Automatic scheduling system for multi-chip parallel burning

Abstract

The invention relates to the technical field of industrial internet, in particular to an automatic scheduling system for multi-chip parallel burning, which is characterized in that a program optimization and deployment module dynamically adapts a basic program according to hardware resources of a target equipment cluster and process parameters of chips, generates an efficient optimized burning program, and sends the efficient optimized burning program to a server; verifying the digital digest through an integrated block chain network and performing safe deployment; the intelligent scheduling decision module utilizes a dynamic digital twinborn model to perceive the equipment state and deployed optimization program characteristics in real time, and intelligently generates a global burning task optimal allocation scheme and a scheduling instruction; the cooperative control module compares actual efficiency with expected efficiency in real time, quantifies deviation and then feeds back the deviation to the reinforcement learning scheduling engine, and the engine automatically adjusts and continuously iteratively optimizes the overall scheduling strategy. According to the invention, the scheduling efficiency of multi-chip parallel burning can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet, and in particular to an automated scheduling system for parallel burning of multiple chips. Background Art

[0002] High-quality programming of industrial chips is a critical step in the production process. Traditional single-chip manual or semi-automatic programming methods are no longer able to meet the growing demand for production capacity and quality control. Therefore, automated scheduling of multi-chip parallel programming is of great significance.

[0003] At present, traditional multi-chip parallel programming methods often adopt fixed or simple polling control strategies, which makes it difficult to achieve dynamic and optimal resource allocation, resulting in limited overall programming efficiency.

[0004] Therefore, an automatic scheduling system for multi-chip parallel programming is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide an automated scheduling system for multi-chip parallel burning. The program optimization and deployment module dynamically adapts the basic program according to the hardware resources of the target device cluster and the process parameters of the chip, generates an efficient and optimized burning program, and integrates the blockchain network to verify the digital summary and deploy it securely. The intelligent scheduling decision module uses a dynamic digital twin model to perceive the device status and the characteristics of the deployed optimization program in real time, and intelligently generates the optimal allocation plan and scheduling instructions for the global burning task. The collaborative control module compares the actual performance with the expected performance in real time, quantifies the deviation, and then feeds it back to the reinforcement learning scheduling engine, which automatically adjusts and continuously iterates to optimize the overall scheduling strategy. The present invention can improve the scheduling efficiency of multi-chip parallel burning.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An automated scheduling system for multi-chip parallel programming, comprising:

[0008] The program optimization and deployment module is used to dynamically adapt the target code of the basic programming control program based on the current hardware resource configuration of the target programming device cluster and the process parameters of the multiple chips to be programmed. This generates an optimized programming control program, verifies the digital summary of the optimized programming control program through the integrated blockchain network, and securely deploys the verified program to the designated parallel programming devices.

[0009] An intelligent scheduling decision module is used to obtain real-time status data and characteristic parameters of the optimized burning control program through the digital twin model, generate a global burning task allocation plan and scheduling instructions, and send them to each burning device;

[0010] The collaborative control module is used to collect actual execution performance data from each burning device in real time, quantify the deviation between actual performance and expected performance through comparative analysis, and feed the deviation back to the reinforcement learning scheduling engine of the digital twin model to automatically adjust the scheduling strategy.

[0011] Preferably, the hardware resource configuration includes the model of each burning device, internal processor, memory, supported physical interface standards, whether there is an onboard dedicated coprocessor, whether there is an FPGA and the number of parallel channels; the process parameters include the specific model of the chip to be burned, memory type, array structure, data image size, programming voltage, current sequence and allowable programming temperature range.

[0012] Preferably, the process of optimizing the burning control program acquisition includes:

[0013] Select the basic burning algorithm template based on the chip model and process parameters;

[0014] Based on the hardware resource configuration of the target burning device, the data flow and control flow in the algorithm template are parallelized and reorganized and instruction-level optimized;

[0015] Combining the device's current load score and comprehensive health score provided by the digital twin model, the program's resource request level and the strength of the fault-tolerant retry logic are dynamically adjusted to generate an optimized burning control program that includes customized drivers and microcode.

[0016] Preferably, the real-time status data includes: the current working mode of each burning device, the real-time temperature and voltage of each burning channel, the number of chips that have completed burning and the success rate, the current task queue length, the estimated remaining processing time, the network connection quality and the data transmission rate; the characteristic parameters include the expected execution time distribution of the optimized burning control program for specific chip models on different burning devices, the CPU and memory resource usage curve, the maximum number of parallel burning instances, the supported dynamic speed regulation range, and the coverage and efficiency evaluation of its built-in error detection and recovery mechanism.

[0017] Preferably, verifying the digital summary of the optimized burning control program through the integrated blockchain network and securely deploying the verified program to each designated parallel burning device includes:

[0018] When the optimized burning control program is generated, its digital summary is calculated and recorded as an atomic transaction in the blockchain's immutable ledger along with the program version number, applicable hardware configuration fingerprint, chip process parameter fingerprint, and generation timestamp to form a digital identity certificate.

[0019] When the target burning device receives the program to be deployed, it first calculates the digital digest and queries the blockchain network for the existence of the digest and the corresponding digital identity certificate to determine whether it is compatible with its own hardware configuration and the process parameters of the chip to be burned. When the blockchain confirms that the digest is valid and the compatibility matches, the target burning device uses a key agreement mechanism based on asymmetric encryption to establish a secure channel with the deployment source, download and decrypt the optimized burning control program.

[0020] After the deployment is completed, the burning device will record the successfully loaded program version and device identification hash as a receipt again on the blockchain, forming a complete deployment audit chain.

[0021] Preferably, the global burning task allocation scheme and scheduling instruction generation process includes:

[0022] Real-time aggregation of dynamic digital twin model data from all burning devices to form a panoramic view of cluster resources and status;

[0023] For each programming task in the waiting queue, based on its process parameters and the characteristic parameters of the optimized programming control program, preliminarily screen out compatible programming devices;

[0024] A decision network based on a heuristic algorithm with multi-factor weighted evaluation is used to calculate a comprehensive performance score for each compatible device. This score takes into account factors such as expected flash completion time, current device load, health status, task priority, preparation time, and minimization of total cluster energy consumption.

[0025] The final task allocation plan is determined by solving a constrained optimization problem with the goal of maximizing the overall comprehensive efficiency score, and a structured scheduling instruction is generated that includes the specific program version, startup parameters, expected completion window and exception handling plan.

[0026] Preferably, the reinforcement learning scheduling engine includes:

[0027] State space: A high-dimensional feature vector consisting of the encoded dynamic digital twin model data and the key attributes of the task to be scheduled;

[0028] Action space: adjust the execution order of tasks in the task queue and dynamically change the number of concurrent programming channels of the programming device;

[0029] Build a dynamic compound reward function for positive rewards related to the successful completion rate of burning tasks, on-time delivery rate, device utilization balance, overall system throughput and energy efficiency; negative rewards are task failure, scheduling delays and device overload;

[0030] Learning algorithm: Deep Q network is used to learn the optimal scheduling policy function offline.

[0031] Preferably, the automatic adjustment scheduling strategy includes:

[0032] Real-time policy execution: directly converts the current optimal action output by the reinforcement learning scheduling engine into the task allocation logic in the intelligent scheduling decision module;

[0033] Dynamic priority adjustment: Based on the reinforcement learning engine, the dynamic priority of each burning task in the waiting queue is periodically adjusted;

[0034] Adaptive resource allocation: The reinforcement learning engine learns device performance patterns from historical data and real-time feedback, and dynamically adjusts the device resource pool size for different types of burning tasks.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. By performing targeted dynamic adaptation processing on the basic burning control program, the present invention can generate a highly optimized burning control program for each specific chip to be burned and each target burning device. This customized program can maximize the potential of specific hardware and adapt according to the chip's process parameters, ensuring that the program behavior accurately matches the chip's optimal burning conditions, thereby improving the scheduling efficiency of multi-chip parallel burning.

[0037] 2. The present invention utilizes an intelligent scheduling decision-making module and a digital twin model to gain real-time and comprehensive insight into the dynamic operating status of each device in the entire programming equipment cluster. By deeply integrating the complex characteristic parameters of the deployed optimized programming control program, the present invention employs a multi-factor weighted evaluation based on a decision network and an advanced constrained optimization algorithm. This method comprehensively considers multiple dimensions, including the intrinsic priority of the task, the estimated completion time, the health of the device, and the impact of energy consumption. It intelligently generates and issues a global programming task allocation plan and highly structured scheduling instructions, including specific execution program versions, refined startup parameters, expected completion time windows, and pre-set exception handling plans. This enables global, real-time optimal allocation and coordinated management of heterogeneous and dynamically changing programming equipment cluster resources, ensuring a highly intelligent scheduling decision-making process, accurate decision-making results, and proactive execution. This significantly improves the overall operational throughput, comprehensive equipment utilization, and accurate achievement of production plans for the entire parallel programming production system. It also greatly enhances the system's ability to quickly respond to and proactively address various disturbances in the production process, making the overall production process more stable, efficient, and predictable, thereby improving the scheduling efficiency of multi-chip parallel programming.

[0038] 3. The present invention uses a collaborative control module to continuously collect the actual execution performance data of each parallel burning device in the cluster in real time, and conducts accurate and quantitative comparative analysis with the expected performance predicted by the digital twin model, so as to identify specific performance deviations and their characteristics; using the deviation information as the key feedback signal, a reinforcement learning scheduling engine with a dynamic compound reward function is constructed; the engine uses a deep Q network to conduct continuous online learning and strategy iteration on massive actual operation data, and then automatically adjusts and optimizes the scheduling strategy of the entire system, and adaptively changes the number of concurrent burning channels available to the burning device, thereby giving the entire automated scheduling system the ability to self-learn, adapt and continuously self-evolve, thereby improving the scheduling efficiency of multi-chip parallel burning. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of the structure of an automated scheduling system for multi-chip parallel programming provided by an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a flow chart for obtaining an optimized burning control program provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] Example 1

[0043] In order to improve the efficiency of automatic scheduling of multi-chip parallel programming, Factory A applied an automatic scheduling system for multi-chip parallel programming. Figure 1 A schematic structural diagram of an automated scheduling system for multi-chip parallel programming provided by an embodiment of the present invention includes:

[0044] The program optimization and deployment module is used to dynamically adapt the target code of the basic programming control program based on the current hardware resource configuration of the target programming device cluster and the process parameters of the multiple chips to be programmed. This generates an optimized programming control program, verifies the digital summary of the optimized programming control program through the integrated blockchain network, and securely deploys the verified program to the designated parallel programming devices.

[0045] Verification includes: preliminary validity verification of the obtained hardware configuration and chip process parameters, for example, checking whether all required parameters are provided and whether the parameter values ​​fall within the expected reasonable range (such as voltage, temperature, etc.).

[0046] In this embodiment, the cluster size is 10 burning devices.

[0047] Furthermore, the hardware resource configuration includes the model of each burning device, internal processor, memory, supported physical interface standards, whether there is an onboard dedicated coprocessor, whether there is an FPGA and the number of parallel channels; the process parameters include the specific model of the chip to be burned, memory type, array structure, data image size, programming voltage, current sequence and allowable programming temperature range.

[0048] like Figure 2 A schematic diagram of a flow chart for obtaining an optimized burning control program provided by an embodiment of the present invention;

[0049] Furthermore, the optimization process of obtaining the burning control program includes:

[0050] Select the basic burning algorithm template based on the chip model and process parameters;

[0051] Specifically, based on the core technical features of the "chip model and process parameters" specified in the task (such as chip architecture, storage type, and required burning protocol), a rule engine based on feature vector matching is used to intelligently select the most suitable basic algorithm template from the preset template library. This engine characterizes the chip requirements and performs similarity calculation and rule matching with the feature vectors supported by each template declaration.

[0052] Based on the hardware resource configuration of the target burning device, the data flow and control flow in the algorithm template are parallelized and reorganized and instruction-level optimized;

[0053] After selecting a template, the template's data and control flows are parallelized and optimized at the instruction level, tailored to the target device's hardware resource configuration (e.g., number of parallel channels, processor architecture). This includes analyzing program dependencies and automatically applying strategies like data sharding and task-level parallelism to restructure the code to accommodate multi-channel parallelism. Instruction-level optimization is performed using compiler optimization options and instruction sets (e.g., SIMD) specific to the target processor. If the device has an FPGA or coprocessor, standardized APIs for these hardware acceleration functions are integrated into the program, or corresponding configuration files (e.g., FPGA bitstreams) are loaded during initialization to offload compute-intensive or I / O-intensive operations.

[0054] Combining the device's current load score and comprehensive health score provided by the digital twin model, the program's resource request level and the strength of the fault-tolerant retry logic are dynamically adjusted to generate an optimized burning control program that includes customized drivers and microcode.

[0055] Furthermore, the current load score is weighted based on CPU usage, memory usage, and queue length, and the comprehensive health score is weighted based on historical error rate and run time normalized).

[0056] Furthermore, adjustments to resource request levels and the strength of fault-tolerant retry logic are analyzed based on a rule-mapping function; this logic takes device load and health prediction data obtained from the digital twin as input and calculates the adjustments to the program's key behavioral parameters based on the rule-mapping function.

[0057] The resource request level adjustment includes dynamically adjusting the burning rate factor; the burning rate factor is:

[0058] SFY=f(PF,FZ);

[0059] Where SFY represents the burning rate factor; f represents the adjustment rule function; PF represents the comprehensive health score; FZ represents the current load;

[0060] If the health score is lower than 0.7 or the load forecast is higher than 85%, the basic burning rate is multiplied by an adjustment factor less than 1 (such as 0.85) to reduce the resource request intensity.

[0061] Example of adjusting the strength of fault-tolerant retry logic: For example, dynamic adjustment is performed by setting the additional number of retries of the function = g (comprehensive health score).

[0062] If the health score is lower than 0.6, the number of retries will be increased on top of the basic number of retries. The specific number of increases is:

[0063] Count = round((1-PF)*MC);

[0064] Among them, Count represents the number of increases; round() represents the rounding function; PF represents the comprehensive health score; MC represents the maximum number of additional retries allowed;

[0065] This embodiment performs targeted dynamic adaptation processing on the basic burning control program. The system can generate a highly optimized burning control program for each specific chip to be burned and each target burning device. This customized program can maximize the potential of specific hardware and adapt according to the chip's process parameters, ensuring that the program behavior accurately matches the chip's optimal burning conditions, thereby improving the scheduling efficiency of multi-chip parallel burning.

[0066] An intelligent scheduling decision module is used to obtain real-time status data and characteristic parameters of the optimized burning control program through the digital twin model, generate a global burning task allocation plan and scheduling instructions, and send them to each burning device;

[0067] Furthermore, the real-time status data includes: the current working mode of each burning device, the real-time temperature and voltage of each burning channel, the number of chips that have completed burning and the success rate, the current task queue length, the estimated remaining processing time, the network connection quality and the data transmission rate; the characteristic parameters include the expected execution time distribution of the optimized burning control program for specific chip models on different burning devices, the CPU and memory resource usage curve, the maximum number of parallel burning instances, the supported dynamic speed regulation range, and the coverage and efficiency evaluation of its built-in error detection and recovery mechanism.

[0068] Furthermore, verifying the digital summary of the optimized burning control program through the integrated blockchain network and securely deploying the verified program to each designated parallel burning device includes:

[0069] After the optimized burning control program is generated, its digital digest (such as the SHA-256 hash value) is calculated. This digest, along with the program version number, the applicable hardware configuration fingerprint, the chip process parameter fingerprint, and the generation timestamp, is recorded as an atomic transaction in the blockchain's immutable ledger to form a digital identity certificate. When the target burning device receives the program to be deployed, it first calculates the digital digest and queries the blockchain network to determine whether the digest exists and the corresponding digital identity certificate, to determine whether it is compatible with its own hardware configuration and the process parameters of the chip to be burned.

[0070] When the blockchain confirms that the digest is valid and the compatibility matches, the target burning device uses a key negotiation mechanism based on asymmetric encryption to establish a secure channel with the deployment source, download and decrypt the optimized burning control program;

[0071] After the deployment is completed, the burning device will record the successfully loaded program version and device identification hash as a receipt again on the blockchain, forming a complete deployment audit chain.

[0072] Furthermore, the allocation scheme and scheduling instruction generation process of the global burning task includes:

[0073] Real-time aggregation of dynamic digital twin model data from all burning devices to form a panoramic view of cluster resources and status;

[0074] For each programming task in the waiting queue, preliminarily screen compatible programming devices based on its process parameters (such as chip model, programming voltage) and the characteristic parameters of the optimized programming control program (such as the maximum number of parallel programming instances);

[0075] A decision network based on a heuristic algorithm with multi-factor weighted evaluation is used to calculate a comprehensive performance score for each compatible device. This score takes into account factors such as expected flash completion time, current device load, health status, task priority, preparation time, and minimization of total cluster energy consumption.

[0076] The comprehensive effectiveness score is obtained by normalizing and weighting each scoring factor;

[0077] The overall performance score indicates the performance of a single burn scheduling task;

[0078] The final task allocation plan is determined by solving a constrained optimization problem with the goal of maximizing the overall comprehensive efficiency score, and a structured scheduling instruction is generated that includes the specific program version, startup parameters, expected completion window and exception handling plan.

[0079] The overall comprehensive performance score represents the comprehensive scheduling performance assigned to all burning scheduling tasks;

[0080] This embodiment utilizes an intelligent scheduling decision-making module and a digital twin model to gain real-time, comprehensive insight into the dynamic operating status of each device in the entire programming equipment cluster. By deeply integrating the complex characteristic parameters of the deployed optimized programming control program, this embodiment employs a multi-factor weighted evaluation based on a decision network and an advanced constrained optimization algorithm. This method comprehensively considers multiple dimensions, including the intrinsic priority of tasks, estimated completion time, device health, and energy consumption impact. It intelligently generates and issues a global programming task allocation plan and highly structured scheduling instructions, including specific execution program versions, refined startup parameters, expected completion time windows, and pre-set exception handling plans. This enables global, real-time optimal allocation and collaborative management of heterogeneous and dynamically changing programming equipment cluster resources, ensuring a highly intelligent scheduling decision-making process, accurate decision results, and proactive execution. This significantly improves the overall operational throughput, comprehensive equipment utilization, and accurate achievement of production plans for the entire parallel programming production system. It also greatly enhances the system's ability to quickly respond to and proactively address various disturbances in the production process, making the overall production process more stable, efficient, and predictable, thereby improving the scheduling efficiency of multi-chip parallel programming.

[0081] The collaborative control module is used to collect actual execution performance data from each burning device in real time, quantify the deviation between actual performance and expected performance through comparative analysis, and feed the deviation back to the reinforcement learning scheduling engine of the digital twin model to automatically adjust the scheduling strategy.

[0082] The actual calculation process of the deviation is:

[0083] S = UA;

[0084] Among them, S represents deviation; U represents actual performance score; A represents expected performance score;

[0085] Furthermore, the reinforcement learning scheduling engine includes:

[0086] State space: A high-dimensional feature vector consisting of the encoded dynamic digital twin model data and the key attributes of the task to be scheduled;

[0087] Furthermore, if the deviation is negative, it means that the current strategy does not fully utilize the device resources. The reinforcement learning scheduling engine will reduce the device load and increase the number of concurrent recording channels. If the deviation is positive, it means that the task execution exceeds expectations. The system will automatically optimize resource allocation and the reinforcement learning scheduling engine will reduce the number of concurrent recording channels.

[0088] Action space: adjust the execution order of tasks in the task queue and dynamically change the number of concurrent programming channels of the programming device;

[0089] Specifically, it adjusts the order of tasks based on task priority, the device's current load, and other information to ensure that high-priority tasks are executed first, avoiding device idleness or overload. It also dynamically adjusts the number of concurrent programming channels based on device load and performance feedback. For example, if the device is overloaded, the number of concurrent channels can be reduced to avoid overload; if the device is lightly loaded, the number of concurrent channels can be increased to improve throughput.

[0090] Build a dynamic compound reward function for positive rewards related to the successful completion rate of burning tasks, on-time delivery rate, device utilization balance, overall system throughput and energy efficiency; negative rewards are task failure, scheduling delays and device overload;

[0091] Learning algorithm: Deep Q network is used to learn the optimal scheduling policy function offline.

[0092] Furthermore, the automatic adjustment scheduling strategy includes:

[0093] Real-time policy execution: directly converts the current optimal action output by the reinforcement learning scheduling engine into the task allocation logic in the intelligent scheduling decision module;

[0094] Dynamic priority adjustment: Based on the reinforcement learning engine, the dynamic priority of each burning task in the waiting queue is periodically adjusted;

[0095] Adaptive resource allocation: The reinforcement learning engine learns device performance patterns from historical data and real-time feedback, and dynamically adjusts the device resource pool size for different burning tasks.

[0096] This embodiment uses a collaborative control module to continuously collect the actual execution performance data of each parallel burning device in the cluster in real time, and conducts precise and quantitative comparative analysis with the expected performance predicted by the digital twin model, thereby identifying specific performance deviations and their characteristics; using this structured deviation information as a key feedback signal, a reinforcement learning scheduling engine with a dynamically configured composite reward function is developed; this engine uses a deep Q network to conduct continuous online learning and strategy iteration on massive amounts of actual operating data, and then automatically adjusts and optimizes the scheduling strategy of the entire system, adaptively changing the number of concurrent burning channels available to the burning device, thereby giving the entire automated scheduling system the ability to self-learn, adapt, and continuously self-evolve, thereby improving the scheduling efficiency of multi-chip parallel burning.

[0097] Example 2

[0098] In order to improve the efficiency of automatic scheduling of multi-chip parallel programming, Factory B applied an automatic scheduling system for multi-chip parallel programming. Figure 1 A schematic structural diagram of an automated scheduling system for multi-chip parallel programming provided by an embodiment of the present invention includes:

[0099] The program optimization and deployment module is used to dynamically adapt the target code of the basic programming control program based on the current hardware resource configuration of the target programming device cluster and the process parameters of the multiple chips to be programmed. This generates an optimized programming control program, verifies the digital summary of the optimized programming control program through the integrated blockchain network, and securely deploys the verified program to the designated parallel programming devices.

[0100] Furthermore, the hardware resource configuration includes the model of each burning device, internal processor, memory, supported physical interface standards, whether there is an onboard dedicated coprocessor, whether there is an FPGA and the number of parallel channels; the process parameters include the specific model of the chip to be burned, memory type, array structure, data image size, programming voltage, current sequence and allowable programming temperature range.

[0101] Furthermore, the optimization process of obtaining the burning control program includes:

[0102] Select the basic burning algorithm template based on the chip model and process parameters;

[0103] Based on the hardware resource configuration of the target burning device, the data flow and control flow in the algorithm template are parallelized and reorganized and instruction-level optimized;

[0104] Combining the device's current load score and comprehensive health score provided by the digital twin model, the program's resource request level and the strength of the fault-tolerant retry logic are dynamically adjusted to generate an optimized burning control program that includes customized drivers and microcode.

[0105] This embodiment performs targeted dynamic adaptation processing on the basic burning control program. The system can generate a highly optimized burning control program for each specific chip to be burned and each target burning device. This customized program can maximize the potential of specific hardware and adapt according to the chip's process parameters, ensuring that the program behavior accurately matches the chip's optimal burning conditions, thereby improving the scheduling efficiency of multi-chip parallel burning.

[0106] An intelligent scheduling decision module is used to obtain real-time status data and characteristic parameters of the optimized burning control program through the digital twin model, generate a global burning task allocation plan and scheduling instructions, and send them to each burning device;

[0107] Furthermore, the real-time status data includes: the current working mode of each burning device, the real-time temperature and voltage of each burning channel, the number of chips that have completed burning and the success rate, the current task queue length, the estimated remaining processing time, the network connection quality and the data transmission rate; the characteristic parameters include the expected execution time distribution of the optimized burning control program for specific chip models on different burning devices, the CPU and memory resource usage curve, the maximum number of parallel burning instances, the supported dynamic speed regulation range, and the coverage and efficiency evaluation of its built-in error detection and recovery mechanism.

[0108] Furthermore, verifying the digital summary of the optimized burning control program through the integrated blockchain network and securely deploying the verified program to each designated parallel burning device includes:

[0109] When the optimized burning control program is generated, its digital summary is calculated and recorded as an atomic transaction in the blockchain's immutable ledger along with the program version number, applicable hardware configuration fingerprint, chip process parameter fingerprint, and generation timestamp to form a digital identity certificate.

[0110] When the target burning device receives the program to be deployed, it first calculates the digital digest and queries the blockchain network for the existence of the digest and the corresponding digital identity certificate to determine whether it is compatible with its own hardware resource configuration and the process parameters of the chip to be burned. When the blockchain confirms that the digest is valid and the compatibility matches, the target burning device uses a key agreement mechanism based on asymmetric encryption to establish a secure channel with the deployment source, download and decrypt the optimized burning control program.

[0111] After the deployment is completed, the burning device will record the successfully loaded program version and device identification hash as a receipt again on the blockchain, forming a complete deployment audit chain.

[0112] Furthermore, the allocation scheme and scheduling instruction generation process of the global burning task includes:

[0113] Real-time aggregation of dynamic digital twin model data from all burning devices to form a panoramic view of cluster resources and status;

[0114] For each programming task in the waiting queue, based on its process parameters and the characteristic parameters of the optimized programming control program, preliminarily screen out compatible programming devices;

[0115] A decision network based on a heuristic algorithm with multi-factor weighted evaluation is used to calculate a comprehensive performance score for each compatible device. This score takes into account factors such as expected flash completion time, current device load, health status, task priority, preparation time, and minimization of total cluster energy consumption.

[0116] The final task allocation plan is determined by solving a constrained optimization problem with the goal of maximizing the overall comprehensive efficiency score, and a structured scheduling instruction is generated that includes the specific program version, startup parameters, expected completion window and exception handling plan.

[0117] This embodiment utilizes an intelligent scheduling decision-making module and a digital twin model to gain real-time, comprehensive insight into the dynamic operating status of each device in the entire programming equipment cluster. By deeply integrating the complex characteristic parameters of the deployed optimized programming control program, this embodiment employs a multi-factor weighted evaluation based on a decision network and an advanced constrained optimization algorithm. This method comprehensively considers multiple dimensions, including the intrinsic priority of tasks, estimated completion time, device health, and energy consumption impact. It intelligently generates and issues a global programming task allocation plan and highly structured scheduling instructions, including specific execution program versions, refined startup parameters, expected completion time windows, and pre-set exception handling plans. This enables global, real-time optimal allocation and collaborative management of heterogeneous and dynamically changing programming equipment cluster resources, ensuring a highly intelligent scheduling decision-making process, accurate decision results, and proactive execution. This significantly improves the overall operational throughput, comprehensive equipment utilization, and accurate achievement of production plans for the entire parallel programming production system. It also greatly enhances the system's ability to quickly respond to and proactively address various disturbances in the production process, making the overall production process more stable, efficient, and predictable, thereby improving the scheduling efficiency of multi-chip parallel programming.

[0118] The collaborative control module is used to collect actual execution performance data from each burning device in real time, quantify the deviation between actual performance and expected performance through comparative analysis, and feed the deviation back to the reinforcement learning scheduling engine of the digital twin model to automatically adjust the scheduling strategy.

[0119] Furthermore, the reinforcement learning scheduling engine includes:

[0120] State space: A high-dimensional feature vector consisting of the encoded dynamic digital twin model data and the key attributes of the task to be scheduled;

[0121] Action space: adjust the execution order of tasks in the task queue and dynamically change the number of concurrent programming channels of the programming device;

[0122] Build a dynamic compound reward function for positive rewards related to the successful completion rate of burning tasks, on-time delivery rate, device utilization balance, overall system throughput and energy efficiency; negative rewards are task failure, scheduling delays and device overload;

[0123] Learning algorithm: Deep Q network is used to learn the optimal scheduling policy function offline.

[0124] Furthermore, the automatic adjustment scheduling strategy includes:

[0125] Real-time policy execution: directly converts the current optimal action output by the reinforcement learning scheduling engine into the task allocation logic in the intelligent scheduling decision module;

[0126] Dynamic priority adjustment: Based on the reinforcement learning engine's judgment of the current system status and future trends, the dynamic priority of each burning task in the waiting queue is periodically adjusted;

[0127] Adaptive resource allocation: The reinforcement learning engine learns device performance patterns from historical data and real-time feedback, and dynamically adjusts the device resource pool size for different types of burning tasks.

[0128] This embodiment uses a collaborative control module to continuously collect the actual execution performance data of each parallel burning device in the cluster in real time, and conducts precise and quantitative comparative analysis with the expected performance predicted by the digital twin model, thereby identifying specific performance deviations and their characteristics; using deviation information as a key feedback signal, a reinforcement learning scheduling engine with a dynamically configured composite reward function is implemented; this engine uses a deep Q network to conduct continuous online learning and strategy iteration on massive amounts of actual operating data, and then automatically adjusts and optimizes the scheduling strategy of the entire system, adaptively changing the number of concurrent burning channels available to the burning device, thereby giving the entire automated scheduling system the ability to self-learn, adapt, and continuously self-evolve, thereby improving the scheduling efficiency of multi-chip parallel burning.

[0129] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An automated scheduling system for multi-chip parallel programming, characterized in that: include: The program optimization and deployment module is used to dynamically adapt the target code of the basic programming control program based on the current hardware resource configuration of the target programming device cluster and the process parameters of the multiple chips to be programmed. This generates an optimized programming control program, verifies the digital summary of the optimized programming control program through the integrated blockchain network, and securely deploys the verified program to the designated programming device. An intelligent scheduling decision module is used to obtain real-time status data and characteristic parameters of the optimized burning control program through the digital twin model, generate a global burning task allocation plan and scheduling instructions, and send them to each burning device; The collaborative control module is used to collect actual execution performance data from each burning device in real time, quantify the deviation between actual performance and expected performance through comparative analysis, and feed the deviation back to the reinforcement learning scheduling engine of the digital twin model to automatically adjust the scheduling strategy.

2. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: The hardware resource configuration includes the model of each burning device, internal processor, memory, supported physical interface standards, whether there is an onboard dedicated coprocessor, whether there is an FPGA and the number of parallel channels; the process parameters include the specific model of the chip to be burned, memory type, array structure, data image size, programming voltage, current sequence and allowable programming temperature range.

3. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: The optimized burning control program acquisition process includes: Select the basic burning algorithm template based on the chip model and process parameters; Based on the hardware resource configuration of the target burning device, the data flow and control flow in the algorithm template are parallelized and reorganized and instruction-level optimized; Combining the device's current load score and comprehensive health score provided by the digital twin model, the program's resource request level and the strength of the fault-tolerant retry logic are dynamically adjusted to generate an optimized burning control program that includes customized drivers and microcode.

4. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: The real-time status data includes: the current working mode of each burning device, the real-time temperature and voltage of each burning channel, the number of chips that have completed burning and the success rate, the current task queue length, the estimated remaining processing time, the network connection quality and the data transmission rate; the characteristic parameters include the expected execution time distribution of the optimized burning control program for specific chip models on different burning devices, the CPU and memory resource usage curve, the maximum number of parallel burning instances, the supported dynamic speed regulation range, and the coverage and efficiency evaluation of its built-in error detection and recovery mechanism.

5. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: Verifying the digital summary of the optimized burning control program through the integrated blockchain network and securely deploying the verified program to each designated parallel burning device includes: When the optimized burning control program is generated, its digital summary is calculated and the summary, program version number, applicable hardware configuration fingerprint, chip process parameter fingerprint, and generation timestamp are recorded as atomic transactions in the blockchain's immutable ledger to form a digital identity certificate; When the target burning device receives the program to be deployed, it first calculates the digital digest and queries the blockchain network for the existence of the digest and the corresponding digital identity certificate to determine whether it is compatible with its own hardware configuration and the process parameters of the chip to be burned; When the blockchain confirms that the digest is valid and the compatibility matches, the target burning device uses a key negotiation mechanism based on asymmetric encryption to establish a secure channel with the deployment source, download and decrypt the optimized burning control program; After the deployment is completed, the burning device will record the successfully loaded program version and device identification hash as a receipt again on the blockchain, forming a complete deployment audit chain.

6. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: The allocation scheme and scheduling instruction generation process of the global burning task include: Real-time aggregation of dynamic digital twin model data from all burning devices; For each programming task in the waiting queue, based on its process parameters and the characteristic parameters of the optimized programming control program, preliminarily screen out compatible programming devices; A decision network based on a heuristic algorithm with multi-factor weighted evaluation is used to calculate a comprehensive performance score for each compatible device. The score takes into account factors such as expected flash completion time, current device load, health status, task priority, preparation time, and minimization of total cluster energy consumption. The final task allocation plan is determined by solving a constrained optimization problem with the goal of maximizing the overall comprehensive efficiency score, and a structured scheduling instruction is generated that includes the specific program version, startup parameters, expected completion window and exception handling plan.

7. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: The reinforcement learning scheduling engine includes: State space: A high-dimensional feature vector consisting of the encoded dynamic digital twin model data and the key attributes of the task to be scheduled; Action space: adjust the execution order of tasks in the task queue and dynamically change the number of concurrent programming channels of the programming device; Build a dynamic compound reward function for positive rewards related to the successful completion rate of burning tasks, on-time delivery rate, device utilization balance, overall system throughput and energy efficiency; negative rewards are task failure, scheduling delays and device overload; Learning algorithm: Deep Q network is used to learn the optimal scheduling policy function offline.

8. The automated scheduling system for multi-chip parallel programming according to claim 1, characterized in that: The automatic adjustment scheduling strategy includes: Real-time policy execution: Converts the current optimal action output by the reinforcement learning scheduling engine into task allocation logic in the intelligent scheduling decision module; Dynamic priority adjustment: Based on the reinforcement learning engine, the dynamic priority of each burning task in the waiting queue is periodically adjusted; Adaptive resource allocation: The reinforcement learning engine learns device performance patterns from historical data and real-time feedback, and dynamically adjusts the device resource pool size for different types of burning tasks.

Images