution, and control the computing node to enable the wireless communication module only within a preset communication time window to complete the batch reception of data and result return required by the execution batch.

[0034] Furthermore, the optimized scheduling module includes:

[0035] A status tracking unit is used to monitor the online status, load rate and power consumption level of each computing node in real time, and to record the execution progress of each task slice.

[0036] The event response unit is used to generate a rescheduling strategy, including task migration, task replication, or task restart, based on the execution progress and remaining computing power of the task shard when a node offline, performance degradation, or abnormal power consumption event is detected.

[0037] The elastic rescheduling unit is used to reselect target nodes for the affected task fragments according to the rescheduling strategy and the node capability profile of the currently available computing nodes, and to trigger the transmission and state synchronization of the task fragments.

[0038] Furthermore, the batch processing unit includes:

[0039] The window adaptive subunit is used to dynamically adjust the preset batch processing duration and the period and length of the preset communication time window based on the remaining power of the computing node, the urgency of the task queue, and the network channel quality.

[0040] The conflict avoidance subunit is used to allocate staggered communication time windows to each computing node when several computing nodes share the same master node, using a time-division multiple access mechanism.

[0041] On the other hand, the present invention also provides a localized low-power computing power aggregation and scheduling method based on heterogeneous SoC, comprising:

[0042] A local communication network is established between the master node and each computing node, wherein the computing nodes include heterogeneous SoC or MCU terminals with different computing power capabilities.

[0043] The computing nodes connected to the local communication network are discovered and authenticated, and the computing power, real-time power consumption status, power information and network communication status of each computing node are collected in real time to generate a node capability profile.

[0044] Based on the node capability profile and the task requirements of the task to be executed, the task to be executed is decomposed into several task fragments. The system energy consumption is used as the core constraint for optimization calculation, task scheduling decision is generated, and the task fragments are allocated to the selected computing nodes.

[0045] The assigned tasks are executed on each of the computing nodes, and execution results are generated.

[0046] All execution results are summarized and merged to generate the final task result;

[0047] The system monitors the operating status of each computing node within the local communication network, and when a computing node is detected to be joining, going offline, or experiencing performance abnormalities, it triggers rescheduling of tasks that have not yet been completed based on the execution status of each task to be executed.

[0048] Compared with existing technologies, the beneficial effects of this invention lie in constructing a technological chain from dynamic perception to intelligent decision-making and then to closed-loop control. This deeply couples multi-dimensional parameters such as the real-time computing power, power consumption, battery level, and network status of computing nodes with demand parameters such as the computational load and data dependencies of tasks, inputting them into a scheduling model with total system energy consumption as the optimization objective. This model accurately depicts the inherent physical constraints and probabilistic relationships between computing energy consumption and execution time, communication energy consumption and data volume and network quality, as well as node reliability and its historical state. This ensures that every scheduling decision is a global balance of these parameters under the goal of optimal energy consumption. In battery-powered local dynamic networks, the system can adaptively select the node combination and task decomposition method with the lowest energy consumption cost. This not only significantly reduces the overall computing and communication power consumption and extends the device's battery life, but also transforms the instability of nodes into manageable scheduling risks through a closed-loop mechanism of "monitoring-evaluation-rescheduling". Thus, in edge scenarios with strictly limited resources and frequent topology changes, it simultaneously improves the efficiency of computing power aggregation and ensures the robustness of long-term system operation. This effectively solves the problem that reliable and efficient persistent computing power aggregation cannot be achieved in such scenarios because the optimization objectives and resource awareness models do not conform to the real constraints of low-power dynamic networks.

[0049] Furthermore, by selecting the IEEE 802.11ah protocol stack optimized for the Internet of Things and implementing proactive link management, a network foundation combining low power consumption, wide coverage, and high dynamic adaptability is constructed. The protocol stack operates in the sub-1GHz frequency band, and its physical layer design provides stronger signal penetration and wider coverage. The media access control layer mechanism significantly reduces the energy consumption of devices during idle listening and channel contention, which aligns with the fundamental requirements of battery-powered devices for communication distance and battery life. The link management unit maintains a dynamic connection table and listens for keep-alive messages, precisely correlating the physical layer signal strength of nodes, the timing patterns of keep-alive messages, and the logical connection status of the network: signal strength reflects real-time changes in spatial path loss and interference, while the regularity of keep-alive messages directly characterizes the node's survival and the health of the protocol stack. This module, through dynamic request correlation and timeout determination mechanisms, transforms the continuous changes at the physical and protocol layers into discrete "join" or "offline" logical events, thereby providing the upper layers with a real-time, accurate, and stable network topology view. This allows the entire computing power scheduling system to run on a reliable communication base, and its scheduling decisions are not affected by frequent interruptions or drastic quality fluctuations in the underlying links, thus ensuring the continuity and success rate of aggregate computing tasks from the network level.

[0050] Furthermore, by combining periodic proactive discovery with fine-grained multi-dimensional perception, a dynamic and accurate node capability profile is constructed. The real-time frequency and available memory of the CPU / NPU directly determine the node's instantaneous computing throughput limit; remaining battery capacity and instantaneous chip power consumption jointly constrain the node's sustainable operating time and energy efficiency ratio; network round-trip latency and packet loss rate quantify the timeliness and reliability costs of data exchange. These parameters are not isolated but interconnected through inherent physical constraints and probabilistic relationships: improved computing performance is often accompanied by increased power consumption, while declining battery power limits sustained performance output; fluctuations in network quality directly affect the communication overhead and completion determinism of remote computing tasks. The profile modeling unit integrates this dynamic information with static hardware configuration, essentially providing the scheduling system with a complete computing entity model that combines real-time status and inherent capabilities. This allows subsequent energy-consumption-centric optimization scheduling to be built on a quantitative foundation that accurately depicts the coupling relationship between "computing power-power consumption-power supply-network," providing a reliable data foundation for achieving truly low-power computing power aggregation.

[0051] Furthermore, by constructing a unified quantitative decision-making framework, the computational load and data volume of tasks are intrinsically coupled and globally optimized with dynamic parameters such as the real-time computing power, power consumption, and network status of nodes. Specifically, the number of floating-point operations in task sharding and the active power consumption of nodes jointly determine the physical basis of computational energy consumption; the data throughput requirements of sharding and the round-trip latency and packet loss rate of the network jointly constrain the actual cost of communication energy consumption; and the wake-up overhead of nodes reflects the additional energy loss caused by state switching. The module flexibly adjusts the relative importance of these different energy consumption sources in the optimization objective through configurable weight coefficients, enabling scheduling decisions to accurately adapt to specific scenarios such as "computationally intensive," "communication-sensitive," or "frequent node hibernation." This allows the system to automatically solve for the precise task allocation scheme with the lowest total energy consumption while meeting deadline constraints based on real-time dynamic parameters. Thus, in complex local dynamic networks, this significantly extends the battery life of devices while ensuring the efficient and reliable completion of computational tasks.

[0052] Furthermore, by designing differentiated sharding and quantization strategies tailored to the inherent characteristics of general-purpose computing and AI inference tasks, the inherent parallelism and computation graph structures of these tasks are precisely mapped into quantifiable parameters directly applicable to physical resource evaluation, such as the number of processor operations, floating-point operations, data block size, and feature graph tensor shape. The analysis of general-purpose computing tasks reveals a direct proportionality between computational load and instruction complexity and data size, while the computation graph deconstruction of AI models explicitly characterizes the inter-layer computational intensity and the dependency path of data flow. This consistent transformation from task logic to physical metrics provides a unified and accurate load representation for downstream scheduling systems. This allows the pre-defined cost calculation function to perform a coordinated trade-off between computation, communication, and storage overhead based on a realistic and complete load profile, thereby fundamentally ensuring the accuracy of subsequent scheduling decisions and the feasibility of overall energy efficiency improvements.

[0053] Furthermore, by decomposing the energy consumption optimization objective into a series of computable parameters based on well-defined physical laws and existing data, and finally deriving a quantified cost through defined algebraic operations, the following parameters are determined: the active power consumption of the computing node, the wireless communication power consumption, and the wake-up overhead are directly derived from hardware measurements or manufacturer datasheets, reflecting the device's basic energy conversion characteristics; the estimated execution time is determined by the ratio of objective computational load to real-time computing power, reflecting the spatiotemporal complexity of the computational task; and the network transmission time is constrained by the amount of data, channel latency, and reliability. These parameters together constitute a complete and observable input set, which is synthesized through a pre-defined linear cost function. Essentially, this process maps physical quantities of different dimensions into a unified and comparable scalar cost through reasonable weight allocation.

[0054] Furthermore, through the collaborative design of isolated execution and communication scheduling, the randomness of task arrival, the continuity of computation execution, and the high power consumption characteristics of wireless communication are systematically decoupled and reorganized. Specifically, the isolated environment provided by the task container unit ensures independent allocation and state clearing of computing resources, providing a definite execution base for continuous, unordered task fragmentation. The batch processing unit utilizes a local queue to absorb fluctuations in task arrival, integrating discrete computation tasks into continuous computation batches through preset durations. This allows the computing unit to maintain a continuous and effective working state, thereby reducing performance loss and static power consumption caused by frequent start-stop cycles. More importantly, this unit forcibly constrains wireless communication activity within a periodic preset time window, precisely aligning the high-power activity time of the communication module with the lifecycle of the computation batches, significantly reducing ineffective power consumption caused by idle listening and frequent protocol interactions. This enables efficient collaboration between computation and communication, two physical processes with vastly different characteristics in terms of power consumption and time scale, ultimately achieving a significant reduction in overall power consumption—the most critical factor for battery-powered devices—while ensuring task execution throughput and security.

[0055] Furthermore, by establishing a closed-loop control chain of "monitoring-diagnosis-recovery," the real-time load, power consumption, and network heartbeat of nodes are deeply correlated with logical states such as task sharding execution progress and remaining computational load. The persistence of the heartbeat signal directly reflects the node's network reachability and basic operational status; abnormal changes in load rate and power consumption reveal chip-level thermal and electrical performance bottlenecks; and the comparison between task execution progress and remaining workload provides crucial evidence for cost-benefit analysis in decision-making. Based on these coupling relationships, when a node is detected to be offline or experiencing performance degradation, this module can automatically and intelligently select differentiated recovery strategies such as migration, replication, or restart based on the proportion of tasks already completed. Essentially, this involves a quantitative trade-off between the sunk cost of computational tasks, the overhead of re-execution, and the timeliness of task completion. Subsequently, the elastic rescheduling unit, based on the latest global resource view, quickly matches suitable new nodes for affected tasks and synchronizes their execution context, thereby transforming dynamic disturbances into a controlled, optimal, or near-optimal rescheduling process. This allows the entire system to proactively manage uncertainty instead of passively accepting node changes, ultimately achieving a significant improvement in task success rate and an effective reduction in overall service downtime in a frequently changing local network environment.

[0056] Furthermore, by coordinating the batch processing duration and communication window, the three core physical and logical constraints—remaining battery power, task urgency, and channel quality—are mapped to a joint optimization of system energy consumption, latency, and reliability. Remaining battery power, acting as an energy budget, directly restricts the total number of times and duration the wireless communication module can be activated. Task urgency defines the lower limit of the service's latency tolerance, exerting reverse pressure on the cumulative batch processing time. Channel quality determines the amount of data reliably transmitted per unit time, affecting the efficiency of a single communication window. The window adaptive subunit dynamically balances these mutually constraining parameters based on preset rules: prioritizing extended batch processing to save communication energy during low battery periods, shortening batch processing for low latency during high-urgency tasks, and compensating for transmission reliability by increasing the window size when network quality is poor. Essentially, this solves a real-time approximate solution to a multi-objective optimization problem. Simultaneously, the conflict avoidance subunit introduces a time-division multiple access mechanism, transforming the random competition among multiple nodes for the shared physical channel into a deterministic, conflict-free timing schedule, fundamentally avoiding retransmissions and energy waste caused by packet collisions. The combination of these two elements enables the system to maintain an optimal balance between communication energy consumption, task latency, and transmission success rate under dynamically changing resource and demand conditions.

[0057] Furthermore, by constructing a local dynamic network between the master node and heterogeneous computing nodes, the computing power, power consumption, battery level, and network status of each node are monitored in real time to form a capability profile. Based on this profile, task decomposition and intelligent scheduling are performed with the core objective of minimizing total system energy consumption. Simultaneously, a rescheduling mechanism is introduced for dynamic node addition, offline status, and abnormal node behavior. This method enables the efficient aggregation and collaboration of distributed, battery-powered heterogeneous computing resources in a local weak network environment. Ultimately, it significantly reduces overall system energy consumption, extends device battery life, and effectively improves the success rate of computing tasks and system robustness in the dynamic network. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the localized low-power computing power aggregation and scheduling system based on heterogeneous SoC in this embodiment;

[0059] Figure 2 This is a schematic diagram of the modeling module in this embodiment;

[0060] Figure 3 This is a schematic diagram of the scheduling and allocation module in this implementation;

[0061] Figure 4 This is a flowchart of the localized low-power computing power aggregation and scheduling method based on heterogeneous SoC in this embodiment. Detailed Implementation

[0062] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0063] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0064] Please see Figure 1 As shown, this is a schematic diagram of the localized low-power computing power aggregation and scheduling system based on heterogeneous SoC in this embodiment. On one hand, this embodiment provides a localized low-power computing power aggregation and scheduling system based on heterogeneous SoC, including:

[0065] The local networking module 201 is used to establish a local communication network between the master node 101 and each computing node, wherein the computing nodes include heterogeneous SoC or MCU terminals with different computing power capabilities.

[0066] The modeling module 202 is connected to the local networking module and is used to discover and authenticate the computing nodes accessing the local communication network, and to collect the computing power, real-time power consumption status, power information and network communication status of each computing node in real time to generate a node capability profile.

[0067] An execution module 204 is set on each computing node to execute the assigned task on each computing node and generate execution results.

[0068] The result aggregation module 205, which is connected to the scheduling and allocation module, is used to summarize and merge all the execution results to generate the final task result;

[0069] The optimization scheduling module 206 is connected to the scheduling allocation module and the local networking module respectively. It is used to monitor the running status of each computing node in the local communication network, and when a computing node is detected to be joining, going offline or experiencing performance abnormalities, it triggers rescheduling of the pending tasks based on the execution status of each pending task.

[0070] In this embodiment, the localized low-power computing power aggregation and scheduling system is deployed in a localized network consisting of a master node and multiple computing nodes (102a, 102b, ..., 102n). The master node serves as the scheduling and control center and is preferably a device integrating a microcontroller unit and a high-performance SoC; the computing nodes are heterogeneous SoC or MCU terminal devices with different computing power capabilities and power consumption characteristics, such as portable WiFi devices, portable gateways, or embedded development boards.

[0071] In this embodiment, heterogeneous SoC refers to a system-on-a-chip that integrates computing units with different architectures, such as a chip combination containing a general-purpose processor (CPU), a neural network processor (NPU), or a graphics processing unit (GPU). These together constitute computing nodes with varying capabilities. Node capability profiling is a comprehensive quantitative representation of the dynamic operating state of each node. Its core lies in fusing multi-dimensional parameters such as the node's real-time computing frequency, active power consumption, remaining battery power, network round-trip latency, and packet loss rate into a standardized model that can be directly utilized by scheduling algorithms. Task sharding scientifically decomposes a complete computing task, such as a lightweight artificial intelligence model inference, into multiple sub-task units that can be executed in parallel or serially, based on its internal data dependencies or computational graph structure. The energy consumption model upon which the system relies is a key mathematical model that precisely expresses the quantitative relationship between the computing energy consumption, communication energy consumption, and additional overhead required to wake a node from a dormant state generated by executing a specific task shard. The optimization calculation based on this model, with energy consumption as the core constraint, essentially solves how to map each task shard to different computing nodes to minimize the total system energy consumption output by the aforementioned energy consumption model, while ensuring the overall task completion time. The entire system has a closed-loop control mechanism, which means that the system can continuously monitor the network and node status. Once it detects dynamic changes such as node offline, it will automatically re-trigger the complete optimization calculation and task migration process based on the latest node capability profile, thereby forming a self-regulating decision loop to ensure the system's continuous stability and efficiency in dynamic environments.

[0072] By constructing a technological chain from dynamic perception to intelligent decision-making and then to closed-loop control, multi-dimensional parameters such as real-time computing power, power consumption, battery level, and network status of computing nodes are deeply coupled with requirement parameters such as task computation volume and data dependencies. These parameters are then input into a scheduling model with total system energy consumption as the optimization objective. This model accurately depicts the inherent physical constraints and probabilistic relationships between computing energy consumption and execution time, communication energy consumption and data volume and network quality, as well as node reliability and its historical state. This ensures that every scheduling decision is a global balance of these parameters under the goal of optimal energy consumption. In battery-powered local dynamic networks, the system can adaptively select the node combination and task decomposition method with the lowest energy consumption cost. This not only significantly reduces the overall computing and communication power consumption and extends the device's battery life, but also transforms the instability of nodes into manageable scheduling risks through a closed-loop mechanism of "monitoring-evaluation-rescheduling". Thus, in edge scenarios with strictly limited resources and frequent topology changes, it simultaneously improves the efficiency of computing power aggregation and ensures the robustness of long-term system operation. This effectively solves the problem that reliable and efficient persistent computing power aggregation cannot be achieved in such scenarios because the optimization objectives and resource awareness models do not conform to the real constraints of low-power dynamic networks.

[0073] Specifically, the local networking module includes:

[0074] A networking protocol unit is used to run a protocol stack conforming to the IEEE 802.11ah standard to build a wireless peer-to-peer network between the master node and each of the computing nodes;

[0075] The link management unit is used to maintain the network connection status with each of the computing nodes and to update the network topology when a computing node joins or leaves.

[0076] In this embodiment, the networking protocol unit runs within the wireless communication chip of the master node. It establishes communication links on unlicensed frequency bands below 1 GHz by configuring and adhering to the physical layer and media access control layer protocols defined by the IEEE 802.11ah standard. This standard is specifically designed for low-power, wide-coverage IoT scenarios, and its protocol stack implementation enables the master node and compute nodes to directly construct a star or peer-to-peer wireless network without traditional wireless access points. The link management unit acts as the control core of this network connection, continuously maintaining a dynamic connection table that records the network identifier, link signal strength, and last active timestamp of each compute node. This unit determines the real-time online status of nodes by listening to customized link layer beacons and keep-alive messages: when a new node's association request is received, its information is registered in the table and marked as available; when a node's keep-alive message times out and is lost, it is marked as offline and removed from the available resource pool, thereby achieving real-time and accurate updates to the network topology and providing a stable network view foundation for upper-layer scheduling.

[0077] By selecting the IEEE 802.11ah protocol stack optimized for the Internet of Things (IoT) and implementing proactive link management, a network foundation combining low power consumption, wide coverage, and high dynamic adaptability is constructed. The protocol stack operates in the sub-1GHz frequency band, and its physical layer design provides stronger signal penetration and wider coverage. The media access control layer mechanism significantly reduces the energy consumption of devices during idle listening and channel contention, which aligns with the fundamental requirements of battery-powered devices for communication distance and battery life. The link management unit maintains a dynamic connection table and listens for keep-alive messages, precisely correlating the physical layer signal strength of nodes, the timing patterns of keep-alive messages, and the logical connection status of the network: signal strength reflects real-time changes in spatial path loss and interference, while the regularity of keep-alive messages directly characterizes the node's survival and the health of the protocol stack. This module, through dynamic request correlation and timeout determination mechanisms, transforms the continuous changes at the physical and protocol layers into discrete "join" or "offline" logical events, thus providing the upper layers with a real-time, accurate, and stable network topology view. This allows the entire computing power scheduling system to run on a reliable communication base, and its scheduling decisions are not affected by frequent interruptions or drastic quality fluctuations in the underlying links, thus ensuring the continuity and success rate of aggregate computing tasks from the network level.

[0078] Please see Figure 2 As shown, this is a schematic diagram of the modeling module in this embodiment. In this embodiment, the modeling module includes:

[0079] The discovery and detection timing unit 2021 is used to broadcast discovery beacons at preset periods and respond to the access requests of the computing nodes;

[0080] The multi-dimensional sensing unit 2022 is connected to the discovery and detection timing unit to actively detect the real-time information of the computing node after the computing node is connected. The real-time information includes CPU / NPU computing power frequency, current available memory, remaining battery capacity, chip instantaneous power consumption, network round-trip latency with the master node, and packet loss rate.

[0081] The portrait modeling unit 2023, which is connected to the multi-dimensional perception unit, is used to fuse the real-time information with the static hardware configuration information of the computing node to generate the node capability portrait.

[0082] In this embodiment, the discovery and detection timing unit specifically broadcasts UDP discovery packets containing network identifiers and security tokens periodically at configurable fixed time intervals through the network interface of the master node. After the node is authenticated and connected, the multi-dimensional perception unit actively acquires its real-time information by remotely calling a lightweight proxy service on the target node: querying the current operating frequency and available memory of the CPU / NPU through the operating system interface; reading the remaining battery capacity and instantaneous power of the chip through the power management chip; and measuring its network round-trip latency and packet loss rate by sending a series of ICMP or customized probe packets with serial numbers to the node. The profiling and modeling unit then fuses this dynamically acquired real-time information with the static hardware configuration information obtained during the node's pre-registration or initial discovery. The static information includes processor model, total memory, peak NPU computing power, etc. After fusion, a structured node capability profile that can be directly parsed by the scheduling algorithm is generated. This profile is usually stored in memory in the form of key-value pairs or specific data structures and is dynamically updated with periodic perception.

[0083] The preset period is a configurable time parameter, the specific value of which depends on the expected frequency of dynamic changes in network nodes and the requirements for system energy consumption control, and is usually set between 5 and 60 seconds. In this embodiment, it is set to 10 seconds, which can effectively control unnecessary communication power consumption caused by frequent broadcasting of discovery beacons while ensuring the responsiveness of new nodes being discovered in a timely manner.

[0084] By combining periodic proactive discovery with fine-grained multi-dimensional perception, a dynamic and accurate node capability profile is constructed. The real-time frequency and available memory of the CPU / NPU directly determine the node's instantaneous computing throughput limit; remaining battery capacity and instantaneous chip power consumption jointly constrain the node's sustainable operating time and energy efficiency ratio; network round-trip latency and packet loss rate quantify the timeliness and reliability costs of data exchange. These parameters are not isolated but interconnected through inherent physical constraints and probabilistic relationships: improved computing performance is often accompanied by increased power consumption, while declining battery power limits sustained performance output; fluctuations in network quality directly affect the communication overhead and completion determinism of remote computing tasks. The profiling modeling unit integrates this dynamic information with static hardware configuration, essentially providing the scheduling system with a complete computing entity model that combines real-time status and inherent capabilities. This allows subsequent energy-consumption-centric optimization scheduling to be built on a quantitative foundation that accurately depicts the coupling relationship between "computing power-power consumption-power supply-network," providing a reliable data foundation for achieving truly low-power computing power aggregation.

[0085] Please see Figure 3 As shown, this is a schematic diagram of the scheduling and allocation module in this embodiment. In this embodiment, the scheduling and allocation module includes:

[0086] The task parsing and sharding unit 2031 is used to receive tasks to be executed, parse them into several task shards with dependencies according to the task type, and calculate the computational load and data throughput requirements of each task shard.

[0087] The energy consumption cost determination unit 2032 is connected to the task parsing and sharding unit. It is used to determine the total execution cost of each task shard and the corresponding computing node based on the node capability profile and the computational amount and data throughput requirements of the task shard, according to a preset cost calculation function.

[0088] The scheduling and allocation unit 2033, which is connected to the energy consumption cost determination unit, is used to minimize the total execution cost as the core optimization objective and meet the overall task completion time limit as the constraint condition. It solves the task allocation scheme under the constraint condition, generates a scheduling decision, and allocates the task fragments to the selected computing nodes according to the scheduling decision.

[0089] In this embodiment, after receiving a task, the task parsing and sharding unit first analyzes its code structure or model definition to determine the task type. For parallelizable tasks, it divides the task based on data dimensions or iterative loops; for AI inference tasks, it breaks down the model into sequential or parallel stages such as preprocessing, feature extraction, and classification heads based on its computation graph. The computational load of each shard is quantified by estimating the required number of floating-point operations, and the data throughput requirement is determined by the size of its input and output tensors. The preset cost calculation function in the energy consumption cost determination unit has the following core inputs: node active power consumption Pa and wake-up overhead Ew obtained from the node capability profile, the estimated task shard execution time tc, and the network transmission time tt and communication power consumption Pr estimated based on network round-trip latency and shard data volume. The typical form of this function is cost C = α × Pa × tc + β × Pr × tt + γ × Ew, where α, β, and γ are adjustable weights. The scheduling and allocation unit takes the cost calculated for all "shard-node" pairings above, the dependency between task shards, and the overall deadline as constraints, and constructs a combinatorial optimization problem. It then uses a mixed-integer linear programming solver or a heuristic algorithm (such as a greedy algorithm) to solve the problem, and finally outputs a scheduling decision list that maps each task shard to a specific computing node.

[0090] In this embodiment, α is the weighting coefficient for computational energy consumption, and its value depends on the system's optimization priority for node computational power consumption, typically set between 0.5 and 1.5. In this embodiment, it is set to 1.0, allowing computational energy consumption to occupy a baseline proportion in cost evaluation, directly reflecting the contribution of the processor's actual energy consumption to the total system energy consumption. β is the weighting coefficient for communication energy consumption, and its value depends on the system's sensitivity to wireless data transmission power consumption, typically set between 0.2 and 1.0. In this embodiment, it is set to 0.8, emphasizing the reduction of communication overhead while balancing with other energy consumption items, adapting to the limited bandwidth characteristics of the local wireless network. γ is the weighting coefficient for wake-up overhead, and its value depends on the system's emphasis on the performance penalty of node sleep-wake cycle, typically set between 0.1 and 0.5. In this embodiment, it is set to 0.3, effectively suppressing arbitrary wake-ups of low-battery nodes or frequently sleeping nodes, thereby extending the overall battery life of battery-powered devices.

[0091] By constructing a unified quantitative decision-making framework, the computational load and data volume of tasks are intrinsically coupled and globally optimized with dynamic parameters such as the real-time computing power, power consumption, and network status of nodes. Specifically, the number of floating-point operations in task sharding and the active power consumption of nodes jointly determine the physical basis of computational energy consumption; the data throughput requirements of sharding and the round-trip latency and packet loss rate of the network jointly constrain the actual cost of communication energy consumption; and the wake-up overhead of nodes reflects the additional energy loss caused by state switching. The module flexibly adjusts the relative importance of these different energy consumption sources in the optimization objective through configurable weight coefficients, enabling scheduling decisions to accurately adapt to specific scenarios such as "computation-intensive," "communication-sensitive," or "frequent node sleep." This allows the system to automatically solve for the precise task allocation scheme with the lowest total energy consumption while meeting deadline constraints based on real-time dynamic parameters. Thus, in complex local dynamic networks, it significantly extends the battery life of devices while ensuring the efficient and reliable completion of computational tasks.

[0092] Specifically, the task parsing and sharding unit includes:

[0093] The task type determination subunit is used to identify whether the task to be executed is a parallelizable general computing task or a lightweight artificial intelligence inference task.

[0094] A general computing sharding subunit, connected to a task type determination subunit, is used to divide the task to be executed into several task shards based on its process parallelism when the task to be executed is identified as the general computing task. The subunit also estimates the number of processor operations required for each task shard as the computational load and the size of its input and output data as the data throughput requirement.

[0095] The AI ​​model sharding subunit is connected to the task type judgment subunit. When the task to be executed is identified as the lightweight artificial intelligence inference task, it is divided according to the structural dependency relationship of its model computation graph to obtain several task shards. The computational volume is estimated based on the number of floating-point operations, and the data throughput requirement is estimated based on the size of the inter-layer feature map.

[0096] In this embodiment, the task type determination subunit can be implemented by parsing the preset metadata fields in the task descriptor. For example, when the value of the field is "general computing" or "AI inference", the corresponding judgment is made. The division of the general computing sharding subunit is based on the standard parallel computing paradigm: for data parallel tasks, it is divided equally or by weight according to the dimension of its input data; for task parallel tasks, the independent function modules or loop iterations in the program are identified and separated into subtasks. The number of processor operations required can be statically estimated by analyzing the loop structure and arithmetic instructions of the code, or dynamically sampled by running performance profiling tools on representative data; its data throughput requirement is the size of the input data block processed and the output data block produced by the subtask. The AI ​​model sharding subunit directly calls the front-end interface provided by the current general deep learning framework, loads the model and obtains its computation graph representation, and then cuts the computation graph into multiple logically continuous subgraph segments according to the data dependencies between the operator nodes in the graph. The computational cost of each sub-segment is calculated using floating-point operations, by summing the FLOPs of all its internal operators; its data throughput requirements are calculated based on the shape, data type, and precision of the input and output feature map tensors of that sub-segment.

[0097] By designing differentiated sharding and quantization strategies tailored to the inherent characteristics of general-purpose computing and AI inference tasks, the inherent parallelism and computation graph structures of these tasks are precisely mapped into quantifiable parameters directly applicable to physical resource evaluation, such as the number of processor operations, floating-point operations, data block size, and feature graph tensor shape. The analysis of general-purpose computing tasks reveals a direct proportionality between computational load and instruction complexity and data size, while the computation graph deconstruction of AI models explicitly characterizes the inter-layer computational intensity and data flow dependency paths. This consistent transformation from task logic to physical metrics provides a unified and accurate load representation for downstream scheduling systems. This allows the pre-defined cost calculation function to coordinate the trade-offs between computation, communication, and storage overhead based on a realistic and complete load profile, thereby fundamentally ensuring the accuracy of subsequent scheduling decisions and the feasibility of overall energy efficiency improvements.

[0098] Specifically, the energy consumption cost determination unit includes:

[0099] The cost parameter acquisition unit is used to calculate the wireless communication power consumption, the active power consumption of the computing node, the overhead of waking up from the dormant state, the estimated execution time of the task fragment, and the network transmission time based on the node capability profile.

[0100] The cost calculation subunit, connected to the cost parameter acquisition unit, is used to input the wireless communication power consumption, the active power consumption of the computing node, the overhead of waking up from the sleep state, the estimated execution time, and the network transmission time into the preset cost calculation function to calculate the total execution cost.

[0101] In this embodiment, the cost parameter acquisition unit directly obtains the active power consumption of the computing node by querying the real-time power consumption status recorded in the node capability profile; the wireless communication power consumption is determined based on the typical power consumption value of the wireless communication chip used by the computing node in the corresponding working mode or a pre-stored power consumption curve table; the overhead of waking up from the sleep state is obtained by pre-calibrating the fixed energy consumption generated by the node switching from the sleep state to the active state. The estimated execution time is estimated by dividing the computational amount of the task fragment by the current effective computing power frequency of the node obtained from the node capability profile; the network transmission time required is estimated based on the data throughput requirements of the task fragment and the network round-trip latency and packet loss rate in the node capability profile. After obtaining all parameters, the cost calculation subunit substitutes them into the preset cost calculation function and performs the corresponding multiplication and addition arithmetic operations to obtain the quantified total execution cost.

[0102] By decomposing the energy consumption optimization objective into a series of computable parameters based on well-defined physical laws and existing data, and finally deriving a quantified cost through defined algebraic operations, the following parameters are used: active power consumption of computing nodes, wireless communication power consumption, and wake-up overhead are directly derived from hardware measurements or manufacturer datasheets, reflecting the basic energy conversion characteristics of the device; the estimated execution time is determined by the ratio of objective computational load to real-time computing power, reflecting the spatiotemporal complexity of the computational task; and the network transmission time is constrained by data volume, channel latency, and reliability. These parameters together constitute a complete and observable input set, which is synthesized through a pre-defined linear cost function. Essentially, this process maps physical quantities of different dimensions to a unified and comparable scalar cost through reasonable weight allocation.

[0103] Specifically, the execution module includes:

[0104] A task container unit, which is used to build an isolated runtime environment on the compute node for receiving, loading and executing the code and data of the task fragments;

[0105] A batch processing unit, connected to a task container unit, is used to maintain a task queue locally on the computing node. It merges all task fragments that belong to the same or different tasks to be executed within a preset batch processing time into an execution batch, and injects each task fragment in the execution batch into the isolated runtime environment in sequence to trigger its loading and execution. It also controls the computing node to enable the wireless communication module only within a preset communication time window to complete the batch reception of data and result return required by the execution batch.

[0106] In this embodiment, the task container unit is implemented by calling the process isolation or lightweight virtualization interface provided by the computing node operating system. Specifically, an independent process or application sandbox can be created for each received task fragment, and the code and dependent libraries corresponding to the fragment can be loaded into the isolated environment for execution, thereby ensuring resource controllability and security isolation of task execution. The batch processing unit maintains a first-in-first-out task queue at the application layer of the computing node. Its workflow is as follows: continuously receiving task fragments from the network and storing them in the queue; a local timer or scheduling thread manages the preset batch processing duration, and when the duration is reached, all task fragments cached in the current queue are retrieved as an execution batch and submitted to the task container unit for sequential or concurrent execution; at the same time, this unit controls the node's wireless communication module to switch to a high-power active state only during periodic, preset-length communication time windows through the operating system's power management or GPIO interface, so as to centrally complete the return of the results of the previous execution batch and the reception of the data of the next batch. Outside the window period, the wireless module is placed in a low-power sleep or off state.

[0107] The preset batch processing duration is a configurable time threshold. Its specific value depends on the system's requirements for real-time task processing and the trade-off between the energy-saving effect of the wireless communication module. It is usually set between 100 milliseconds and 2 seconds. In this embodiment, it is set to 500 milliseconds, which can effectively accumulate enough task fragments to form a significant communication batch effect without excessively increasing the task queuing delay, thereby maximizing the energy utilization rate of a single wireless module activation window.

[0108] By employing a collaborative design of isolated execution and communication scheduling, the randomness of task arrival, the continuity of computation execution, and the high power consumption of wireless communication are systematically decoupled and reorganized. Specifically, the isolated environment provided by the task container unit ensures independent allocation and state clearing of computing resources, providing a definite execution base for continuous, unordered task fragmentation. The batch processing unit utilizes a local queue to absorb fluctuations in task arrival, integrating discrete computation tasks into continuous computation batches through preset durations. This allows the computing unit to maintain a continuous and effective working state, thereby reducing performance loss and static power consumption caused by frequent start-stop cycles. More importantly, by forcibly constraining wireless communication activity within a periodic preset time window, this unit precisely aligns the high-power activity time of the communication module with the lifecycle of the computation batches, significantly reducing ineffective power consumption caused by idle listening and frequent protocol interactions. This enables efficient collaboration between computation and communication, two physical processes with vastly different characteristics in terms of power consumption and time scale, ultimately achieving a significant reduction in overall power consumption—the most critical factor for battery-powered devices—while ensuring task execution throughput and security.

[0109] Specifically, the optimized scheduling module includes:

[0110] A status tracking unit is used to monitor the online status, load rate and power consumption level of each computing node in real time, and to record the execution progress of each task slice.

[0111] An event response unit, connected to a state tracking unit, is used to generate a rescheduling strategy, including task migration, task replication, or task restart, based on the execution progress and remaining computational capacity of the task shards when an offline node, performance degradation, or abnormal power consumption event is detected.

[0112] The elastic rescheduling unit, connected to the event response unit, is used to reselect target nodes for the affected task fragments according to the rescheduling strategy and the node capability profile of the currently available computing nodes, and to trigger the transmission and state synchronization of the task fragments.

[0113] In this embodiment, the state tracking unit monitors online status by maintaining a periodic heartbeat protocol between the node and the master control unit. It obtains the real-time load rate and power consumption level of the node by reading performance counters provided by the operating system, and calculates the execution progress by recording the start timestamp of the assigned task shard and the amount of work confirmed to be completed. The event response unit analyzes and judges the above monitoring data based on preset thresholds: when the number of heartbeat losses exceeds a set number, the node is determined to be offline; when the load rate is continuously higher than a preset load rate threshold or power consumption increases abnormally sharply, the node's performance is determined to be degraded. Based on the execution progress of the affected task shards, this unit generates a migration strategy if the progress is low, a replication strategy if the progress is high, and a restart strategy if the task cannot continue. After receiving the rescheduling strategy, the elastic rescheduling unit, based on the updated available node profile, quickly selects a target node for the affected task shards using heuristic methods such as greedy algorithms, and transmits the task shards and their context state to the new node via network communication protocols, while simultaneously notifying relevant parties to update the task state mapping.

[0114] The set number of lost connections is the consecutive loss threshold for determining node offline in the heartbeat protocol. Its value depends on the expected stability of the network environment and the system's real-time requirements for fault diagnosis, and is typically set between 2 and 5. In this embodiment, it is set to 3, which can effectively filter out brief network jitter interference while responding quickly to substantial node offline faults. The preset load rate threshold is the load level limit for determining node performance degradation. Its value depends on the design margin of the node's hardware performance and the system's sensitivity to computational task response latency, and is typically set between 70% and 90%. In this embodiment, it is set to 85%, which strikes a balance between fully utilizing node computing resources and preventing task execution timeouts or failures due to overload.

[0115] By establishing a closed-loop control chain of "monitoring-diagnosis-recovery," the module deeply correlates dynamic signals such as real-time load, power consumption, and network heartbeats of nodes with logical states such as task fragment execution progress and remaining computational load. The persistence of the heartbeat signal directly reflects the node's network reachability and basic operational status; abnormal changes in load rate and power consumption reveal chip-level thermal and electrical performance bottlenecks; and the comparison between task execution progress and remaining workload provides crucial evidence for cost-benefit analysis in decision-making. Based on these coupling relationships, when a node is detected to be offline or experiencing performance degradation, the module can automatically and intelligently select differentiated recovery strategies such as migration, replication, or restart based on the proportion of tasks already completed. Essentially, this involves a quantitative trade-off between the sunk cost of computational tasks, the overhead of re-execution, and the timeliness of task completion. Subsequently, the elastic rescheduling unit, based on the latest global resource view, quickly matches suitable new nodes for affected tasks and synchronizes their execution context, thereby transforming dynamic disturbances into a controlled, optimal, or near-optimal rescheduling process. This allows the entire system to proactively manage uncertainty instead of passively accepting node changes, ultimately achieving a significant improvement in task success rate and an effective reduction in overall service downtime in a frequently changing local network environment.

[0116] Specifically, the batch processing unit includes:

[0117] The window adaptive subunit is used to dynamically adjust the preset batch processing duration and the period and length of the preset communication time window based on the remaining power of the computing node, the urgency of the task queue, and the network channel quality.

[0118] The conflict avoidance subunit is used to allocate staggered communication time windows to each computing node when several computing nodes share the same master node, using a time-division multiple access mechanism.

[0119] In this embodiment, the window adaptive subunit dynamically adjusts parameters according to a set of predefined rules by monitoring the battery level of the computing nodes, analyzing the preset deadlines of the fragments to be executed in the task queue, and evaluating the signal strength and bit error rate of the current wireless channel: when the battery is low, it tends to extend the batch processing time to merge more tasks, thereby reducing the activation frequency of the communication window; when there are high-urgency tasks in the queue, it shortens the batch processing time and opens the communication window in advance to prioritize their low latency; when the network quality is poor, it may extend the batch processing time and appropriately increase the communication window length to improve the success rate of a single transmission. The conflict avoidance subunit is centrally coordinated by the master node. Based on the time division multiple access principle, it calculates and allocates a unique and non-overlapping communication time slot for each associated computing node, and broadcasts a scheduling table containing the start time and duration of the time slot to all nodes; each computing node strictly follows the scheduling table and only wakes up the wireless module to communicate in its allocated time slot, thereby achieving conflict-free and orderly uplink communication between multiple nodes and a single master node. The implementation of the above rules and the time-division scheduling mechanism can be completed through conventional state judgment, arithmetic calculation and communication protocol interaction.

[0120] By coordinating the batch processing duration and communication window, the three core physical and logical constraints—remaining battery power, task urgency, and channel quality—are mapped to a joint optimization of system energy consumption, latency, and reliability. Remaining battery power, acting as an energy budget, directly limits the total number of times and duration the wireless communication module can be activated. Task urgency defines the lower limit of the service's latency tolerance, exerting reverse pressure on the cumulative batch processing time. Channel quality determines the amount of data reliably transmitted per unit time, affecting the efficiency of a single communication window. The window adaptive subunit dynamically balances these mutually constraining parameters based on preset rules: extending batch processing to save communication energy during low battery periods, shortening batch processing for low latency during high-urgency tasks, and increasing the window to compensate for transmission reliability when network quality is poor. Essentially, this solves a real-time approximate solution to a multi-objective optimization problem. Simultaneously, the conflict avoidance subunit introduces a time-division multiple access (TDMA) mechanism, transforming the random competition for a shared physical channel among multiple nodes into a deterministic, conflict-free time-series scheduling, fundamentally avoiding retransmissions and energy waste caused by packet collisions. The combination of these two elements enables the system to maintain an optimal balance between communication energy consumption, task latency, and transmission success rate under dynamically changing resource and demand conditions.

[0121] Please see Figure 4 As shown, this is a flowchart of the localized low-power computing power aggregation and scheduling method based on heterogeneous SoC in this embodiment. Furthermore, this embodiment also provides a localized low-power computing power aggregation and scheduling method based on heterogeneous SoC, including:

[0122] A local communication network is established between the master node and each computing node, wherein the computing nodes include heterogeneous SoC or MCU terminals with different computing power capabilities.

[0123] The computing nodes connected to the local communication network are discovered and authenticated, and the computing power, real-time power consumption status, power information and network communication status of each computing node are collected in real time to generate a node capability profile.

[0124] Based on the node capability profile and the task requirements of the tasks to be executed, optimization calculations are performed with system energy consumption as the core constraint to generate task scheduling decisions, and the task fragments after the task to be executed are allocated to the selected computing nodes.

[0125] The assigned tasks are executed on each of the computing nodes, and execution results are generated.

[0126] All execution results are summarized and merged to generate the final task result;

[0127] The system monitors the operating status of each computing node within the local communication network, and when a computing node is detected to be joining, going offline, or experiencing performance abnormalities, it triggers rescheduling of tasks that have not yet been completed based on the execution status of each task to be executed.

[0128] By constructing a local dynamic network between the master node and heterogeneous computing nodes, the computing power, power consumption, battery level, and network status of each node are monitored in real time to form a capability profile. Based on this profile, tasks are decomposed and intelligently scheduled with the core objective of minimizing total system energy consumption. A rescheduling mechanism is also introduced to handle dynamic node addition, offline status, and anomalies. This method enables the efficient aggregation and collaboration of distributed, battery-powered heterogeneous computing resources in a local weak network environment. Ultimately, it significantly reduces overall system energy consumption, extends device battery life, and effectively improves the success rate of computing tasks and system robustness in the dynamic network.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A localized low-power computing power aggregation and scheduling system based on a heterogeneous SoC, characterized in that, Comprise: a local networking module configured to establish a local communication network between a master node and each computing node, wherein the computing node comprises a heterogeneous SoC or MCU terminal with different computing power capabilities; a modeling module configured to discover and authenticate the computing nodes accessing the local communication network, and collect the computing power capabilities, real-time power consumption states, power information, and network communication states of each computing node in real time to generate a node capability profile; a scheduling and distribution module configured to decompose a to-be-executed task into a plurality of task fragments according to the node capability profile and the task requirements of the to-be-executed task, perform optimization calculation with system energy consumption as a core constraint, generate a task scheduling decision, and distribute the task fragments to selected computing nodes; an execution module configured to execute the distributed to-be-executed task on each computing node and generate an execution result; a result aggregation module configured to aggregate and merge all the execution results to generate a final task result; an optimized scheduling module configured to monitor the running states of each computing node in the local communication network, and trigger re-scheduling of a to-be-executed task that has not been executed to completion based on the execution state of each to-be-executed task when detecting that a computing node joins, goes offline, or has a performance anomaly; the scheduling and distribution module comprises: a task analysis and fragmentation unit configured to receive a to-be-executed task, analyze it into a plurality of task fragments with dependency relationships according to the task type, and calculate the computation amount and data throughput requirement of each task fragment; an energy consumption cost determination unit configured to determine the total execution cost of a pair formed by each task fragment and the corresponding computing node according to a preset cost calculation function based on the node capability profile and in combination with the computation amount and data throughput requirement of the task fragment; a scheduling and distribution unit configured to solve a task distribution scheme under the constraint condition of meeting the overall completion time limit of the task with the minimization of all the total execution costs as a core optimization objective, generate a scheduling decision, and distribute the task fragments to selected computing nodes according to the scheduling decision; the task analysis and fragmentation unit comprises: a task type judgment subunit configured to identify the to-be-executed task as a general-purpose computing task or a lightweight artificial intelligence inference task that can be parallelized; a general-purpose computing fragmentation subunit configured to, when the to-be-executed task is identified as the general-purpose computing task, divide it according to its process parallelism to obtain a plurality of task fragments, and estimate the number of processor operations required by each task fragment as the computation amount and the size of its input data and output data as the data throughput requirement; an AI model fragmentation subunit configured to, when the to-be-executed task is identified as the lightweight artificial intelligence inference task, divide it according to the structural dependency relationship of its model computation graph to obtain a plurality of task fragments, and estimate its computation amount based on the number of floating-point operations and estimate its data throughput requirement based on the size of the inter-layer feature map.

2. The heterogeneous SoC-based localized low-power computing power aggregation and scheduling system according to claim 1, wherein, the local networking module comprises: a networking protocol unit configured to run a protocol stack compliant with IEEE 802.11ah standard to build a wireless peer-to-peer network between the master node and each of the computing nodes; a link management unit configured to maintain network connection status with each of the computing nodes and update network topology when a computing node joins or leaves; 3. The heterogeneous SoC-based localized low-power computing power aggregation and scheduling system according to claim 1, wherein, the modeling module comprises: a discovery and probe timing unit configured to broadcast a discovery beacon at a preset period and respond to an access request from the computing node; a multi-dimensional awareness unit configured to actively probe real-time information of the computing node after the computing node accesses, the real-time information including CPU / NPU computing power frequency, current available memory, battery remaining capacity, chip instantaneous power consumption, network round-trip delay to the master node, and packet loss rate; an image modeling unit configured to fuse the real-time information with static hardware configuration information of the computing node to generate the node capability image.

4. The heterogeneous SoC-based localized low-power computing power aggregation and scheduling system according to claim 1, wherein, the energy consumption cost determination unit comprises: a cost parameter acquisition unit configured to calculate wireless communication power consumption, active power consumption of the computing node, overhead of waking up from sleep state, estimated execution duration of the task fragment, and required duration of network transmission according to the node capability image; a cost calculation sub-unit configured to input the wireless communication power consumption, the active power consumption of the computing node, the overhead of waking up from sleep state, the estimated execution duration, and the required duration of network transmission into the preset cost calculation function to calculate the total execution cost.

5. The heterogeneous SoC-based localized low-power computing power aggregation and scheduling system according to claim 1, wherein, the execution module comprises: a task container unit configured to build an isolated runtime environment on the computing node for receiving, loading, and executing code and data of the task fragment; a batch processing unit configured to maintain a task queue locally on the computing node, combine all task fragments received within a preset batch processing duration and belonging to the same or different to-be-executed tasks into an execution batch, and sequentially inject each task fragment in the execution batch into the isolated runtime environment to trigger loading and execution, and control the computing node to enable the wireless communication module only within a preset communication time window to complete batch data reception and result return required by the execution batch.

6. The heterogeneous SoC-based localized low-power computing power aggregation and scheduling system according to claim 5, wherein, the optimization scheduling module comprises: a state tracking unit configured to monitor online state, load rate, and power consumption level of each of the computing nodes in real time, and record execution progress of each of the task fragments; an event response unit configured to generate a rescheduling strategy including task migration, task replication, or task restart according to execution progress and computing remaining amount of the task fragments when detecting a node offline, performance degradation, or power consumption anomaly event; a flexible rescheduling unit configured to reselect a target node for the affected task fragment by a preset rescheduling algorithm according to the rescheduling strategy and the node capability image of the currently available computing node, and trigger transmission and state synchronization of the task fragment.

7. The heterogeneous SoC-based localized low-power computing power aggregation and scheduling system according to claim 5, wherein, the batch processing unit comprises: a window self-adaptive subunit configured to dynamically adjust the preset batch processing time length and the period and length of the preset communication time window according to the remaining power of the computing node, the urgency of the task queue, and the network channel quality; a conflict avoidance subunit configured to allocate communication time windows for each computing node in a time division multiple access mechanism when a plurality of computing nodes share the same master node.

8. A method for local low-power computing power aggregation scheduling based on a heterogeneous SoC, applied to the local low-power computing power aggregation scheduling system based on a heterogeneous SoC in any one of claims 1-7, characterized in that, The method comprises: establishing a local communication network between the master node and each computing node, wherein the computing node comprises a heterogeneous SoC or MCU terminal with different computing power capabilities; discovering and authenticating the computing nodes accessing the local communication network, and collecting the computing power capability, real-time power consumption state, power information, and network communication state of each computing node in real time to generate a node capability profile; decomposing the to-be-executed task into a plurality of task fragments according to the node capability profile and the task requirement of the to-be-executed task, performing optimization calculation with system energy consumption as the core constraint to generate a task scheduling decision, and allocating the task fragments to selected computing nodes; executing the allocated to-be-executed task on each computing node to generate an execution result; summarizing and merging all the execution results to generate a final task result; monitoring the running state of each computing node in the local communication network, and triggering re-scheduling of the to-be-executed task that has not been executed to completion based on the execution state of each to-be-executed task when a computing node is detected to be joined, offline, or have a performance anomaly.

Images