Multi-soc cooperative task allocation method based on quantum ant colony optimization

Abstract

The application discloses a multi-SoC cooperative task allocation method based on quantum ant colony optimization, and collects node resource portrait set and task attribute set of each system-on-a-chip node in a multi-SoC system; a task-resource mapping diagram is constructed; a quantum encoding task set is obtained; the quantum encoding task set is loaded to a search state space of a quantum ant individual, an information element matrix is set, a quantum rotation angle matrix is set, an initial population of quantum ant colony is obtained; an updated quantum state population is obtained; delay indexes, power consumption indexes and reliability indexes of each deterministic task allocation result are calculated according to the node resource portrait set, a candidate feasible solution set meeting real-time constraints is screened out; and a global scheduling scheme is obtained. The application significantly improves modeling granularity and allocation accuracy of task allocation under multi-domain heterogeneous resources, effectively avoids bandwidth contention and delay mutation problems in cross-node and cross-domain migration, and realizes real-time and high adaptability of task allocation.

Description

Technical Field

[0001] This invention relates to the field of radar technology, and in particular to a multi-SoC collaborative task allocation method based on quantum ant colony optimization. Background Technology

[0002] With the widespread application of heterogeneous multi-core system-on-a-chip (SoC) technology, distributed computing architecture based on multi-SoC collaboration has become the mainstream choice for handling complex computing and multi-task concurrency. Modern multi-SoC systems are typically composed of multiple system-on-a-chip nodes that integrate processing system domains (PS) and programmable logic domains (PL), and each node achieves resource collaboration through high-speed on-chip interconnects.

[0003] For collaborative task allocation in multi-SoC systems, existing technologies mainly employ a combination of traditional heuristic algorithms, classic ant colony optimization algorithms, and genetic algorithms. While these methods can achieve a certain degree of resource allocation optimization in homogeneous clusters or single-domain environments, they have the following limitations in multi-SoC heterogeneous constraints, PS / PL collaborative allocation, task migration across nodes, node congestion and bandwidth management, and dynamic load balancing:

[0004] Traditional task allocation models struggle to perform fine-grained modeling of PS and PL dual-domain resources, failing to fully characterize the engineering constraints between different on-chip system nodes. Existing ant colony optimization and intelligent scheduling methods often fail to effectively utilize quantum coding or dynamic pheromone mechanisms to improve search efficiency, resulting in difficulties in ensuring the real-time performance and global feasibility of scheduling results when facing complex dynamic loads, cross-domain migrations, and node resets. Summary of the Invention

[0005] One objective of this invention is to propose a multi-SoC collaborative task allocation method based on quantum ant colony optimization. This invention significantly improves the modeling granularity and allocation accuracy of task allocation under multi-domain heterogeneous resources, effectively avoids bandwidth contention and latency mutation problems in cross-node and cross-domain migration, and achieves real-time and high adaptability of task allocation.

[0006] A multi-SoC collaborative task allocation method based on quantum ant colony optimization includes:

[0007] In a multi-SoC system, collect node resource profiles and task attribute sets for each on-chip system node;

[0008] Construct a task-resource mapping graph based on the node resource profile set and the task attribute set;

[0009] Perform quantum state encoding on the task-resource mapping graph, and encode each task to be scheduled using a set of qubit strings to obtain a set of quantum-encoded tasks.

[0010] Initialize the quantum ant colony, randomly generate individual quantum ants, load the quantum coding task set into the search state space of the individual quantum ants, set the pheromone matrix and the quantum rotation angle matrix to obtain the initial population of the quantum ant colony;

[0011] An iterative search operation is performed in the initial population of the quantum ant colony. Controlled-domain differential quantum rotation gate operation and cross-task controlled rotation gate operation are performed on the qubit string to obtain the updated quantum state population.

[0012] The updated quantum state population is subjected to quantum state collapse operation to convert the qubit string into deterministic task allocation results. Based on the node resource profile set, the delay index, power consumption index and reliability index of each deterministic task allocation result are calculated, and a set of candidate feasible solutions that meet the real-time constraints are selected.

[0013] Perform local refinement operations on the candidate feasible solution set, adjust the task-resource mapping relationship according to the current system load status and interconnect bandwidth occupancy status, write the adjusted task allocation results back to the pheromone matrix and update the quantum rotation angle matrix to obtain the global scheduling scheme.

[0014] Optionally, the resource profile set and task attribute set for the data collection nodes include:

[0015] Resource profiles are collected for each on-chip system node in a multi-SoC system, and node resource feature vectors are constructed.

[0016] The node resource feature vectors corresponding to each on-chip system node are summarized to form a set of node resource profiles;

[0017] For each task in the set of tasks to be scheduled, extract attributes and construct a task attribute vector.

[0018] The task attribute vectors of each task to be scheduled are aggregated to form a task attribute set;

[0019] The task attribute set describes the resource constraints and demand distribution of all scheduled tasks within a multi-SoC system.

[0020] Optionally, the construction task-resource mapping graph includes:

[0021] Based on the set of node resource profiles and the set of task attributes, a task-resource mapping graph is constructed.

[0022] For each edge in the task-resource mapping graph, construct a composite edge weight vector that includes the expected response latency, average power consumption, and reliability evaluation.

[0023] By weighted summing the ratio of expected response delay to task real-time deadline, the ratio of average power consumption to system power consumption limit, and the value after subtracting reliability evaluation, a task-resource matching cost function is constructed.

[0024] Write the task-resource matching cost function value of each pair of tasks and on-chip system nodes into the composite edge weight in the task-resource mapping graph.

[0025] Optionally, encoding each task to be scheduled using a set of qubit strings includes:

[0026] Based on the task-resource mapping graph, the total number of tasks in the task node set is determined to be N, the total number of nodes in the on-chip system node set is determined to be K, and the deployment location sets are defined as PS and PL, where PS represents the processing system domain and PL represents the programmable logic domain.

[0027] For each task in the task set, a quantum state code is constructed. The quantum state code consists of two sets of qubits. The first set of qubits represents the target on-chip system node to which the task is assigned, and the second set of qubits represents the deployment domain of the task in the target on-chip system node.

[0028] The first group of qubits is indexed and encoded using binary encoding. All on-chip system nodes are numbered, and the node encoding of each task is initialized to a quantum superposition state with equal probability amplitude. All states corresponding to the node numbers have the same probability of being selected at the initial moment.

[0029] The deployment position of the second qubit group is encoded, and one binary qubit is used to represent the deployment state of the processing system domain and the programmable logic domain respectively. The processing system domain corresponds to 0 and the programmable logic domain corresponds to 1. The sum of the initial distribution probabilities of each task in the two deployment domains is equal to 1.

[0030] The quantum state encoding step is performed on each task separately to generate quantum state codes for all tasks in sequence, forming a set of quantum state codes used to represent the search state space of node allocation and domain deployment of all tasks in the multi-SoC system.

[0031] Optionally, the initialization of the quantum ant colony includes:

[0032] The quantum ant colony is initialized for multi-SoC system collaborative tasks based on the quantum state encoding set. The total number of quantum ant individuals is set, and each quantum ant individual corresponds to a set of quantum encoded task copies as the search state representation of the quantum ant individual, thus constructing the initial population of the quantum ant colony.

[0033] Assign a corresponding task search state vector to each individual quantum ant;

[0034] Construct a pheromone matrix, where each element represents the pheromone intensity when a task is assigned to a specific on-chip system node and deployed in a specific deployment domain.

[0035] For each qubit of each task, construct a quantum rotation angle matrix. Each element in the quantum rotation angle matrix represents the rotation angle used by a quantum ant for a qubit of a certain task at a certain number of iterations.

[0036] The task search state vector, pheromone matrix, and quantum rotation angle matrix of each quantum ant individual are bound together to form a quantum ant individual state structure. All quantum ant individual state structures are integrated to construct the initial population structure of the quantum ant colony.

[0037] Optionally, performing controlled-domain differential quantum rotation gate operations and cross-task controlled rotation gate operations on the qubit string includes:

[0038] At the beginning of each iteration round, the operating status of the multi-SoC system is monitored, and the congestion ratio of each on-chip system node is calculated.

[0039] Calculate the adaptive rotational gain based on the current iteration round and thermal state indices;

[0040] For each individual quantum ant in the quantum ant colony, perform a congestion-suppressing quantum rotation update for the node qubit group:

[0041] ;

[0042] in, The rotation angle of the previous round node qubit. The rotation angle of the updated node qubit. For pheromone intensity, For the task Average pheromones For the first Task in the next iteration Assigned to on-chip system nodes Task—Resource matching cost function value, The optimal matching cost function value, For symbolic functions, This is the congestion ratio;

[0043] For each task's node qubit set and deployment domain qubit set, perform a controlled-domain differential quantum rotation gate operation;

[0044] ;

[0045] Among them, the control bit is the highest bit of the node qubit. , The coefficient of difference between regions. The deadline for the task in real time;

[0046] For task pairs with communication dependencies, if the corresponding element of the dependency matrix is ​​one, then a cross-task controlled rotation gate operation is applied between the node qubit groups of the two tasks:

[0047] ;

[0048] in, For elements in the dependency matrix, if Indicates task With the task There is a direct communication dependency; a value of 0 indicates no dependency. For the first Task in the next iteration and tasks The cross-task qubit coupling rotation angle between them For the first Task in the next iteration and tasks The cross-task qubit coupling rotation angle between them For the dependent adjustment coefficient, For the first Task in the next iteration Assigned to on-chip system nodes Task—Resource matching cost function value, For the task With the task The predicted average communication power consumption required to complete one data communication between nodes on the same on-chip system or between adjacent nodes;

[0049] After completing all controlled-domain differential quantum rotation gate operations and cross-task controlled rotation gate operations, the node qubit group and the deployment domain qubit group of each quantum ant individual are normalized respectively;

[0050] The updated quantum state population is obtained by summing all individual quantum ants. Then, according to the improved pheromone volatilization enhancement rule, the pheromone intensity of each task assigned to each on-chip system node and deployment domain is updated until the termination condition is met, and the updated quantum state population is output.

[0051] Optionally, the step of filtering out the set of candidate feasible solutions that satisfy the real-time constraints includes:

[0052] Perform a quantum state collapse operation on the quantum state population updated by the quantum rotation gate operation;

[0053] The collapse results of all tasks of each quantum ant individual are combined according to the task number to obtain the complete task allocation decision vector of the quantum ant individual in this round of search;

[0054] For each quantum ant individual, the complete task allocation decision vector is used to calculate the latency, power consumption and reliability indicators for all tasks based on the resource feature vector of each on-chip system node in the node resource profile set.

[0055] For the evaluation results of each task, determine whether the real-time constraints are met;

[0056] All complete task assignment decision vectors that meet the real-time constraints are aggregated to form a set of candidate feasible solutions.

[0057] Optionally, the real-time constraints include:

[0058] Only when the latency of all tasks is less than or equal to the task deadline, the power consumption of all tasks is less than or equal to the maximum allowable average power consumption of the node, and the reliability of all tasks is greater than or equal to the preset reliability threshold, is the complete task allocation decision vector of the corresponding quantum ant individual marked as a candidate feasible solution.

[0059] Optionally, performing a local refinement operation on the candidate feasible solution set includes:

[0060] From the set of candidate feasible solutions, the optimal scheduling result within the current scheduling period is selected as the starting solution for local refinement.

[0061] By combining the current system's operational status information, the current computing load percentage of the on-chip system node is obtained by summing the ratio of the computing workload of all tasks mapped to the node's processing system domain to the node's processing system domain computing capacity.

[0062] For each task on the allocated on-chip system node, the relative interconnect bandwidth pressure generated when the task is mapped to the on-chip system node is obtained by the ratio of the task's bandwidth requirement to the interconnect bandwidth capability of the on-chip system node.

[0063] Calculate the local adaptation cost for each task;

[0064] For each task, while keeping other task allocation schemes unchanged in the candidate scheduling neighborhood, all combinations of on-chip system node number and deployment domain number are traversed, the local adaptation cost under each combination is calculated, the on-chip system node number and deployment domain number with the minimum local adaptation cost are selected as the local refinement result of the task, and the original allocation is replaced with the local refinement result.

[0065] Once all tasks have been locally refined, the refinement results of all tasks are combined into a new task allocation decision vector. The pheromone values ​​corresponding to the on-chip system node number and deployment domain number of each task are written back. The locally refined task allocation decision vector is then output as the final global scheduling scheme.

[0066] For the refinement results of each task, the on-chip system node number and deployment domain number are converted into binary code, and the binary code is initialized into a quantum measurement state with a probability of one. The refinement results of all tasks are combined to form a refined quantum state set, which is used to guide the next round of quantum rotation gate operation.

[0067] The beneficial effects of this invention are:

[0068] This invention introduces dual-domain qubit encoding to synchronously map the node allocation and deployment state of the processing system domain / programmable logic domain for each task into a set of qubit strings. This fully characterizes the heterogeneity of multi-SoC systems, node resource distribution, internal and external domain constraints, and task executability in the search space. Combined with local refinement and adaptive quantum rotation gate strategies, it can dynamically adjust the task-resource mapping relationship in real time for hard constraints, significantly improving the modeling granularity and allocation accuracy of task allocation under multi-domain heterogeneous resources. It effectively avoids bandwidth contention and latency mutation problems in cross-node and cross-domain migration, achieving real-time and high adaptability of task allocation.

[0069] In the quantum ant colony optimization process, this invention combines the node load state, on-chip interconnect bandwidth pressure, and inter-task communication dependencies of a multi-SoC system. It introduces congestion suppression quantum rotation gate, controlled-domain differential quantum rotation gate, and cross-task controlled rotation gate operations. By adaptively adjusting the rotation gain based on the node congestion ratio and thermal state, the global search efficiency and convergence stability of the algorithm in dynamic scenarios are improved. At the same time, the task communication dependency matrix is ​​used to apply linkage adjustment to highly coupled task pairs, which significantly reduces the energy consumption and scheduling conflicts of high communication load tasks after allocation. Experimental results show that in complex SoC clusters, the average system task latency and energy consumption of this invention are significantly lower than those of traditional ant colony optimization and heuristic allocation methods.

[0070] This invention achieves dynamic fine-tuning by combining the current system load, interconnect bandwidth, node reliability, and task real-time objectives through local refinement of the candidate feasible solution set and refined local adaptation cost evaluation. The refined allocation result is directly fed back and the pheromone matrix is ​​strengthened, realizing the empirical inheritance of historical allocation paths and the dynamic enhancement of the optimal allocation result, ensuring that the global scheduling result can still output a highly reliable solution stably under complex working conditions. Attached Figure Description

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0072] Figure 1 This is a flowchart of a multi-SoC collaborative task allocation method based on quantum ant colony optimization proposed in this invention. Detailed Implementation

[0073] Example 1: Reference Figure 1 A multi-SoC collaborative task allocation method based on quantum ant colony optimization includes:

[0074] In a multi-SoC system, collect node resource profiles and task attribute sets for each on-chip system node;

[0075] In this embodiment, the collection of node resource profiles and task attribute sets includes:

[0076] Resource profiles are collected for each on-chip system node in a multi-SoC system, and node resource feature vectors are constructed.

[0077] The node resource feature vector describes the resource capabilities of the k-th system-on-a-chip (SoC) node. The node resource feature vector consists of three parts: the number of computing cores in the processing system domain, the size of the logic units in the programmable logic domain, and the on-chip interconnect bandwidth capability. The number of computing cores in the processing system domain represents the number of cores in the processing system domain of the k-th SoC node. The size of the logic units in the programmable logic domain represents the total number of lookup table units in the programmable logic domain of the k-th SoC node. The on-chip interconnect bandwidth capability represents the data throughput capability of the on-chip interconnect within the k-th SoC node.

[0078] The node resource feature vectors corresponding to each on-chip system node are summarized to form a set of node resource profiles;

[0079] The node resource profile set describes the resource distribution characteristics of all on-chip system nodes within a multi-SoC system.

[0080] For each task in the set of tasks to be scheduled, extract attributes and construct a task attribute vector.

[0081] The task attribute vector describes the resource requirements and constraints of the i-th scheduled task. The task attribute vector consists of three parts: task real-time deadline, processing system domain computational workload, and memory bandwidth requirement. The task real-time deadline describes the maximum allowed response time for the i-th scheduled task to complete. The processing system domain computational workload describes the instruction execution requirements of the i-th scheduled task for the processing system domain computational resources. The memory bandwidth requirement describes the data throughput requirements of the i-th scheduled task for the on-chip interconnect within the system-on-chip node.

[0082] The task attribute vectors of each task to be scheduled are aggregated to form a task attribute set;

[0083] The task attribute set describes the resource constraints and demand distribution of all scheduled tasks within a multi-SoC system.

[0084] Construct a task-resource mapping graph based on the node resource profile set and the task attribute set;

[0085] In this embodiment, constructing the task-resource mapping graph includes:

[0086] Based on the set of node resource profiles and the set of task attributes, a task-resource mapping graph is constructed.

[0087] In Example 1, the task-resource mapping graph is obtained by: determining the resource feasibility of task allocation on each on-chip system node for all tasks to be scheduled within the multi-SoC system.

[0088] For each combination of task and on-chip system node, if the computational workload of the processing system domain required by the task attribute vector is less than or equal to the number of processing system domain computational cores contained in the node resource feature vector, and the memory bandwidth requirement required by the task attribute vector is less than or equal to the on-chip interconnect bandwidth capability contained in the node resource feature vector, then an edge from the task node to the on-chip system node is established in the task-resource mapping graph.

[0089] The task-resource mapping graph includes a set of task nodes, a set of resource nodes, and a set of edges. The set of task nodes represents all tasks to be scheduled, the set of resource nodes represents all on-chip system nodes, and the set of edges describes the mapping relationships between all task nodes and resource nodes that meet resource constraints. Each edge connects a task node and a resource node.

[0090] For each edge in the task-resource mapping graph, construct a composite edge weight vector that includes the expected response latency, average power consumption, and reliability evaluation.

[0091] The expected response latency is calculated by using the ratio between the computational workload of the task processing system domain in the task attribute vector and the number of computational cores in the processing system domain in the node resource feature vector. The expected response latency describes the estimated response time required to complete the task when it is executed on the on-chip system node.

[0092] The average power consumption is obtained based on the processing instruction volume of the task and the node power consumption curve prediction model. The average power consumption describes the average energy consumption required by the system when the task is executed on the on-chip system node.

[0093] The reliability assessment describes the reliability of a task when it is executed on an on-chip system node. The reliability assessment value ranges from zero to one. The reliability assessment is based on manual modeling of the congestion risk between node temperature margin, task bandwidth request and node bandwidth capacity.

[0094] By weighted summing the ratio of expected response delay to task real-time deadline, the ratio of average power consumption to system power consumption limit, and the value after subtracting reliability evaluation, a task-resource matching cost function is constructed.

[0095] The task-resource matching cost function comprehensively measures the performance of tasks allocated to on-chip system nodes in terms of response latency, power consumption, and reliability. Weighted summation factors balance the impact of latency, power consumption, and reliability, and the sum of all weighted factors equals one. A smaller task-resource matching cost function value indicates a better allocation effect.

[0096] Write the task-resource matching cost function value of each pair of tasks and on-chip system nodes into the composite edge weight in the task-resource mapping graph.

[0097] Perform quantum state encoding on the task-resource mapping graph, and encode each task to be scheduled using a set of qubit strings to obtain a set of quantum-encoded tasks.

[0098] In this embodiment, each task to be scheduled is encoded using a set of qubit strings, including:

[0099] Based on the task-resource mapping graph, the total number of tasks in the task node set is determined to be N, the total number of nodes in the on-chip system node set is determined to be K, and the deployment location sets are defined as PS and PL, where PS represents the processing system domain and PL represents the programmable logic domain.

[0100] For each task in the task set, a quantum state code is constructed. The quantum state code consists of two sets of qubits. The first set of qubits represents the target on-chip system node to which the task is assigned, and the second set of qubits represents the deployment domain of the task in the target on-chip system node.

[0101] The deployment domain can be a processing system domain or a programmable logic domain.

[0102] The first group of qubits is indexed and encoded using binary encoding. All on-chip system nodes are numbered, and the node encoding of each task is initialized to a quantum superposition state with equal probability amplitude. All states corresponding to the node numbers have the same probability of being selected at the initial moment.

[0103] ;

[0104] in, Indicates task The initial superposition state of the node qubit group, Indicates node number In the corresponding binary state, all nodes have the same probability amplitude.

[0105] The deployment position of the second qubit group is encoded, and one binary qubit is used to represent the deployment state of the processing system domain and the programmable logic domain respectively. The processing system domain corresponds to 0 and the programmable logic domain corresponds to 1. The sum of the initial distribution probabilities of each task in the two deployment domains is equal to 1.

[0106] The quantum state encoding step is performed on each task separately to generate quantum state codes for all tasks in sequence, forming a set of quantum state codes used to represent the search state space of node allocation and domain deployment of all tasks in the multi-SoC system.

[0107] Initialize the quantum ant colony, randomly generate individual quantum ants, load the quantum coding task set into the search state space of the individual quantum ants, set the pheromone matrix and the quantum rotation angle matrix to obtain the initial population of the quantum ant colony;

[0108] In this embodiment, initializing the quantum ant colony includes:

[0109] The quantum ant colony is initialized for multi-SoC system collaborative tasks based on the quantum state encoding set. The total number of quantum ant individuals is set, and each quantum ant individual corresponds to a set of quantum encoded task copies as the search state representation of the quantum ant individual, thus constructing the initial population of the quantum ant colony.

[0110] Assign a corresponding task search state vector to each individual quantum ant;

[0111] The task search state vector consists of quantum state encoded copies of all tasks. Each quantum state encoded copy of a task includes the task's node qubit set and deployment domain qubit set.

[0112] Construct a pheromone matrix, where each element represents the pheromone intensity when a task is assigned to a specific on-chip system node and deployed in a specific deployment domain.

[0113] The task number, on-chip system node number, and deployment domain number of the pheromone matrix correspond to the task sequence number, on-chip system node sequence number, and deployment domain sequence number, respectively.

[0114] For each qubit of each task, construct a quantum rotation angle matrix. Each element in the quantum rotation angle matrix represents the rotation angle used by a quantum ant for a qubit of a certain task at a certain number of iterations.

[0115] The initial value range of the quantum rotation angle is from the negative maximum rotation angle to the positive maximum rotation angle. The quantum rotation angle matrix is ​​used for quantum rotation gate operations to guide the direction of quantum state evolution.

[0116] The task search state vector, pheromone matrix, and quantum rotation angle matrix of each quantum ant individual are bound together to form a quantum ant individual state structure. All quantum ant individual state structures are integrated to construct the initial population structure of the quantum ant colony.

[0117] An iterative search operation is performed in the initial population of the quantum ant colony. Controlled-domain differential quantum rotation gate operation and cross-task controlled rotation gate operation are performed on the qubit string to obtain the updated quantum state population.

[0118] In this embodiment, controlled-domain differential quantum rotation gate operations and cross-task controlled rotation gate operations are performed on the qubit string, including:

[0119] At the beginning of each iteration round, the operating status of the multi-SoC system is monitored, and the congestion ratio of each on-chip system node is calculated.

[0120] The congestion ratio is equal to the ratio of the actual AXI-DDR bandwidth usage of the on-chip system node to the node's bandwidth capacity. The congestion ratio ranges from zero to one.

[0121] Calculate the adaptive rotational gain based on the current iteration round and thermal state indices;

[0122] The adaptive rotational gain is obtained by multiplying the initial rotational gain value by two adjustment terms. The first adjustment term is the negative exponent corresponding to the ratio of the current iteration round to the maximum number of iterations, and the second adjustment term is the ratio of one unit minus the current average temperature to the maximum allowable operating temperature. The adaptive rotational gain reflects the adjustment range of the quantum rotation gate under the current scheduling environment.

[0123] ;

[0124] in, This is the initial rotational gain. For adaptive rotational gain, The maximum number of iterations is given by t, where t is the current iteration round. For the maximum permissible operating temperature, This is the current average temperature.

[0125] For each individual quantum ant in the quantum ant colony, perform a congestion-suppressing quantum rotation update for the node qubit group:

[0126] ;

[0127] in, The rotation angle of the previous round node qubit. The rotation angle of the updated node qubit. For pheromone intensity, For the task Average pheromones For the first Task in the next iteration Assigned to on-chip system nodes Task—Resource matching cost function value, The optimal matching cost function value, For symbolic functions, This is the congestion ratio;

[0128] For each task's node qubit set and deployment domain qubit set, perform a controlled-domain differential quantum rotation gate operation;

[0129] ;

[0130] Among them, the control bit is the highest bit of the node qubit. , The coefficient of difference between regions. The deadline for the task in real time;

[0131] For task pairs with communication dependencies, if the corresponding element of the dependency matrix is ​​one, then a cross-task controlled rotation gate operation is applied between the node qubit groups of the two tasks:

[0132] ;

[0133] in, For elements in the dependency matrix, if Indicates task With the task There is a direct communication dependency; a value of 0 indicates no dependency. For the first Task in the next iteration and tasks The cross-task qubit coupling rotation angle between them For the first Task in the next iteration and tasks The cross-task qubit coupling rotation angle between them For the dependent adjustment coefficient, For the first Task in the next iteration Assigned to on-chip system nodes Task—Resource matching cost function value, For the task With the task The predicted average communication power consumption required to complete one data communication between nodes on the same on-chip system or between adjacent nodes;

[0134] After completing all controlled-domain differential quantum rotation gate operations and cross-task controlled rotation gate operations, the node qubit group and the deployment domain qubit group of each quantum ant individual are normalized respectively;

[0135] The updated quantum state population is obtained by summing all individual quantum ants. Then, according to the improved pheromone volatilization enhancement rule, the pheromone intensity of each task assigned to each on-chip system node and deployment domain is updated until the termination condition is met, and the updated quantum state population is output.

[0136] The rule for enhancing pheromone volatilization is: the amount of pheromone corresponding to the current pheromone intensity minus a certain proportion of the volatilization factor, plus the reciprocal of the node's optimal matching cost function value in this round multiplied by the volatilization factor.

[0137] The updated quantum state population is subjected to quantum state collapse operation to convert the qubit string into deterministic task allocation results. Based on the node resource profile set, the delay index, power consumption index and reliability index of each deterministic task allocation result are calculated, and a set of candidate feasible solutions that meet the real-time constraints are selected.

[0138] In this embodiment, a set of candidate feasible solutions that satisfy the real-time constraints is selected, including:

[0139] Perform a quantum state collapse operation on the quantum state population updated by the quantum rotation gate operation;

[0140] The quantum state collapse operation refers to the following: For each individual quantum ant, the qubit string of each task in its task search state vector is measured once according to the probability amplitude, and the qubit string is converted into a unique node allocation number and deployment domain number. The node allocation number indicates the number of the on-chip system node to which the task is assigned, and the deployment domain number indicates the position of the task in the processing system domain or programmable logic domain inside the on-chip system node.

[0141] The collapse results of all tasks of each quantum ant individual are combined according to the task number to obtain the complete task allocation decision vector of the quantum ant individual in this round of search;

[0142] The complete task allocation decision vector includes a set of node allocation numbers and deployment domain numbers for all tasks, which fully describes the task scheduling scheme for each quantum ant individual.

[0143] For each quantum ant individual, the complete task allocation decision vector is used to calculate the latency, power consumption and reliability indicators for all tasks based on the resource feature vector of each on-chip system node in the node resource profile set.

[0144] The latency metric represents the actual response time required for each task to complete on the on-chip system node corresponding to its node assignment number. It is obtained by dividing the computational workload of the task's processing system domain by the number of processing system domain cores of the on-chip system node and adding the latency compensation due to communication dependencies.

[0145] The power consumption metric represents the average energy consumption actually consumed by each task on the on-chip system node corresponding to its node allocation number. It is obtained by dividing the computational workload of the task's processing system domain by the number of cores in the processing system domain of the on-chip system node and multiplying by the CPU power consumption mapping coefficient, and then dividing the task's bandwidth requirement by the interconnect bandwidth capacity of the on-chip system node and multiplying by the AXI interconnect power consumption mapping coefficient.

[0146] The reliability index represents the reliability of each task running on the on-chip system node corresponding to its node assignment number. The reliability index is obtained by subtracting the congestion factor from one and the ratio of the current node temperature to the highest operating temperature from one. The value of the reliability index ranges from zero to one.

[0147] For the evaluation results of each task, determine whether the real-time constraints are met;

[0148] In this embodiment, the real-time constraints include:

[0149] Only when the latency of all tasks is less than or equal to the task deadline, the power consumption of all tasks is less than or equal to the maximum allowable average power consumption of the node, and the reliability of all tasks is greater than or equal to the preset reliability threshold, is the complete task allocation decision vector of the corresponding quantum ant individual marked as a candidate feasible solution.

[0150] All complete task assignment decision vectors that meet the real-time constraints are aggregated to form a set of candidate feasible solutions.

[0151] The candidate feasible solution set contains all complete task assignment decision vectors that satisfy the condition that the delay index is less than or equal to the deadline.

[0152] Perform local refinement operations on the candidate feasible solution set, adjust the task-resource mapping relationship according to the current system load status and interconnect bandwidth occupancy status, write the adjusted task allocation results back to the pheromone matrix and update the quantum rotation angle matrix to obtain the global scheduling scheme.

[0153] In this embodiment, a local refinement operation is performed on the candidate feasible solution set, including:

[0154] From the set of candidate feasible solutions, the optimal scheduling result within the current scheduling period is selected as the starting solution for local refinement.

[0155] The local refinement starting solution is defined as a set consisting of the on-chip system node number and deployment domain number of all tasks, which is used as the starting point for local refinement operations.

[0156] By combining the current system's operational status information, the current computing load percentage of the on-chip system node is obtained by summing the ratio of the computing workload of all tasks mapped to the node's processing system domain to the node's processing system domain computing capacity.

[0157] For each task on the allocated on-chip system node, the relative interconnect bandwidth pressure generated when the task is mapped to the on-chip system node is obtained by the ratio of the task's bandwidth requirement to the interconnect bandwidth capability of the on-chip system node.

[0158] Calculate the local adaptation cost for each task;

[0159] In Example 1, the local adaptation cost is obtained by weighting and summing the ratio of the computational load percentage of the on-chip system node in the current task allocation scheme to one percent, the relative interconnect bandwidth pressure generated when the task is mapped to the on-chip system node, and the reliability index of the task minus one. The local adaptation cost reflects the comprehensive performance of the task allocation scheme in terms of current system load, bandwidth pressure and reliability.

[0160] For each task, while keeping other task allocation schemes unchanged in the candidate scheduling neighborhood, all combinations of on-chip system node number and deployment domain number are traversed, the local adaptation cost under each combination is calculated, the on-chip system node number and deployment domain number with the minimum local adaptation cost are selected as the local refinement result of the task, and the original allocation is replaced with the local refinement result.

[0161] Once all tasks have been locally refined, the refinement results of all tasks are combined into a new task allocation decision vector. The pheromone values ​​corresponding to the on-chip system node number and deployment domain number of each task are written back. The locally refined task allocation decision vector is then output as the final global scheduling scheme.

[0162] In Example 1, the pheromone value is written back as follows: based on the pheromone value corresponding to the on-chip system node number and deployment domain number of the same task in the previous round, the product of the pheromone value and the volatility coefficient is subtracted to represent the natural volatility effect of the pheromone. Then, the product of the inverse of the local adaptation cost and the volatility coefficient is added to represent the enhancement effect of the quality of the refinement result on the pheromone. The new pheromone value obtained reflects both the forgetting effect of the historical allocation path and the priority of the refined allocation scheme.

[0163] For the refinement results of each task, the on-chip system node number and deployment domain number are converted into binary code, and the binary code is initialized into a quantum measurement state with a probability of one. The refinement results of all tasks are combined to form a refined quantum state set, which is used to guide the next round of quantum rotation gate operation.

[0164] Example 2: In a large-scale intelligent manufacturing plant, a newly built intelligent production line for automotive electronic controllers has been put into operation. To meet the growing demand for production capacity and flexible production, the production line uses multiple high-performance industrial control units distributed at various key workstations. Each control unit is based on the JYX4S020 SoC platform. Each SoC includes a PS (processing system domain, dual-core ARM Cortex-A9) and a PL (programmable logic domain, FPGA), supporting high-speed DDR3 / DDR3L, Gigabit Ethernet, PCIe and other high-bandwidth interconnects. The workshop business tasks include robotic arm motion planning, AI detection of solder joint quality, real-time image acquisition, adaptive scheduling of process parameters, self-diagnosis of equipment faults, and data synchronization.

[0165] The enterprise employs a traditional static task allocation method, where all tasks are pre-mapped to different SoC units according to fixed priorities and node capabilities. In Example 2, AI inference and image processing tasks are preferentially assigned to nodes with abundant PL resources, while motion control and communication synchronization tasks are assigned to nodes with more idle PS resources. The task allocation results for each node remain unchanged for a long period, with manual adjustments only made during extreme peak periods or node failures. The entire system is generally usable during stable production periods, but uneven resource utilization frequently occurs. Approximately two SoC nodes operate at low loads year-round, while some nodes performing high-density AI tasks experience prolonged high-temperature alarms in their CPUs or PL units. The average resource utilization per node is only around 61%, with a load standard deviation as high as 21%. During peak periods, if a master control node or a high-load node fails, task takeover typically takes more than 5 seconds, accompanied by the loss of several task statuses, requiring on-site intervention by maintenance personnel, severely impacting production efficiency and production line continuity. During peak production line operation, the response time for some tasks even rises to 68 milliseconds, and throughput declines by more than 27% compared to stable periods. During communication and collaborative scheduling among multiple nodes, network bottlenecks or node overload can cause local deadlocks or scheduling conflicts, affecting the overall production rhythm.

[0166] To address the aforementioned challenges, the technical team introduced this invention and conducted a week-long real-world application and comparative test. The new generation scheduling platform first comprehensively profiles the PS and PL resource capabilities of each JYX4S020 SoC node, quantifying the number of processing system cores, logic unit size, bandwidth capacity, and power consumption characteristics. All business tasks are monitored in real-time for features such as task granularity, instruction quantity, latency, and bandwidth requirements. During actual operation, the scheduling platform dynamically reassesses the resource distribution and task execution status of all nodes every 10 seconds, encoding all tasks to be scheduled using qubits in the node number and deployment domain (PS / PL) dimensions. The system utilizes the quantum ant colony algorithm to perform parallel search and optimization of the global mapping space, adaptively quantizing the load, bandwidth, real-time performance, and communication dependencies of each node. Each round of scheduling outputs a set of optimal or suboptimal task allocation schemes and supports task hot migration, automatic takeover of faulty nodes, and task state rollback.

[0167] During a week of continuous production testing, the resource utilization of all SoC nodes significantly improved after adopting the method of this invention, averaging 87%, with the standard deviation of fluctuation reduced to 7%. The balance between nodes under extreme high and low loads was significantly improved, and previously idle nodes were fully activated to participate in collaborative tasks. The number of overload alarms for high-load nodes decreased from 13 per week to 2, resulting in better overall system energy consumption control. During stable production periods, the average response time for each task was reduced from 15.4 milliseconds under the traditional method to 8.9 milliseconds, and the throughput increased from 39.2 megabytes per second to 52.6 megabytes per second, significantly improving production line stability.

[0168] During peak periods, the system experiences high concurrency and dynamic business expansion. Under the original static method, the queue latency of the master node surged, reaching a maximum of 68 milliseconds, causing persistent node overload. After adopting this invention, the system automatically detects node congestion and temperature status, allowing tasks to migrate multiple times within 10 seconds. The load on critical nodes is rapidly balanced, the maximum task latency is only 23 milliseconds, and the throughput decrease is reduced from 27% with the traditional method to only 6%, ensuring the production line continues to operate efficiently during peak periods.

[0169] The most striking contrast occurs in node failure and task migration scenarios. Under traditional static mapping, when the master node fails, backup nodes take an average of 5.3 seconds to take over all tasks, resulting in a task migration failure rate as high as 8.2% and the loss of up to three task states. Multiple instances of manual intervention are required to restore production. However, with this invention, after node failure detection, the quantum ant colony algorithm quickly identifies the optimal migration node and deployment domain for the task and automatically rolls back all task states. The average task takeover time is only 0.9 seconds, the migration failure rate is reduced to less than 1%, and no task states are lost. The entire process requires no manual intervention. Even in the event of large-scale AI task migrations, the reliability indicators of nodes and tasks remain at high thresholds, and the system achieves uninterrupted smooth recovery in 99.1% of scenarios.

[0170] In all experimental samples, taking peak production as an example, with 18 complex task pools, the average task response time under the traditional method was 21.8 milliseconds, and the system throughput was 38.5 megabytes / second; under the method of this invention, the response time was only 10.2 milliseconds, the throughput was increased to 51.6 megabytes / second, the production line's fault-free operation time was improved, and the maintenance and management difficulty was greatly reduced.

[0171] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-SoC collaborative task allocation method based on quantum ant colony optimization, characterized in that, include: In a multi-SoC system, collect node resource profiles and task attribute sets for each on-chip system node; Construct a task-resource mapping graph based on the node resource profile set and the task attribute set; Perform quantum state encoding on the task-resource mapping graph, and encode each task to be scheduled using a set of qubit strings to obtain a set of quantum-encoded tasks. Initialize the quantum ant colony, randomly generate individual quantum ants, load the quantum coding task set into the search state space of the individual quantum ants, set the pheromone matrix and the quantum rotation angle matrix to obtain the initial population of the quantum ant colony; An iterative search operation is performed in the initial population of the quantum ant colony. Controlled-domain differential quantum rotation gate operation and cross-task controlled rotation gate operation are performed on the qubit string to obtain the updated quantum state population. The updated quantum state population is subjected to quantum state collapse operation to convert the qubit string into deterministic task allocation results. Based on the node resource profile set, the delay index, power consumption index and reliability index of each deterministic task allocation result are calculated, and a set of candidate feasible solutions that meet the real-time constraints are selected. Perform local refinement on the candidate feasible solution set, adjust the task-resource mapping relationship according to the current system load status and interconnect bandwidth occupancy status, write the adjusted task allocation result back to the pheromone matrix and update the quantum rotation angle matrix to obtain the global scheduling scheme; The step of encoding each task to be scheduled using a set of qubit strings includes: Based on the task-resource mapping graph, the total number of tasks in the task node set is determined to be N, the total number of nodes in the on-chip system node set is determined to be K, and the deployment location sets are defined as PS and PL, where PS represents the processing system domain and PL represents the programmable logic domain. For each task in the task set, a quantum state code is constructed. The quantum state code consists of two sets of qubits. The first set of qubits represents the target on-chip system node to which the task is assigned, and the second set of qubits represents the deployment domain of the task in the target on-chip system node. The first group of qubits is indexed and encoded using binary encoding. All on-chip system nodes are numbered, and the node encoding of each task is initialized to a quantum superposition state with equal probability amplitude. All states corresponding to the node numbers have the same probability of being selected at the initial moment. The deployment position of the second qubit group is encoded, and one binary qubit is used to represent the deployment state of the processing system domain and the programmable logic domain respectively. The processing system domain corresponds to 0 and the programmable logic domain corresponds to 1. The sum of the initial distribution probabilities of each task in the two deployment domains is equal to 1. Perform the quantum state encoding step for each task separately to generate quantum state codes for all tasks in sequence, forming a set of quantum state codes to represent the search state space of node allocation and domain deployment of all tasks in the multi-SoC system; The initialization of the quantum ant colony includes: The quantum ant colony is initialized for multi-SoC system collaborative tasks based on the quantum state encoding set. The total number of quantum ant individuals is set, and each quantum ant individual corresponds to a set of quantum encoded task copies as the search state representation of the quantum ant individual, thus constructing the initial population of the quantum ant colony. Assign a corresponding task search state vector to each individual quantum ant; Construct a pheromone matrix, where each element represents the pheromone intensity when a task is assigned to a specific on-chip system node and deployed in a specific deployment domain. For each qubit of each task, construct a quantum rotation angle matrix. Each element in the quantum rotation angle matrix represents the rotation angle used by a quantum ant for a qubit of a certain task at a certain number of iterations. The task search state vector, pheromone matrix, and quantum rotation angle matrix of each quantum ant individual are bound together to form a quantum ant individual state structure. All quantum ant individual state structures are integrated to construct the initial population structure of the quantum ant colony. The controlled-domain differential quantum rotation gate operation and cross-task controlled rotation gate operation on the qubit string include: At the beginning of each iteration round, the operating status of the multi-SoC system is monitored, and the congestion ratio of each on-chip system node is calculated. Calculate the adaptive rotational gain based on the current iteration round and thermal state indices; For each individual quantum ant in the quantum ant colony, perform a congestion-suppressing quantum rotation update for the node qubit group: ； in, The rotation angle of the previous round node qubit. The rotation angle of the updated node qubit. For pheromone intensity, For the task Average pheromones For the first Task in the next iteration Assigned to on-chip system nodes Task—Resource matching cost function value, The optimal matching cost function value, For symbolic functions, This is the congestion ratio; For each task's node qubit set and deployment domain qubit set, perform a controlled-domain differential quantum rotation gate operation; ； Among them, the control bit is the highest bit of the node qubit. , The coefficient of difference between regions. The deadline for the task in real time; For task pairs with communication dependencies, if the corresponding element of the dependency matrix is ​​one, then a cross-task controlled rotation gate operation is applied between the node qubit groups of the two tasks: ； in, For elements in the dependency matrix, if Indicates task With the task There is a direct communication dependency; a value of 0 indicates no dependency. For the first Task in the next iteration and tasks The cross-task qubit coupling rotation angle between them For the first Task in the next iteration and tasks The cross-task qubit coupling rotation angle between them For the dependent adjustment coefficient, For the first Task in the next iteration Assigned to on-chip system nodes Task—Resource matching cost function value, For the task With the task The predicted average communication power consumption required to complete one data communication between nodes on the same on-chip system or between adjacent nodes; After completing all controlled-domain differential quantum rotation gate operations and cross-task controlled rotation gate operations, the node qubit group and the deployment domain qubit group of each quantum ant individual are normalized respectively; The updated quantum state population is obtained by summing all individual quantum ants. Then, according to the improved pheromone volatilization enhancement rule, the pheromone intensity of each task assigned to each on-chip system node and deployment domain is updated until the termination condition is met, and the updated quantum state population is output.

2. The multi-SoC collaborative task allocation method based on quantum ant colony optimization according to claim 1, characterized in that, The resource profile set and task attribute set of the data collection nodes include: Resource profiles are collected for each on-chip system node in a multi-SoC system, and node resource feature vectors are constructed. The node resource feature vectors corresponding to each on-chip system node are summarized to form a set of node resource profiles; For each task in the set of tasks to be scheduled, extract attributes and construct a task attribute vector. The task attribute vectors of each task to be scheduled are aggregated to form a task attribute set; The task attribute set describes the resource constraints and demand distribution of all scheduled tasks within a multi-SoC system.

3. The multi-SoC collaborative task allocation method based on quantum ant colony optimization according to claim 2, characterized in that, The construction task-resource mapping graph includes: Based on the set of node resource profiles and the set of task attributes, a task-resource mapping graph is constructed. For each edge in the task-resource mapping graph, construct a composite edge weight vector that includes the expected response latency, average power consumption, and reliability evaluation. By weighted summing the ratio of expected response delay to task real-time deadline, the ratio of average power consumption to system power consumption limit, and the value after subtracting reliability evaluation, a task-resource matching cost function is constructed. Write the task-resource matching cost function value of each pair of tasks and on-chip system nodes into the composite edge weight in the task-resource mapping graph.

4. The multi-SoC collaborative task allocation method based on quantum ant colony optimization according to claim 3, characterized in that, The selection of the candidate feasible solution set that satisfies the real-time constraints includes: Perform a quantum state collapse operation on the quantum state population updated by the quantum rotation gate operation; The collapse results of all tasks of each quantum ant individual are combined according to the task number to obtain the complete task allocation decision vector of the quantum ant individual in this round of search; For each quantum ant individual, the complete task allocation decision vector is used to calculate the latency, power consumption and reliability indicators for all tasks based on the resource feature vector of each on-chip system node in the node resource profile set. For the evaluation results of each task, determine whether the real-time constraints are met; All complete task assignment decision vectors that meet the real-time constraints are aggregated to form a set of candidate feasible solutions.

5. The multi-SoC collaborative task allocation method based on quantum ant colony optimization according to claim 4, characterized in that, The real-time constraints include: Only when the latency of all tasks is less than or equal to the task deadline, the power consumption of all tasks is less than or equal to the maximum allowable average power consumption of the node, and the reliability of all tasks is greater than or equal to the preset reliability threshold, is the complete task allocation decision vector of the corresponding quantum ant individual marked as a candidate feasible solution.

6. The multi-SoC collaborative task allocation method based on quantum ant colony optimization according to claim 5, characterized in that, The local refinement operation on the candidate feasible solution set includes: From the set of candidate feasible solutions, the optimal scheduling result within the current scheduling period is selected as the starting solution for local refinement. By combining the current system's operational status information, the current computing load percentage of the on-chip system node is obtained by summing the ratio of the computing workload of all tasks mapped to the node's processing system domain to the node's processing system domain computing capacity. For each task on the allocated on-chip system node, the relative interconnect bandwidth pressure generated when the task is mapped to the on-chip system node is obtained by the ratio of the task's bandwidth requirement to the interconnect bandwidth capability of the on-chip system node. Calculate the local adaptation cost for each task; For each task, while keeping other task allocation schemes unchanged in the candidate scheduling neighborhood, all combinations of on-chip system node number and deployment domain number are traversed, the local adaptation cost under each combination is calculated, the on-chip system node number and deployment domain number with the minimum local adaptation cost are selected as the local refinement result of the task, and the original allocation is replaced with the local refinement result. Once all tasks have been locally refined, the refinement results of all tasks are combined into a new task allocation decision vector. The pheromone values ​​corresponding to the on-chip system node number and deployment domain number of each task are written back, and the locally refined task allocation decision vector is output as the final global scheduling scheme.

Images