A distributed computing power sharing method and system based on a mobile terminal cluster

Abstract

The application relates to the field of distributed computing, and particularly discloses a distributed computing power sharing method and system based on a mobile terminal cluster, which relies on a system architecture comprising a cloud / local computing power scheduling platform, an encrypted communication channel and a mobile terminal node cluster, and comprises the following steps: step one: obtaining an original computing task to be processed, decomposing the original computing task into a plurality of subtask units carrying input parameters and execution environments through a task analysis module, and establishing a directed acyclic graph topology; step two: obtaining device state data of registered mobile terminal nodes through a node management module, and performing multi-layer access evaluation on the device state data based on a composite access decision engine; the technical scheme can solve the problems of large-scale mobile node dynamic management, abnormal node fast fault tolerance, heterogeneous subtask result synchronization alignment and local area network computing power data complete isolation.

Description

Technical Field

[0001] This invention relates to the field of distributed computing, and in particular to a method and system for sharing distributed computing power based on mobile terminal clusters. Background Technology

[0002] The number of mobile devices in use worldwide has now exceeded several billion. The vast majority of smartphones, tablets, and other devices are equipped with high-performance CPUs, GPUs, and dedicated NPU computing units. A single device can achieve a peak AI computing power of 30-50 TOPS, and these computing resources are completely idle for most of the time during daily use. The theoretical scale of this idle computing power far exceeds the total computing power of existing centralized data centers. Existing distributed computing power aggregation solutions, such as the BOINC distributed computing platform, general edge computing scheduling frameworks, and federated learning systems, cannot directly adapt to the special operational constraints of mobile devices: the BOINC platform is natively designed for PCs with stable power supplies and heat dissipation, and is not adapted for the power consumption and heat dissipation characteristics of mobile devices; the federated learning solution focuses on privacy protection scenarios where data does not leave the device, and is not designed for general computing power sharing.

[0003] Looking further, existing computing power scheduling solutions for mobile terminals still have several unresolved substantial defects: First, they lack access experience protection mechanisms for mobile terminals, failing to consider physical constraints such as limited battery life, passive heat dissipation leading to overheating and frequency throttling, frequent network environment switching, and random user behavior. This makes computing power sharing tasks highly susceptible to interfering with the normal user experience. Second, they lack lightweight runtime fault-tolerant monitoring mechanisms. Mobile nodes are prone to disconnection due to user device unlocking, process reclaiming by the system, network fluctuations, etc. Existing solutions cannot quickly detect node anomalies, leading to long-term blocking of subtasks and significantly reducing overall computing efficiency. Third, the computing power performance of heterogeneous mobile terminals varies greatly. Existing scheduling solutions lack a time-series alignment mechanism for results returned by multiple nodes, resulting in time differences of several times between the returned results, failing to guarantee result consistency. Fourth, all existing solutions require nodes to connect to the public network and communicate with cloud servers, which cannot support self-organizing network computing power scheduling in LAN scenarios. This incurs high public network bandwidth costs and fails to meet the computing needs of enterprises and communities for complete physical isolation of sensitive data.

[0004] Therefore, there is an urgent need for a distributed computing power sharing method and system based on mobile terminal clusters that can securely, stably, and with low latency aggregate dispersed idle mobile computing power without interfering with the normal user experience, and solve the problems of dynamic management of large-scale mobile nodes, rapid fault tolerance of abnormal nodes, synchronization and alignment of heterogeneous subtask results, and complete isolation of local area network computing power data. Summary of the Invention

[0005] This invention provides a distributed computing power sharing method based on mobile terminal clusters, which can safely, stably, and with low latency aggregate dispersed idle mobile computing power without interfering with the normal user experience. It solves the problems of dynamic management of large-scale mobile nodes, rapid fault tolerance of abnormal nodes, synchronization and alignment of heterogeneous subtask results, and complete isolation of local area network computing power data.

[0006] To solve the above-mentioned technical problems, this application provides the following technical solution: A distributed computing power sharing method based on mobile terminal clusters, relying on a system architecture that includes a scheduling platform for cloud / local computing power, an encrypted communication channel, and a mobile terminal node cluster, includes the following steps: Step 1: Obtain the original computation task to be processed, and decompose the original computation task into multiple sub-task units carrying input parameters and execution environment through the task parsing module, and establish a directed acyclic graph topology; Step 2: Obtain the device status data of the registered mobile terminal nodes through the node management module, and perform multi-level access evaluation on the device status data based on the composite access decision engine to screen out mobile terminal nodes that meet the access conditions for computing power sharing and build an executable computing power pool. Step 3: Input the computational requirements of the sub-task unit, the computing power performance, historical reputation, and estimated latency of the mobile terminal nodes in the executable computing power pool into the scheduling matching degree function to calculate the scheduling matching degree score between each mobile terminal node and each sub-task unit. Step 4: Based on the scheduling matching score, the sub-task units are distributed to the corresponding target mobile terminal nodes for computing power sharing; and during the calculation, the heartbeat monitoring module sends heartbeat detection packets to the target mobile terminal nodes at a preset half-cycle. The target mobile terminal node is determined to be abnormally disconnected based on whether a heartbeat response packet is received within a consecutive preset cycle. If an abnormal disconnection is determined, the backup distribution logic is triggered to redistribute the unfinished sub-task units to the backup mobile terminal nodes. Step 5: Open the synchronization waiting time window, receive the subtask calculation result packets fed back by each mobile terminal node, perform timing coordination alignment and consistency verification on the received subtask calculation result packets within the synchronization waiting time window, and reorder and concatenate them according to the absolute sequence number of the subtasks to output the final calculation result of the original calculation task.

[0007] Furthermore, the step of performing multi-layer access evaluation on the device status data based on the composite access decision engine, filtering out mobile terminal nodes that meet the computing power sharing admission conditions, and constructing an executable computing power pool specifically includes: obtaining the current network connection status, current remaining battery power, and foreground application running status of the mobile terminal node; if the current network connection status is a preset network, the current remaining battery power is greater than a preset battery power threshold, and there are no sensitive applications running in the foreground, then it is determined that the basic conditions for admission are met; after meeting the basic conditions for admission, the current charging status and historical power consumption rate of the mobile terminal node are obtained, and the current charging status and historical power consumption rate are jointly evaluated by a joint condition discrimination formula, wherein the historical power consumption rate is obtained by collecting the battery discharge current value in multiple consecutive sampling periods and calculating the discharge current difference; if the current slope corresponding to the discharge current difference is continuously negative in multiple consecutive sampling periods, it is determined that the user usage intensity of the mobile terminal node is weakening, the joint condition discrimination formula outputs true, and it is determined that the charging and energy consumption admission conditions are met.

[0008] Furthermore, the step of performing multi-layer access evaluation on the device status data based on the composite access decision engine, selecting mobile terminal nodes that meet the access conditions for computing power sharing, and constructing an executable computing power pool further includes: real-time monitoring of the processor core junction temperature of the mobile terminal node, and executing corresponding thermal state management strategies according to the temperature range of the processor core junction temperature. Specifically, when the processor core junction temperature is less than a first temperature threshold, the processor is controlled to execute calculations at full frequency; when the processor core junction temperature is greater than or equal to the first temperature threshold and less than a second temperature threshold, the processor is controlled to reduce its execution frequency or suspend receiving new subtasks until the processor core junction temperature drops below the temperature difference; when the processor core junction temperature is greater than or equal to the second temperature threshold, overheat protection is triggered, security sandbox data is cleared, and the computing power sharing is exited; the user calendar habits and sleep patterns of the mobile terminal node are obtained, the user's usage status within a future time window is predicted, and subtask units with corresponding computing power scale are allocated to the mobile terminal node according to the prediction results.

[0009] Furthermore, the step of performing multi-layer access evaluation on the device status data based on the composite access decision engine, selecting mobile terminal nodes that meet the access conditions for computing power sharing, and constructing an executable computing power pool also includes a dynamic exit trigger mechanism: during the execution of sub-task units by mobile terminal nodes, after obtaining user authorization, the system monitors user interaction events, physical motion status, power level changes, and charging connection status of the mobile terminal nodes in real time; when it detects that the user unlocks the screen to start a foreground application, the acceleration change detected by the accelerometer exceeds a preset acceleration threshold, the current remaining power level drops to the user-defined minimum participation power level threshold, or the charger is unplugged and the current remaining power level is lower than the set high power level limit, the system immediately releases the computing resource control of the mobile terminal node, interrupts the currently executing sub-task unit, and sends the unfinished sub-task unit back to the scheduling platform so that the scheduling platform can redistribute the unfinished sub-task unit to other backup mobile terminal nodes.

[0010] Furthermore, the scheduling matching degree function is calculated based on the CPU and GPU / NPU computational requirements of the sub-task unit, as well as the CPU computing power score, NPU / GPU computing power score, historical reputation score, and estimated data transmission latency of the mobile terminal node. The weight values ​​corresponding to each parameter in the scheduling matching degree function satisfy normalization constraints. When selecting a node using the scheduling matching degree function, preset scheduling constraints must be met. These preset constraints include: the current battery percentage of the mobile terminal node is greater than a user-defined minimum participation battery threshold, and the estimated total execution time of the mobile terminal node for the corresponding sub-task unit is less than or equal to the maximum allowable latency of that sub-task unit. The estimated total execution time is calculated based on the computational requirements of the sub-task unit, the computing power score of the mobile terminal node, and the estimated data transmission latency.

[0011] Furthermore, the weight values ​​corresponding to each parameter in the scheduling matching function are dynamically adjusted according to the load status of the scheduling platform and the fluctuation of node reputation. Specifically, when the scheduling platform detects that the backlog of overall computing tasks is lower than a preset low load threshold, it automatically increases the transmission efficiency weight corresponding to the estimated data transmission latency and correspondingly decreases the computing power score weight and reputation score weight to reduce network transmission time. When it detects that the historical reputation score fluctuation of a preset batch of mobile terminal nodes in the executable computing power pool exceeds a preset fluctuation threshold, it automatically increases the reputation score weight corresponding to the historical reputation score and correspondingly decreases the transmission efficiency weight and computing power score weight to improve the sensitivity to prevent disconnection and avoid assigning tasks to mobile terminal nodes with unstable reputation.

[0012] Furthermore, if the system architecture is in a dedicated networking mode, the method also includes a self-organizing network step for the local dedicated computing power center: the scheduling platform is deployed as an independent background process on a resident physical device within the local area network; a description file carrying the cluster service set identifier (SSID) and task type is broadcast or multicast within the local area network through the scheduling platform; mobile terminal nodes within the local area network listen to a specific port through a client, and register to access the scheduling platform after receiving and confirming the description file, thus constructing a localized dedicated computing power cluster; the input data of the original computing task is sliced ​​by the edge server within the local area network and distributed to the mobile terminal nodes in the dedicated computing power cluster via a wireless channel based on the local area network unicast protocol; the computing process and result aggregation are all completed in a closed loop within the local area network to achieve physical isolation of sensitive data within the local area network.

[0013] Furthermore, the mobile terminal node executes the sub-task unit through a built-in lightweight security sandbox. The lightweight security sandbox establishes an independent user identifier (UID) isolation mechanism based on the operating system's application layer and restricts the lightweight security sandbox's access permissions to the mobile terminal node's local file system and hardware sensors to prevent computing tasks from reading user privacy data. When the sub-task unit completes execution or triggers an abnormal exit, the lightweight security sandbox automatically erases and reclaims temporary data, cached data, and memory usage generated during execution. At the same time, the lightweight security sandbox sets upper limits on the CPU usage and memory usage of the computing process during operation to prevent computing tasks from excessively consuming system resources, which could cause system lag or interface response delays on the mobile terminal node.

[0014] Furthermore, after outputting the final calculation result of the original computing task, the method further includes a computing power measurement and revenue distribution step: Through the computing power measurement and settlement module, the actual computing time, hardware resource consumption, and number of successfully uploaded result packets of each mobile terminal node during the execution of the sub-task unit are collected to calculate the standard computing power unit (SCU) contribution value of each mobile terminal node; based on the standard computing power unit (SCU) contribution value and the corresponding node historical reputation score, the revenue distribution ratio of each mobile terminal node is calculated; if in a dedicated network mode, the standard computing power unit (SCU) contribution value is converted into corresponding enterprise work points or community shared rights, and recorded and settled within the local area network; if in a public network scheduling mode, tokens or virtual asset rewards are issued to the corresponding mobile terminal nodes based on the revenue distribution ratio. In the step of the heartbeat monitoring module sending heartbeat probe data packets to the target node at a fixed time period, the fixed time period is dynamically adjusted according to the network connection quality and historical reputation score of the target node. Specifically, it includes: periodically acquiring the network latency and packet loss rate of the target node; if the network latency is greater than a preset latency threshold or the packet loss rate is greater than a preset packet loss rate threshold, then the fixed time period is extended to a first preset period to reduce false judgments of heartbeat response timeouts caused by instantaneous network fluctuations; if the historical reputation score of the target node is lower than a preset reputation threshold, then the fixed time period is shortened to a second preset period to increase the monitoring frequency of the offline status of low-reputation nodes, wherein the first preset period is greater than the fixed time period, and the second preset period is less than the fixed time period.

[0015] A distributed computing power sharing system based on mobile terminal clusters includes: The computing power demander interface layer provides task submission, status query, and result retrieval interfaces, enabling demanders to submit computing tasks, query subtask distribution status, and retrieve the final aggregated results. The cloud / local computing power scheduling platform is used to receive raw computing tasks, parse and split the raw computing tasks into multiple sub-task units and establish a directed acyclic graph topology, manage registered mobile terminal nodes, distribute the sub-task units to the optimal available mobile terminal nodes based on the scheduling matching degree function, perform runtime monitoring and abnormal rescheduling through the heartbeat monitoring module, and perform collaborative alignment and splicing of received computing results within the synchronization waiting time window; An encrypted communication layer is used to establish an encrypted communication channel between the cloud / local computing power scheduling platform and the mobile terminal node for data transfer.

[0016] The system described above already provides the basic server-side functionality. By adding mobile terminal nodes as users, all computing power sharing capabilities can be achieved. Typically, this is done through a mobile terminal node cluster. The cluster consists of multiple independent mobile terminal nodes, each equipped with a computing power contribution client. This client uses a composite access decision engine to perform multi-layered access evaluation to determine whether to join the executable computing power pool. It then executes distributed sub-task units within a lightweight security sandbox and sends heartbeat response packets and computation result packets back to the cloud / local computing power scheduling platform.

[0017] The principle and effect of the solution are as follows: The present invention decomposes the original computing task into multiple sub-task units carrying input parameters and execution environment through the task parsing module in step one, and establishes a directed acyclic graph topology. Then, through steps three and four, the sub-task units are distributed to an executable computing power pool composed of massive heterogeneous mobile terminal nodes for distributed computing.

[0018] This invention utilizes directed acyclic graphs (DAGs) to decouple and topologically arrange complex computing tasks, transforming the computing power requirements that originally required centralized processing on high-configuration servers into lightweight subtask flows that can be executed concurrently by multiple mobile devices; at the same time, it utilizes the existing hardware resources (CPU, GPU, NPU) of mobile terminals for on-site computing.

[0019] Ultimately, this achieves the goal of eliminating the need to build high-cost centralized data centers or deploy dedicated edge servers, significantly reducing the construction cycle and hardware procurement costs of computing infrastructure. Furthermore, because computing tasks are distributed to various idle mobile terminals, it enables elastic aggregation and on-demand scheduling of computing resources, significantly reducing the energy consumption pressure caused by centralized power supply and heat dissipation, and realizing low-carbon, low-cost computing power sharing.

[0020] The present invention uses a composite access decision engine in step two to perform multi-level access evaluation on the device status data of registered mobile terminal nodes, and dynamically selects mobile terminal nodes that meet the access conditions for computing power sharing to build an executable computing power pool.

[0021] This means that through a multi-level composite decision-making mechanism, before and during device access, cross-layer collaborative evaluation is conducted on physical constraints such as network status, remaining power, charging trend, processor junction temperature, and foreground interaction status to ensure that computing tasks are only undertaken when the device is in an "unobtrusive" state where it is idle, has sufficient power, is connected to a preset network, and is at a suitable temperature.

[0022] It effectively avoids problems such as rapid battery drain, device overheating and frequency reduction, system lag or foreground application response delay caused by shared computing power on mobile terminals. It achieves secure and smooth access to mobile terminal computing power without interfering with the user's normal experience and daily use.

[0023] In step four of this invention, during the execution of a subtask, a heartbeat detection packet is sent to the target mobile terminal node at a preset half-cycle by a heartbeat monitoring module. The target mobile terminal node is determined to be abnormally disconnected based on whether a heartbeat response packet is received within a consecutive preset cycle. After the disconnection is determined, the backup distribution logic is immediately triggered to redistribute the unfinished subtask unit to the backup mobile terminal node.

[0024] Because it adopts a lightweight half-cycle heartbeat detection mechanism, it can track the active status of mobile terminals in real time with extremely low network bandwidth overhead. Once it detects that a node has lost contact due to network switching, process killing or user unlocking, it immediately reschedules the task through the backup distribution logic, so that the unfinished subtasks can be quickly resumed on the backup node.

[0025] It solves the fault tolerance problem caused by the high dynamism and susceptibility of mobile terminals as computing nodes, avoids the situation where the entire DAG topology task is blocked for a long time due to the failure of a single node, and significantly improves the overall execution efficiency and robustness of distributed computing.

[0026] In step five of this invention, a synchronization waiting time window is opened to receive subtask calculation result packets fed back by each mobile terminal node. Within the synchronization waiting time window, the received subtask calculation result packets are subjected to timing coordination alignment and consistency verification, and then reordered and concatenated according to the absolute sequence number of the subtask.

[0027] By introducing a synchronous waiting time window mechanism, a buffer is established for heterogeneous nodes with huge performance differences; the out-of-order result packets are rearranged in ascending order using the absolute sequence number (Sequence ID) of the subtask, and the data integrity and authenticity are determined by combining a consistent hash verification algorithm.

[0028] It solves the problem of result timing disorder caused by the difference in computing performance of heterogeneous hardware, ensures the output order and logical correctness of distributed computing results, and prevents the risk of malicious nodes tampering with results through consistency verification, thus ensuring the security and reliability of the final output results. Attached Figure Description

[0029] Figure 1 A flowchart illustrating a distributed computing power sharing method based on mobile terminal clusters; Figure 2 This is a schematic diagram of the architecture of a distributed computing power sharing system based on mobile terminal clusters. Detailed Implementation

[0030] The following detailed description illustrates the specific implementation method: A distributed computing power sharing method based on mobile terminal clusters (e.g.) Figure 1 As shown), relying on a system architecture that includes a scheduling platform with cloud / local computing power, encrypted communication channels, and mobile terminal node clusters, the following steps are included: Step 1: Obtain the original computation task to be processed, and decompose the original computation task into multiple sub-task units carrying input parameters and execution environment through the task parsing module, and establish a directed acyclic graph topology; Step 2: Obtain the device status data of the registered mobile terminal nodes through the node management module, and perform multi-level access evaluation on the device status data based on the composite access decision engine to screen out mobile terminal nodes that meet the access conditions for computing power sharing and build an executable computing power pool. Step 3: Input the computational requirements of the sub-task unit, the computing power performance, historical reputation, and estimated latency of the mobile terminal nodes in the executable computing power pool into the scheduling matching degree function to calculate the scheduling matching degree score between each mobile terminal node and each sub-task unit. Step 4: Based on the scheduling matching score, the sub-task unit is distributed to the corresponding target mobile terminal node for computing power sharing (after authorization or a request is made by the user of the mobile terminal); and during the calculation, the heartbeat monitoring module sends heartbeat detection packets to the target mobile terminal node at a preset half-cycle. The target mobile terminal node is determined to be abnormally disconnected based on whether a heartbeat response packet is received within a consecutive preset cycle. If an abnormal disconnection is determined, the backup distribution logic is triggered to redistribute the unfinished sub-task unit to the backup mobile terminal node. Step 5: Open the synchronization waiting time window, receive the subtask calculation result packets fed back by each mobile terminal node, perform timing coordination alignment and consistency verification on the received subtask calculation result packets within the synchronization waiting time window, and reorder and concatenate them according to the absolute sequence number of the subtasks to output the final calculation result of the original calculation task.

[0031] In practical use, during the task parsing and topology generation stage, the original computing task to be processed is obtained. The task parsing module decomposes the original computing task into multiple sub-task units carrying input parameters and execution environment, and establishes a directed acyclic graph topology.

[0032] In this embodiment, the "original computing task" refers to a complex computing task submitted by the computing power requester through an external interface, whose computational load exceeds the processing limit of a single ordinary device. The "task parsing module" is the core logical unit deployed on the cloud / local computing power scheduling platform, used for structured task decomposition. The "subtask unit" is the smallest computational closed loop that can be independently distributed and executed after decomposition. Each subtask unit is assigned a unique "subtask absolute sequence number (Sequence ID)" to ensure accurate rearrangement after distributed out-of-order execution. The "directed acyclic graph (DAG) topology" is used to characterize the execution dependencies between subtask units, ensuring that dependent subtasks are executed sequentially according to the topological order.

[0033] For task parsing and topology generation, this embodiment provides a general task decomposition process: Receive the original computation task and use the parser to construct the abstract syntax tree (AST) or computation dependency graph of the task; Identify acyclic data flows in the computation graph and determine the input-output dependencies between various computation nodes; Based on the independence of computing nodes, nodes without dependencies or whose dependencies are already satisfied are divided into independent subtask units; Construct a directed acyclic graph topology using subtask units as vertices and data dependencies as directed edges.

[0034] In this embodiment, the original computational task is a distributed inference task of a Large Language Model (LLM). Upon receiving the inference task, the task parsing module decomposes it layer-wise, generating 32 sub-task units. Each sub-task unit is encapsulated as an independent task package, carrying the weight parameter indexes (input parameters) required for that layer's inference, the required CUDA / PyTorch runtime library version (execution environment), and the absolute sub-task number from 001 to 032. Next, the task parsing module establishes the corresponding DAG topology, where sub-task 002 depends on the output activation value of sub-task 001, and so on.

[0035] During the node status access assessment and computing power pool construction phase, the device status data of registered mobile terminal nodes is obtained through the node management module, and multi-level access assessment is performed on the device status data based on the composite access decision engine to screen out mobile terminal nodes that meet the computing power sharing admission conditions and construct an executable computing power pool.

[0036] The "node management module" maintains a dynamic database of online mobile terminals, recording the hardware configuration and real-time status of the devices. The "composite access decision engine" is a multi-dimensional admission evaluation algorithm running locally on the mobile terminal or on a scheduling platform. Through multi-level physical constraint filtering, it ensures that the computing power sharing process is "completely seamless" for mobile terminal users. The "executable computing power pool" refers to the set of active mobile terminals that currently meet all admission conditions and are ready to accept task assignments at any time.

[0037] The general construction process of the composite access decision engine is as follows: The mobile terminal node sends a registration request to the platform, and the platform assigns a unique node identifier to it and establishes a status listening channel. Periodically or before task distribution, collect the mobile terminal's hardware configuration (CPU / GPU / NPU computing power score) and real-time physical status (network, battery level, temperature, foreground status). The physical state is sequentially input into the multi-layer evaluation module of the composite access decision engine to calculate whether the admission requirements are met. Nodes that pass all levels of evaluation will be added to the executable computing power pool, while nodes that fail will either remain on the listening list or be removed from the computing power pool.

[0038] In this embodiment, when the mobile terminal node first launches the computing power sharing client, it registers with the cloud / local computing power scheduling platform 2. The node management module 22 records the NPU computing power score of each node in the database.

[0039] The multi-layer access evaluation of the device status data based on the composite access decision engine includes: obtaining the current network connection status, current remaining battery power, and foreground application running status of the mobile terminal node; if the current network connection status is a preset network, the current remaining battery power is greater than a preset battery power threshold, and there are no sensitive applications running in the foreground, then it is determined that the basic conditions for access are met.

[0040] In this embodiment, the composite access decision engine first performs a first-layer basic condition admission assessment. It obtains the current network connection status of the mobile terminal, determines whether it is a Wi-Fi connection (in the local dedicated computing center self-organizing network mode, it determines whether it is connected to the specified local area network SSID); and obtains the current remaining battery power. Determine whether it exceeds the preset power threshold. (In this embodiment, the percentage is set to 60%); Obtain the running status of the foreground application and determine whether there are sensitive applications such as phone calls, video conferencing, or large-scale games running in the foreground. Only when the network is Wi-Fi, the battery level is greater than 60%, and there are no sensitive applications in the foreground, is the basic condition for admission met, and the system proceeds to the next level of evaluation.

[0041] The multi-layer access evaluation of the device status data based on the composite access decision engine further includes: obtaining the current charging status and historical power consumption rate of the mobile terminal node, and jointly evaluating the current charging status and historical power consumption rate through a joint conditional discrimination formula. If the output of the joint conditional discrimination formula is true, it is determined that the charging and energy consumption access conditions are met; the joint conditional discrimination formula is: ; in, This indicates the current charging status. It is True when the device is connected to a charger and False when it is not connected. To show the downward trend of historical power consumption rates, data was collected from the most recent continuous... The current slope is obtained by calculating the difference in battery discharge current within each sampling period. When the current slope is continuously negative for multiple consecutive sampling periods, it is determined to be less than 0.

[0042] In this embodiment, the composite access decision engine performs the second-layer charging and energy consumption assessment. The sampling period is set to 30 seconds, and the number of consecutive sampling periods is... The platform obtains the current charging status of the device. If not charged ( Then calculate the slope of the battery discharge current over the last 180 seconds. If the slope is negative, it indicates that the intensity of the user's equipment use is decreasing, and the power consumption rate is trending downwards. In this case, the joint conditional discrimination formula... If the output is True, it indicates that the charging and energy consumption access conditions are met, and the node is allowed to enter the computing power pool.

[0043] The multi-layer access evaluation of the device status data based on the composite access decision engine further includes: real-time monitoring of the processor's current core junction temperature; when the current core junction temperature is less than a first temperature threshold, allowing the processor to execute calculations at full frequency; when the current core junction temperature is greater than or equal to the first temperature threshold and less than a second temperature threshold, controlling the processor to reduce its execution frequency or suspend receiving new subtasks; when the current core junction temperature is greater than or equal to the second temperature threshold, triggering overheat protection and exiting computing power sharing.

[0044] In this embodiment, the composite access decision engine performs a third-layer dynamic thermal state assessment. It reads the core junction temperature of the mobile terminal processor in real time. Set the first temperature threshold. Second temperature threshold .when When it is in a green and safe state, the processor is allowed to run at full frequency; when When the system is in a yellow alert state, the local client controls the computing process to reduce CPU usage by 30% to limit heat generation; when When the overheating occurs, the red overheat protection is triggered, immediately interrupting the calculation, erasing the sandbox temporary data, and exiting the computing pool. The specific values ​​in this embodiment are only illustrative examples. Those skilled in the art can make adaptive adjustments based on the hardware configuration of the mobile terminal node, the type of computing task, and user experience requirements. These adjustments do not require any creative effort.

[0045] The multi-layer access evaluation of the device status data based on the composite access decision engine also includes: acquiring the user's calendar habits and sleep patterns of the mobile terminal node, predicting the user's usage status within a future time window, and allocating sub-task units with corresponding computational load levels according to the prediction results; during the execution of the computing power sharing task, real-time monitoring of user interaction events, physical motion status, power change status, and charging connection status; when a preset exit event is detected (triggering a dynamic exit trigger mechanism), immediately releasing control of computing resources, interrupting the currently executing sub-task unit, and sending back the data.

[0046] In this embodiment, the composite access decision engine performs fourth and fifth layer evaluations. It predicts future time windows by analyzing the user's historical unlock logs. The system monitors the user's usage status. If a user is predicted to be asleep, a subtask with a "high" computational load is assigned to them; if the user is predicted to use the phone soon, only fragmented tasks with a "low" computational load are assigned. During task execution, the client listens for system broadcasts in real time. Once the system detects a user unlocking the screen (user interaction event) or an accelerometer detecting a change in acceleration, it will initiate a monitoring process. When any of the preset exit events (triggering the dynamic exit mechanism) occur, such as (physical motion state), battery level dropping below 30% (battery change state), or charger being unplugged (charging connection state), the dynamic exit mechanism is immediately triggered, interrupting calculations and releasing resources.

[0047] In the heterogeneous multi-dimensional intelligent scheduling stage, the computational requirements of the sub-task units, as well as the computing power performance, historical reputation, and estimated latency of the mobile terminal nodes in the executable computing power pool, are input into the scheduling matching degree function to calculate the scheduling matching degree score between each mobile terminal node and each sub-task unit.

[0048] The "scheduling matching degree function" is a mathematical model used to quantify the matching degree between subtasks and mobile terminal nodes. The "computing performance" refers to the normalized capability index of the mobile terminal hardware when processing a specific computing task. The "historical reputation" refers to the comprehensive score of the node's completion rate and disconnection frequency in historical task execution. The "estimated latency" includes the sum of network transmission latency and expected computation latency.

[0049] The general process for calculating the matching degree in intelligent scheduling is as follows: Extract the computational complexity feature vector of the subtask to be assigned; Extract the performance feature vectors and reputation scores of candidate nodes in the computing power pool; The overall score for each node in performing the subtask is calculated using a preset matching formula. The node with the highest score is selected as the target execution node, and hard constraints on latency and power consumption must be met.

[0050] In this embodiment, when distributing subtasks, the intelligent scheduling engine traverses and calculates each candidate node in the computing power pool to obtain the optimal node-task mapping relationship.

[0051] The formula for calculating the scheduling matching degree function is as follows: ; in, For mobile terminal nodes With sub-task units The scheduling matching score between them; and Sub-task units CPU computing requirements and NPU computing requirements; and Mobile terminal nodes CPU computing power score and NPU computing power score; For mobile terminal nodes Historical reputation score; Sub-task unit Assigned to mobile terminal nodes Estimated data transmission delay; The corresponding weight coefficients, and satisfying ; Furthermore, the scheduling selection must meet the following constraints: ; in, For mobile terminal nodes The current battery percentage; A user-defined minimum power threshold for participation; Sub-task unit The maximum allowable delay.

[0052] In this embodiment, a weighting coefficient is set. For subtask 001 ( , ), candidate nodes The parameters are: , , , Substituting the values ​​into the formula yields the matching score for that node. At the same time, verify the constraint condition: the current power level of this node. Greater than the custom minimum battery threshold And estimate the total delay. Less than the maximum allowed latency of the subtask If the constraints are satisfied, the node is determined to be usable.

[0053] The weight values ​​of each parameter in the scheduling matching function are dynamically adjusted according to the load status of the scheduling platform and the fluctuation of node reputation: when the scheduling platform detects that the backlog of the overall computing tasks is lower than the preset low load threshold, the transmission efficiency weight is automatically increased. The weighting of computing power score and reputation score will be reduced proportionally; when the historical reputation score fluctuation of a preset batch of mobile terminal nodes in the executable computing power pool exceeds a preset fluctuation threshold, the reputation score weighting will be automatically increased. The weights for transmission efficiency and computing power are reduced proportionally.

[0054] In this embodiment, the scheduling platform monitors system load and node reputation in real time. When the number of backlogged tasks in the task queue is lower than a preset low-load threshold (e.g., 5 tasks), it indicates that the system has sufficient computing power, and the bottleneck lies in network transmission. At this time, the platform automatically assigns transmission efficiency weights. The value was increased from 0.2 to 0.4, and... The weighting is adjusted proportionally to prioritize near-end nodes with extremely low network latency. When the standard deviation (fluctuation range) of the reputation score of a node in the computing power pool exceeds a preset fluctuation threshold (e.g., 0.15) within the most recent statistical period, it indicates an unstable network environment and frequent node disconnections. In this case, the platform automatically adjusts the reputation score weighting. The value was increased to 0.5 to prioritize high-reputation nodes with extremely stable historical performance, thus ensuring computational stability.

[0055] During the heartbeat monitoring and timeout rescheduling phase, based on the scheduling matching score, the sub-task units are distributed to the corresponding target mobile terminal nodes for shared computing power. During the computing period, the heartbeat monitoring module sends heartbeat detection packets to the target mobile terminal nodes at preset half-cycles. Whether the target mobile terminal node has abnormally disconnected is determined based on whether a heartbeat response packet is received within a consecutive preset cycle. If abnormal disconnection is determined, the backup distribution logic is triggered to redistribute the unfinished sub-task units to the backup mobile terminal nodes.

[0056] The "heartbeat monitoring module" is a lightweight daemon process on the scheduling platform responsible for detecting the liveness status of nodes. The "heartbeat probe packet" refers to a network probe data packet with extremely low bandwidth consumption, used to trigger an immediate response from the node. The "backup distribution logic" refers to a fault-tolerant mechanism in which the scheduling engine quickly activates a backup node to take over the unfinished task when the original execution node fails.

[0057] The general state machine process for heartbeat monitoring and rescheduling is as follows: After the task is distributed, a heartbeat timer is started for that node; Send probe packets according to the set period and listen for responses; Maintain a continuous non-response counter. The counter is reset to zero when a response is received and incremented by one if no response is received within a timeout period. When the counter reaches the set threshold, the node is determined to be offline, the task is redistributed, and the node's reputation is updated.

[0058] In this embodiment, the heartbeat monitoring module of the scheduling platform monitors the subtasks every preset half-cycle after the subtasks are distributed. A 32-byte UDP heartbeat probe packet is sent to the target mobile terminal. The preset half-cycle refers to a preset basic heartbeat probe time interval, which is set to 3 seconds in this embodiment and can be dynamically adjusted according to network quality.

[0059] The preset half-cycle is dynamically adjusted based on the network connection quality and historical reputation score of the target mobile terminal node: the network latency and packet loss rate of the target mobile terminal node are periodically obtained; if the network latency is greater than a preset latency threshold or the packet loss rate is greater than a preset packet loss rate threshold, the heartbeat detection cycle is extended to the first preset cycle; if the historical reputation score of the target mobile terminal node is lower than a preset reputation threshold, the heartbeat detection cycle is shortened to the second preset cycle.

[0060] In this embodiment, the heartbeat detection period employs an adaptive adjustment algorithm. The platform periodically (e.g., every 60 seconds) acquires the network status of the target node. If the network latency of the node is detected to be greater than 150ms or the packet loss rate is greater than 5%, it indicates poor network quality. To avoid frequent detections exacerbating network congestion or leading to misjudgments, the platform extends the heartbeat detection period from 3 seconds to a first preset period of 6 seconds. Conversely, if the node's historical reputation score... This indicates that the device is highly prone to disconnection. The platform shortens the heartbeat detection cycle to a second preset cycle of 1.5 seconds to achieve high-frequency monitoring, enabling immediate detection of disconnections. If a disconnection is detected, the node's reputation score is reduced by 0.05.

[0061] During the synchronization and result sorting phase of the waiting window, a synchronization waiting time window is opened to receive subtask calculation result packets from each mobile terminal node. Within the synchronization waiting time window, the received subtask calculation result packets are subjected to timing coordination alignment and consistency verification, and are reordered and concatenated according to the absolute sequence number of the subtasks to output the final calculation result of the original calculation task.

[0062] The "synchronization waiting time window" refers to the maximum time limit set by the platform for waiting for the results of all subtasks to return. The "timing coordination alignment" refers to reorganizing data packets that arrive out of order due to differences in device performance according to their logical sequence. The "consistency verification" refers to using a hash algorithm to verify the integrity and authenticity of the returned data, preventing malicious nodes from tampering with it.

[0063] The general process for result synchronization and rearrangement is as follows: Initialize the result collection buffer and start the synchronization window timer; Receive the result packets returned by each node and extract the absolute sequence number of the subtasks; Perform hash verification on the result packets and discard packets that fail the verification; If all serial numbers have been collected before the timer expires, they are concatenated in ascending order. If there are still missing blocks by the timeout, a preemptive rescheduling is triggered to retrieve the missing blocks and complete the assembly.

[0064] In this embodiment, a length of The synchronization waiting time window is defined. All arriving subtask result packets are temporarily stored in the aggregation buffer, and their Sequence IDs are extracted.

[0065] If, when the synchronization waiting time window expires, some sub-task units have not yet returned their calculation result packets, a fast preemption command is sent to the scheduling engine to redistribute the input data packets corresponding to the non-returning sub-task units to the currently online high-performance standby mobile terminal nodes for execution, and the final queue splicing is completed after obtaining the response.

[0066] In this embodiment, when the 60-second synchronization waiting window expires, if subtask 015 (due to frequency throttling caused by overheating of the mobile phone undertaking the task) still has not returned a result, the system will not continue waiting but will immediately send a "fast preemption command" to the intelligent scheduling engine 23. The scheduling engine will then quickly redistribute the input data packet of subtask 015 to the currently online NPU with the highest computing power score in the computing power pool (e.g., ...). The backup mobile terminal node executes the calculation. The backup node completes the calculation within 3 seconds and returns a result packet. This result packet is then reordered and concatenated with the other 31 existing result packets in ascending order according to their absolute sequence numbers 001-032, and the final LLM inference result is output.

[0067] During the security sandbox isolation and computing power measurement phase, the mobile terminal node executes the sub-task unit through a built-in lightweight security sandbox. The lightweight security sandbox establishes an independent user identification isolation mechanism based on the application layer of the operating system and restricts the access permissions of the lightweight security sandbox to the mobile terminal node's local file system and hardware sensors. When the sub-task unit completes execution or triggers an abnormal exit, the lightweight security sandbox automatically erases and reclaims temporary data generated during execution. Furthermore, the lightweight security sandbox imposes upper limits on the processor utilization and memory usage of the computing process during operation.

[0068] The "lightweight security sandbox" refers to an isolated operating environment built within the user space of a mobile terminal operating system. Its core innovation lies in protecting the local privacy data of mobile terminal users from being stolen by computing tasks through "two-way isolation," while also limiting the excessive consumption of system resources by computing tasks and ensuring the smooth operation of the foreground system.

[0069] The general operating procedure of a security sandbox is as follows: After receiving the subtask unit, the client initializes the sandbox container and assigns a unique UID; Configure the sandbox's resource quotas (CPU core limit, memory limit); Mount read-only compute dependency libraries and cut off access to local photo album, address book and sensors; Start the computation process within the sandbox; When the calculation ends or an abnormal exit occurs, the sandbox instance is destroyed, and the allocated memory space is filled with zeros and erased.

[0070] In this embodiment, the lightweight security sandbox deployed within the mobile terminal node is built based on the Android system's UID (User Identifier) ​​isolation mechanism. During sandbox operation, its CPU utilization is strictly limited to no more than 50%, and its memory usage is capped at 1GB. The sandbox is prohibited from calling hardware sensor APIs such as the system's Camera and Location functions, and cannot read user-shared storage directories such as / sdcard. Once the subtask is completed and the result package is successfully uploaded, the sandbox immediately calls the memory erasure function to overwrite and erase the temporary matrix data generated during the calculation process, completely releasing memory and ensuring user privacy and security.

[0071] After the task is completed, the computing power measurement and settlement module 25 collects the actual computing time, hardware resource consumption, and number of successfully uploaded result packets of each node when executing the sub-task unit. It calculates the standard computing unit (SCU) contribution value of each mobile terminal node and issues tokens or virtual asset rewards based on this.

[0072] This embodiment also discloses a distributed computing power sharing system based on a mobile terminal cluster, such as... Figure 2 As shown, it includes: The computing power demander interface layer provides task submission, status query, and result retrieval interfaces, enabling demanders to submit computing tasks, query subtask distribution status, and retrieve the final aggregated results. The cloud / local computing power scheduling platform is used to receive raw computing tasks, parse and split the raw computing tasks into multiple sub-task units and establish a directed acyclic graph topology, manage registered mobile terminal nodes, distribute the sub-task units to the optimal available mobile terminal nodes based on the scheduling matching degree function, perform runtime monitoring and abnormal rescheduling through the heartbeat monitoring module, and perform collaborative alignment and splicing of received computing results within the synchronization waiting time window; An encrypted communication layer is used to establish an encrypted communication channel between the cloud / local computing power scheduling platform and the mobile terminal node for data transfer.

[0073] The system described above already provides the basic server-side functionality. By adding mobile terminal nodes as users, all computing power sharing capabilities can be achieved. Typically, this is done through a mobile terminal node cluster. The cluster consists of multiple independent mobile terminal nodes, each equipped with a computing power contribution client. This client uses a composite access decision engine to perform multi-layered access evaluation to determine whether to join the executable computing power pool. It then executes distributed sub-task units within a lightweight security sandbox and sends heartbeat response packets and computation result packets back to the cloud / local computing power scheduling platform.

[0074] In practical use, this system architecture adopts a three-layer loosely coupled distributed design pattern. The top layer is the standard service interface layer facing the demand side; the middle layer is the core control hub of the "cloud / local computing power scheduling platform" deployed on a cloud server or local area network edge gateway; and the bottom layer is a massive heterogeneous mobile terminal node cluster. This architecture centralizes the complex task scheduling, fault tolerance management, and result aggregation logic in the computing power scheduling platform, greatly reducing the computing burden on the mobile terminal side, allowing ordinary smartphones to join the cluster as computing power contributing nodes. The "encrypted communication layer" is built based on the TLS 1.3 security protocol to ensure that subtask data and computing results are not eavesdropped on or tampered with when transmitted over the public network or local area network.

[0075] The physical deployment and execution entity affiliation of each functional module in this system are as follows: Task parsing module: Deployed and running on a cloud / local computing power scheduling platform, it is a core logical component of the platform. Its hardware carrier is the server's CPU and memory, responsible for the structured decomposition of raw computing tasks input from external sources.

[0076] Node Management Module: Deployed and runs on a cloud / local computing power scheduling platform. It maintains a high-concurrency online node database to record the hardware configuration, real-time physical status, and historical reputation scores of all registered mobile terminal nodes in real time.

[0077] Intelligent scheduling engine: Deployed and running on a cloud / local computing power scheduling platform. It is responsible for performing the aforementioned scheduling matching function calculation and outputting the optimal task-node mapping allocation scheme.

[0078] Heartbeat monitoring module: Deployed and runs on a cloud / local computing power scheduling platform. As a background daemon on the platform side, it concurrently listens for heartbeat responses from hundreds or thousands of mobile terminal nodes using asynchronous non-blocking I / O.

[0079] Result aggregation module: Deployed and runs on the cloud / local computing power scheduling platform. Responsible for maintaining the result collection buffer, performing time-series co-alignment, consistent hash verification, and final result reordering and concatenation.

[0080] Computing power metering and settlement module: Deployed and running on a cloud / local computing power scheduling platform. Responsible for generating a revenue / equity settlement list based on the actual computing power contribution of each node after all tasks are completed.

[0081] The composite access decision engine has one part of its evaluation logic (such as basic condition admission, dynamic evaluation of hot status, and dynamic exit triggering mechanism) running locally on the computing power contribution client of the mobile terminal node to achieve millisecond-level local physical status awareness; the other part (such as scheduling matching degree score calculation and historical reputation score query) runs on the cloud / local computing power scheduling platform to achieve global optimal decision-making.

[0082] Lightweight security sandbox: Fully deployed and running inside each mobile terminal node, existing as a sub-container of the client process.

[0083] In this embodiment, the "cloud / local computing power scheduling platform" is deployed on an edge computing server equipped with a quad-core Intel Xeon processor and 16GB of DDR4 memory. This server runs Ubuntu Linux 22.04 and uses Docker containerization to independently encapsulate and run the task parsing module, node management module, intelligent scheduling engine, result aggregation module, computing power metering and settlement module, and heartbeat monitoring and fault-tolerant scheduling module. The computing power requester interface layer is exposed on port 8080 of the server. Requesters submit LLM distributed inference tasks via a RESTful API call to this interface. The encrypted communication layer uses the TLS 1.3 protocol, and all subtask data packets transmitted between the platform and mobile terminal nodes are symmetrically encrypted using the AES-256-GCM algorithm. The mobile terminal node cluster 4 consists of 120 idle smartphones of different brands and models. They connect to the router in the local area network via Wi-Fi, discover the locally deployed computing power scheduling platform using the UDP multicast protocol, and register and access the platform through the computing power contribution client installed on their phones.

[0084] The complete workflow of the various modules in this system when executing distributed computing power sharing tasks is as follows: Computing power requesters upload their original computing tasks through the task submission interface; The task parsing module in the cloud / local computing power scheduling platform receives the task, breaks it down into 32 sub-task units, and establishes a DAG topology; The node management module traverses the list of currently online mobile terminal nodes and synchronizes the real-time status data of each node to the intelligent scheduling engine. The intelligent scheduling engine runs the scheduling matching function, calculates the optimal allocation scheme, and distributes the subtask units to the corresponding mobile terminal nodes through an encrypted communication channel; After the mobile terminal node's computing power contribution client calls the local composite access decision engine to complete the final access verification, it loads the subtask unit into the lightweight security sandbox to perform calculations. The heartbeat monitoring and fault-tolerant scheduling module sends heartbeat detection packets to each mobile terminal node at a preset half-cycle to monitor the node's liveness status in real time. After each mobile terminal node has completed its calculations, it sends the subtask result packet back to the result aggregation module via an encrypted communication channel. The result aggregation module completes the temporal alignment and reordering of the results within the synchronous waiting time window, and finally returns the complete LLM inference results to the computing power requester through the result acquisition interface. Finally, the computing power measurement and settlement module calculates the standard computing power unit (SCU) contribution value of each mobile terminal node based on the actual computing time and result contribution of each node, and distributes virtual asset rewards to the corresponding node accounts.

[0085] In another embodiment, because iPhones take an average of 2-5 seconds to switch between WiFi and cellular networks, and Android's WiFi scanning cycle is approximately 15-60 seconds, UDP packets are 100% lost during the switching process. Network interruptions of 10-30 seconds are common in scenarios such as subway tunnels and elevators. If the heartbeat cycle is set to 3 seconds and N=3 times is used to determine a disconnection, the above scenarios will trigger 100% false positives.

[0086] The node management module registers listener callbacks for changes in the system's network status. When a change in the network connection type of a mobile terminal node is detected (such as WiFi disconnection, 4G fallback, airplane mode resumption, etc.), the node's heartbeat determination timer is automatically suspended for 1 second (T1 can be 5-10 seconds) to wait for the new link to stabilize. Heartbeat listening resumes after T1 has elapsed to avoid misjudgments caused by packet loss during the switchover.

[0087] A three-level node health status is set: Healthy (Online): Normal heartbeat response, counter = 0; Suspected Disconnection: No response received N times consecutively (e.g., N=3), counter = N, start a T2-second grace period timer (e.g., T2=15 seconds); Confirmed Disconnection: No response received within the grace period, counter > N, determined to be disconnected, triggering redistribution. If a node responds again in the suspected disconnection state, and the originally assigned subtask has not exceeded the maximum allowed delay, the node is allowed to continue executing the original task, the counter is reset to zero, and the node returns to the healthy state.

[0088] PC-based distributed computing (such as BOINC) has a stable node network connection, so this optimization is not necessary. However, not doing so on mobile devices will lead to frequent misjudgments.

[0089] In yet another embodiment, when a mobile terminal node registers with the scheduling platform for the first time, the computing power contribution client automatically executes a set of lightweight benchmark test processes (total time not exceeding 5 seconds), including the following representative operators: Matrix multiplication (M×N×K, M=N=K=1024, FP16): Characterizes general computing power; Convolution operation (3×3 convolution, 1×128×128 → 32×126×126): characterizes image processing capability; Transformer forward reasoning (single layer, hidden=768, FFN=3072): represents the AI ​​reasoning ability.

[0090] After the benchmark test is completed, the client reports the measured throughput (OPS) of each operator to the scheduling platform. Based on this, the platform generates an operator-level performance profile vector for that node: P_node = [CPU_GFLOPs, GPU_OPS, NPU_TOPS, Conv_OPS, Transformer_OPS,...] In subsequent task distribution, this performance profile will be used for matching. At the same time, the platform will trigger a benchmark update every certain period of time (e.g., every 24 hours) to reflect the real performance after node hardware aging or system state changes.

[0091] In another embodiment, the following two-way authentication process is performed when the mobile terminal node registers for the first time: The scheduling platform sends a randomly generated Challenge Token (32 bytes) to the terminal. The terminal client uses the device's hardware fingerprint (generated by a mixed hash of IMEI / AndroidID / device serial number / motherboard ID) to sign the Challenge Token and sends the signing result back to the platform; After the platform verifies the signature, it assigns a unique node ID to the terminal and records the hardware fingerprint in the whitelist; if the signature verification fails, registration is rejected; each heartbeat packet includes a re-signed token fragment to prevent session hijacking. Afterward, the platform can block malicious nodes based on the hardware fingerprint whitelist—even if the attacker changes the application account and re-registers, as long as the hardware fingerprint remains unchanged, access will still be denied.

[0092] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for sharing distributed computing power based on mobile terminal clusters, characterized in that, Based on a system architecture that includes a scheduling platform with cloud / local computing power, encrypted communication channels, and mobile terminal node clusters, the following steps are included: Step 1: Obtain the original computation task to be processed, and decompose the original computation task into multiple sub-task units carrying input parameters and execution environment through the task parsing module, and establish a directed acyclic graph topology; Step 2: Obtain the device status data of the registered mobile terminal nodes through the node management module, and perform multi-level access evaluation on the device status data based on the composite access decision engine to screen out mobile terminal nodes that meet the access conditions for computing power sharing and build an executable computing power pool. Step 3: Input the computational requirements of the sub-task unit, the computing power performance, historical reputation, and estimated latency of the mobile terminal nodes in the executable computing power pool into the scheduling matching degree function to calculate the scheduling matching degree score between each mobile terminal node and each sub-task unit. Step 4: Based on the scheduling matching score, the sub-task units are distributed to the corresponding target mobile terminal nodes for computing power sharing; and during the calculation, the heartbeat monitoring module sends heartbeat detection packets to the target mobile terminal nodes at a preset half-cycle. The target mobile terminal node is determined to be abnormally disconnected based on whether a heartbeat response packet is received within a consecutive preset cycle. If an abnormal disconnection is determined, the backup distribution logic is triggered to redistribute the unfinished sub-task units to the backup mobile terminal nodes. Step 5: Open the synchronization waiting time window, receive the subtask calculation result packets fed back by each mobile terminal node, perform timing coordination alignment and consistency verification on the received subtask calculation result packets within the synchronization waiting time window, and reorder and concatenate them according to the absolute sequence number of the subtasks to output the final calculation result of the original calculation task.

2. The distributed computing power sharing method based on mobile terminal clusters according to claim 1, characterized in that, The steps of performing multi-layer access evaluation on the device status data based on the composite access decision engine, filtering out mobile terminal nodes that meet the computing power sharing admission conditions, and constructing an executable computing power pool specifically include: obtaining the current network connection status, current remaining battery power, and foreground application running status of the mobile terminal node; if the current network connection status is a preset network, the current remaining battery power is greater than a preset battery power threshold, and there are no sensitive applications running in the foreground, then it is determined that the basic conditions for admission are met; after meeting the basic conditions for admission, the current charging status and historical power consumption rate of the mobile terminal node are obtained, and the current charging status and historical power consumption rate are jointly evaluated by a joint condition discrimination formula, wherein the historical power consumption rate is obtained by collecting the battery discharge current value in multiple consecutive sampling periods and calculating the discharge current difference; if the current slope corresponding to the discharge current difference is continuously negative in multiple consecutive sampling periods, it is determined that the user usage intensity of the mobile terminal node is weakening, the joint condition discrimination formula outputs true, and it is determined that the charging and energy consumption admission conditions are met.

3. The distributed computing power sharing method based on mobile terminal clusters according to claim 1, characterized in that, The step of performing multi-layer access evaluation on the device status data based on the composite access decision engine, selecting mobile terminal nodes that meet the access conditions for computing power sharing, and constructing an executable computing power pool further includes: real-time monitoring of the processor core junction temperature of the mobile terminal node, and executing corresponding thermal state management strategies according to the temperature range of the processor core junction temperature. Specifically, when the processor core junction temperature is less than a first temperature threshold, the processor is controlled to execute calculations at full frequency; when the processor core junction temperature is greater than or equal to the first temperature threshold but less than a second temperature threshold, the processor is controlled to reduce its execution frequency or suspend receiving new subtasks until the processor core junction temperature drops below the temperature difference; when the processor core junction temperature is greater than or equal to the second temperature threshold, overheat protection is triggered, security sandbox data is cleared, and the computing power sharing is exited; the user calendar habits and sleep patterns of the mobile terminal node are obtained, the user's usage status within a future time window is predicted, and subtask units with corresponding computing power scale are allocated to the mobile terminal node according to the prediction results.

4. The distributed computing power sharing method based on mobile terminal clusters according to claim 1, characterized in that, The step of performing multi-level access evaluation on the device status data based on the composite access decision engine, selecting mobile terminal nodes that meet the access conditions for computing power sharing, and constructing an executable computing power pool also includes a dynamic exit trigger mechanism: during the execution of sub-task units by the mobile terminal node, after obtaining user authorization, the user interaction events, physical motion status, power change status, and charging connection status of the mobile terminal node are monitored in real time. When the system detects that a user unlocks the screen to launch a foreground application, the acceleration change detected by the accelerometer exceeds a preset acceleration threshold, the remaining battery power drops to a user-defined minimum participation battery power threshold, or the charger is unplugged and the remaining battery power is below a set high battery power limit, the system immediately releases control of the mobile terminal node's computing resources, interrupts the currently executing sub-task unit, and sends the unfinished sub-task unit back to the scheduling platform so that the scheduling platform can redistribute the unfinished sub-task unit to other backup mobile terminal nodes.

5. The distributed computing power sharing method based on mobile terminal clusters according to claim 1, characterized in that, The scheduling matching degree function is calculated based on the CPU and GPU / NPU computing requirements of the sub-task unit, as well as the CPU computing power score, NPU / GPU computing power score, historical reputation score, and estimated data transmission latency of the mobile terminal node. The weight values ​​corresponding to each parameter in the scheduling matching degree function satisfy normalization constraints. Furthermore, when selecting a node using the scheduling matching degree function, preset scheduling constraints must be met. These preset constraints include: the current battery percentage of the mobile terminal node is greater than a user-defined minimum participation battery threshold, and the estimated total execution time of the mobile terminal node for the corresponding sub-task unit is less than or equal to the maximum allowable latency of that sub-task unit. The estimated total execution time is calculated based on the computing power requirements of the sub-task unit, the computing power score of the mobile terminal node, and the estimated data transmission latency.

6. The distributed computing power sharing method based on mobile terminal clusters according to claim 5, characterized in that, The weight values ​​of each parameter in the scheduling matching function are dynamically adjusted according to the load status of the scheduling platform and the fluctuation of node reputation. Specifically, when the scheduling platform detects that the backlog of overall computing tasks is lower than a preset low load threshold, it automatically increases the transmission efficiency weight corresponding to the estimated data transmission latency and proportionally decreases the computing power score weight and reputation score weight to reduce network transmission time. When it detects that the historical reputation score fluctuation of a preset batch of mobile terminal nodes in the executable computing power pool exceeds a preset fluctuation threshold, it automatically increases the reputation score weight corresponding to the historical reputation score and proportionally decreases the transmission efficiency weight and computing power score weight to improve the sensitivity to prevent disconnection and avoid assigning tasks to mobile terminal nodes with unstable reputation.

7. A distributed computing power sharing method based on a mobile terminal cluster according to claim 1, characterized in that, If the system architecture is in a dedicated networking mode, the method further includes a local dedicated computing center self-organizing network step: deploying the scheduling platform as an independent background process on a resident physical device within the local area network; and broadcasting or multicasting a description file carrying the cluster service set identifier (SSID) and task type within the local area network through the scheduling platform. Mobile terminal nodes within the local area network (LAN) listen to specific ports via clients. After receiving and confirming the description file, they register and access the scheduling platform to build a localized dedicated computing power cluster. The input data for the original computing task is sliced ​​by edge servers within the LAN and distributed to the mobile terminal nodes in the dedicated computing power cluster via wireless channels based on the LAN unicast protocol. The computing process and result aggregation are completed in a closed loop within the LAN to achieve physical isolation of sensitive data within the LAN.

8. The distributed computing power sharing method based on mobile terminal clusters according to claim 1, characterized in that, The mobile terminal node executes the sub-task unit through a built-in lightweight security sandbox. The lightweight security sandbox establishes an independent user identifier (UID) isolation mechanism based on the operating system's application layer and restricts the lightweight security sandbox's access permissions to the mobile terminal node's local file system and hardware sensors to prevent computing tasks from reading user privacy data. When the sub-task unit completes execution or triggers an abnormal exit, the lightweight security sandbox automatically erases and reclaims temporary data, cached data, and memory usage generated during execution. At the same time, the lightweight security sandbox sets upper limits on the CPU usage and memory usage of the computing process during operation to prevent computing tasks from excessively consuming system resources, which could cause system lag or interface response delays on the mobile terminal node.

9. A distributed computing power sharing method based on a mobile terminal cluster according to claim 1, characterized in that, After outputting the final calculation result of the original computing task, the method further includes a computing power measurement and revenue distribution step: through the computing power measurement and settlement module, the actual computing time, hardware resource consumption, and number of successfully uploaded result packets of each mobile terminal node when executing the sub-task unit are collected, and the standard computing power unit (SCU) contribution value of each mobile terminal node is calculated. The revenue distribution ratio of each mobile terminal node is calculated based on the contribution value of the standard computing power unit (SCU) and the corresponding historical reputation score of the node. In a dedicated network mode, the contribution value of the standard computing power unit (SCU) is converted into corresponding enterprise work points or community shared rights and interests, and recorded and settled within the local area network. In the public network scheduling mode, tokens or virtual asset rewards are issued to the corresponding mobile terminal nodes based on the aforementioned revenue distribution ratio. In the step of the heartbeat monitoring module sending heartbeat probe data packets to the target node at a fixed time period, the fixed time period is dynamically adjusted according to the network connection quality and historical reputation score of the target node. Specifically, it includes: periodically obtaining the network latency and packet loss rate of the target node; if the network latency is greater than a preset latency threshold or the packet loss rate is greater than a preset packet loss rate threshold, the fixed time period is extended to a first preset period to reduce false judgments of heartbeat response timeout caused by instantaneous network fluctuations. If the historical reputation score of the target node is lower than the preset reputation threshold, the fixed time period is shortened to a second preset period to increase the monitoring frequency of the offline status of low-reputation nodes. The first preset period is longer than the fixed time period, and the second preset period is shorter than the fixed time period.

10. A distributed computing power sharing system based on mobile terminal clusters, characterized in that, include: The computing power demander interface layer provides task submission, status query, and result retrieval interfaces, enabling demanders to submit computing tasks, query subtask distribution status, and retrieve the final aggregated results. The cloud / local computing power scheduling platform is used to receive raw computing tasks, parse and split the raw computing tasks into multiple sub-task units and establish a directed acyclic graph topology, manage registered mobile terminal nodes, distribute the sub-task units to the optimal available mobile terminal nodes based on the scheduling matching degree function, perform runtime monitoring and abnormal rescheduling through the heartbeat monitoring module, and perform collaborative alignment and splicing of received computing results within the synchronization waiting time window; An encrypted communication layer is used to establish an encrypted communication channel between the cloud / local computing power scheduling platform and the mobile terminal node for data transfer.

Images