Heterogeneous computing power resource dynamic scheduling system and method based on multi-objective optimization

Abstract

The invention discloses a heterogeneous computing power resource dynamic scheduling system and method based on multi-objective optimization, and the system is deployed in a digital ecological platform complex based on multivariate consensus and embedded intelligent management. Comprising a platform access module, a data acquisition and perception module, a multi-target modeling and optimization module, a dynamic scheduling and execution module and a fault-tolerant mechanism module. The system registers as a computing power scheduling service node through an intelligent contract interface, obtains and verifies the compliance of a computing task, collects heterogeneous computing power resource node state data, and generates a scheduling decision by dynamically adjusting the weight of a target function; evaluating task suitability by using an AMCU suitability scoring model, and executing a scheduling AMCU strategy; according to the method, the technical problems of low resource scheduling efficiency and poor fault-tolerant capability in a heterogeneous computing power environment are solved, and the system throughput and the resource utilization rate are improved.

Description

Technical Field

[0001] This invention relates to the fields of distributed computing, resource scheduling and artificial intelligence, and in particular to a dynamic scheduling system and method for heterogeneous computing resources based on multi-objective optimization, implemented in a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance. Background Technology

[0002] With the development of cloud computing, edge computing, and heterogeneous computing technologies, current computing infrastructure includes various heterogeneous computing powers such as central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and analog matrix computing units (AMCUs). These diverse computing units constitute a heterogeneous computing resource pool, enabling computing resources to flexibly adapt to different computing needs. Traditional scheduling systems, such as rule-based or single-objective optimization schedulers, often face the following technical problems when handling mixed workloads:

[0003] First, it is difficult to achieve a dynamic balance among multiple conflicting objectives such as performance, energy consumption, and cost. Second, the lack of awareness of the specialized computing modes of new computing units such as AMCU leads to inaccurate resource matching. Third, in cross-domain collaborative environments such as digital ecosystem platform consortia based on multi-party consensus and embedded intelligent governance, the traditional scheduling system architecture is disconnected from the platform-level governance system, failing to effectively utilize the platform's global resource catalog, dynamic resource type library, and embedded supervisory nodes (iRegNode). This results in isolated scheduling decisions, making it difficult to perform cross-platform collaborative optimization while meeting global policy constraints, and also limiting the maximization of the performance of new computing units such as AMCU. Summary of the Invention

[0004] To address the challenges of existing technologies in achieving dynamic balance among conflicting objectives such as performance, energy consumption, and cost when handling mixed workloads, the lack of awareness of specialized computing modes of new computing units like AMCUs, and the inability to meet global compliance policy constraints in real time in cross-platform collaborative scenarios, this paper proposes a multi-objective optimization-based dynamic scheduling system and method for heterogeneous computing resources. This system aims to improve resource utilization, optimize energy efficiency, enhance scheduling efficiency and collaboration, improve system reliability, and ensure compliance.

[0005] According to a first aspect of the present invention, a heterogeneous computing resource dynamic scheduling system based on multi-objective optimization is provided. The system is deployed within a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance, comprising: The platform access module is used to interact with the consortium consensus network and iRegNode data in the platform; call the service node registration smart contract interface provided by the platform to register the scheduling system as a computing power scheduling service node; obtain computing tasks that have passed the iRegNode audit, call the iRegNode interface to perform compliance verification on the computing tasks, and send the verified computing tasks to the dynamic scheduling and execution module. The data acquisition and sensing module is used to collect status data of heterogeneous computing power resource nodes in the global resource directory published by the platform; The multi-objective modeling and optimization module is used to obtain multi-objective optimization functions based on deep reinforcement learning models and generate initial scheduling decisions by dynamically adjusting the objective weights. The dynamic scheduling and execution module is used to parse the verified computing tasks, extract feature vectors to form feature fingerprints, evaluate the feature fingerprints using a preset AMCU adaptability scoring model, and generate a scheduling decision to prioritize scheduling to the AMCU if the adaptability score exceeds a preset threshold. When executing the scheduling decision, if the target AMCU node fails, the scheduling is performed with a lower priority.

[0006] Furthermore, it also includes a fault tolerance mechanism module, which is used to determine whether there is a fault in the status of heterogeneous computing power resource nodes. If a fault is found, a rescheduling process is triggered, which is as follows: The node failure event is reported to iRegNode for inspection and confirmation. iRegNode marks the confirmed faulty node as faulty or isolated and updates the global resource catalog through the consortium consensus network. Query the computing task allocation record, identify the faulty computing task, add the faulty computing task to the scheduling queue, re-execute the scheduling process, generate a new scheduling decision, allocate the computing task to the standby node, and resume execution from the breakpoint or the starting point. An audit log is generated and sent to the consortium consensus network via iRegNode for evidence storage.

[0007] According to a second aspect of the present invention, a method for dynamic scheduling of heterogeneous computing resources based on multi-objective optimization is provided. The method is implemented using a scheduling system deployed in a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance, and includes the following steps: S1. The platform access module calls the smart contract registration interface provided by the platform to register the scheduling system as a computing power scheduling service node. S2. Obtain the computing tasks audited by iRegNode in the platform, call the iRegNode interface to perform compliance verification on the computing tasks, and send the computing tasks that pass the verification to the dynamic scheduling and execution module. S3. The data acquisition and sensing module collects the status data of heterogeneous computing power resource nodes in the global resource directory published by the platform. S4. The dynamic scheduling and execution module parses the verified computational tasks, extracts feature vectors to form feature fingerprints, and evaluates the feature fingerprints using a preset AMCU adaptability scoring model. If the adaptability score exceeds a preset threshold, a scheduling decision is generated to prioritize scheduling to the AMCU. When executing the scheduling decision, if the target AMCU node fails, a lower priority scheduling is performed. S5. The multi-objective modeling and optimization module obtains a multi-objective optimization function based on a deep reinforcement learning model, and generates an initial scheduling decision by dynamically adjusting the objective weights. S6. The platform access module calls the iRegNode interface to determine whether the initial scheduling decision is compliant. If not, it returns to step S5 for correction. S7, the dynamic scheduling and execution module, makes the final compliant scheduling decision and allocates computing tasks to matching heterogeneous computing resource nodes.

[0008] This invention's technical dependency declaration: The effectiveness and completeness of this invention rely on the consortium consensus network, iRegNode, and data-value closed-loop control protocol defined in the digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance. Without the platform's ability to provide real-time compliance verification, global state consistency, and trusted auditing, the technical solution provided by this invention cannot achieve the cross-domain collaboration and dynamic optimization effects described herein.

[0009] Specifically: First, the scheduling system of this invention needs to undergo real-time compliance verification through iRegNode in the platform to ensure that scheduling decisions comply with the dynamic governance strategy. Second, the global heterogeneous computing power resource status information required by this invention comes from the dynamic resource type library established and maintained by the platform. Third, the scheduling audit logs and performance contribution data generated by this invention need to be encapsulated through the data-value closed-loop control protocol defined by the platform and submitted to the consortium consensus network in the platform for notarization and circulation in order to participate in the platform's value distribution cycle.

[0010] Therefore, the technical solution provided by this invention is a specific function optimization system and method that runs on the consortium infrastructure defined by the digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance. Together, they constitute a complete trusted collaborative computing power scheduling solution.

[0011] Compared with the prior art, the present invention has the following significant advantages: 1. By using a multi-objective deep reinforcement learning model and an AMCU adaptability scoring model, we avoid the waste of resources and performance degradation caused by forcibly scheduling computing tasks that are not suitable for AMCU to AMCU. At the same time, we ensure that these computing tasks can find the most suitable other heterogeneous computing resources, thereby improving the overall resource utilization efficiency and task completion rate of the system. 2. By introducing a compliance verification mechanism that links with the iRegNode within the platform, the system ensures that scheduling failures and efficiency losses due to non-compliance are avoided while strictly adhering to platform governance and security policies. This improves the robustness and availability of the system, especially in cross-domain collaborative environments, where the invention can effectively address complex compliance challenges. 3. The fault tolerance mechanism module ensures the authority and global consistency of fault information. As an embedded supervisory node, iRegNode can quickly verify the status of computing nodes and update the global computing resource catalog in real time through the consortium consensus network. This can quickly prevent new tasks from being scheduled to faulty nodes, avoid fault propagation and resource waste, and thus improve the reliability and fault recovery efficiency of the system. Attached Figure Description

[0012] Figure 1 This invention provides a schematic diagram of a heterogeneous computing resource dynamic scheduling system architecture based on multi-objective optimization, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a method for dynamic scheduling of heterogeneous computing resources based on multi-objective optimization, provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0014] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0015] To more clearly illustrate the embodiments of this application, the technical terms used in the embodiments will be briefly introduced below: A cross-domain trusted collaborative computing platform, namely a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance, includes at least service nodes, iRegNodes, a consortium consensus network, and a data-value closed-loop control protocol. Service nodes are used for identity registration and service discovery for cross-platform transactions via smart contracts; iRegNodes are used for risk assessment of cross-platform transactions, generating signed audit trails, and verifying compliance; the consortium consensus network is jointly maintained by service nodes and iRegNodes; and the data-value closed-loop control protocol is used for the transfer, processing, and value-added functions among heterogeneous service nodes, and for completing value distribution.

[0016] The dynamic resource type library is used to provide a unified standard for resource classification and description for the platform. Its content is the type definition of various heterogeneous computing resources supported by the platform, which includes at least the type definitions of central processing units (CPU), graphics processing units (GPU), field-programmable gate arrays (FPGA), processing units (PE), quantum processing units (QPU), and analog matrix computing units (AMCU).

[0017] The global resource catalog is based on the type standard of the dynamic resource type library and includes real-time status information of each specific heterogeneous computing power resource computing node in the platform. Specifically, it includes node ID, resource type (corresponding type library), IP address, current load, available computing power, health status, compliance status, etc.

[0018] The Analog Matrix Computing Unit (AMCU) is a hardware unit designed for efficient matrix operations, utilizing analog computing principles to process continuous signals.

[0019] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Example 1 To address the problems existing in the prior art, this invention provides a dynamic scheduling system for heterogeneous computing resources based on multi-objective optimization. The system is deployed within a consortium of digital ecosystem platforms based on multi-party consensus and embedded intelligent governance, serving as a computing power scheduling service node with specific functions on the platform. (See also...) Figure 1 The system includes a platform access module, a data acquisition and sensing module, a multi-objective modeling and optimization module, and a dynamic scheduling and execution module.

[0021] The platform access module is used to interact with the consortium consensus network and iRegNode data within the platform. This module calls the service node registration smart contract interface provided by the platform to register the scheduling system as a computing power scheduling service node, establishing the system's identity and service permissions within the platform; it obtains computing tasks that have passed iRegNode auditing, calls the iRegNode interface to perform compliance verification on the computing tasks, and sends the verified computing tasks to the dynamic scheduling and execution module.

[0022] Specifically, after the platform access module registers as a computing power scheduling service node, it performs service registration and authentication, task acquisition and auditing, and data reporting and evidence storage processes.

[0023] The specific workflow for service registration and authentication is as follows: When the system starts, the platform access module prepares registration information including node identity, compliance protocol version, and service capability declaration. Then, it calls the service node registration smart contract interface provided by the platform, and the system completes its registration as a computing power scheduling service node. After successful registration, the system obtains the corresponding access credentials from the platform, establishing its identity and service permissions within the platform. These credentials serve as the basis for all subsequent interactions with the platform.

[0024] The service capability declaration may specifically include service type, such as computing power scheduling and function description, which supports AMCU adaptability scoring and multi-objective optimization scheduling; the compliance protocol version can be understood as the compliance protocol version followed by the system, such as the version number of the data-value closed-loop control protocol.

[0025] The specific workflow for task acquisition and auditing is as follows: Computational tasks that have been audited by iRegNode are periodically retrieved from the platform's task queue. The received tasks are parsed and their formats validated. A secondary compliance check is then performed by calling the iRegNode interface to verify whether the task's data source, computational content, and output requirements comply with the platform's governance standards. Computational tasks that pass the check are encapsulated into a standard format and sent to the dynamic scheduling and execution module for further processing. For computational tasks that fail the check, the reason for the failure is recorded and stored on the platform.

[0026] The specific workflow for data reporting and notarization is as follows: all scheduling decisions, task execution status, and heterogeneous computing resource status data must be encapsulated into standardized messages through the data-value closed-loop control protocol and then submitted to the consortium consensus network for notarization to ensure the credibility and consistency of the global heterogeneous computing resource status.

[0027] The data acquisition and sensing module is used to collect the status data of heterogeneous computing power resource nodes in the global resource directory published by the platform.

[0028] Specifically, the data acquisition and sensing module periodically accesses various computing resource nodes, including CPUs, GPUs, AMCUs, FPGAs, PEs, and QPUs, through a polling mechanism to collect key indicators such as node computing load, memory utilization, network latency, energy consumption, and availability status. The data acquisition and sensing module then standardizes and filters outliers from the collected raw data to form a unified resource status dataset, providing an accurate data foundation for subsequent scheduling decisions.

[0029] The multi-objective modeling and optimization module is used to obtain multi-objective optimization functions based on deep reinforcement learning models and generate initial scheduling decisions by dynamically adjusting the objective weights.

[0030] Specifically, the multi-objective modeling and optimization module is the core of the scheduling decision. It derives a multi-objective optimization function based on a deep reinforcement learning model. Maximizing system throughput, maximizing resource utilization, maximizing energy efficiency, and minimizing average task latency are the main optimization objectives. The module learns optimal strategy patterns from historical scheduling data through a neural network. Based on the current platform load and business priority requirements, the module dynamically adjusts the weight coefficients of each objective, flexibly switching between performance-first, energy-saving-first, and balanced modes. Based on real-time resource status data and task characteristics, the module calculates the Pareto optimal solution set, generating an initial scheduling decision that includes resource allocation schemes and execution sequences.

[0031] The reward function R of a deep reinforcement learning model is defined as: R = w1·Perf + w2·Util + w3·Eneg + w4·Comp; Among them, Perf represents the normalized performance score, Util represents the resource utilization rate, Eneg represents the reciprocal of the energy efficiency score, Comp represents the compliance score, and the weight coefficients w1-w4 can be dynamically adjusted through the data-value closed-loop control protocol in the platform to adapt to the optimization focus at different times.

[0032] After generating the initial decision, the iRegNode's compliance verification interface is immediately invoked to verify the decision. If it does not comply with the current platform policy, a correction mechanism is triggered to re-optimize until a compliant final scheduling decision is generated.

[0033] The dynamic scheduling and execution module is used to parse the verified computing tasks, extract feature vectors to form feature fingerprints, evaluate the feature fingerprints using a preset AMCU adaptability scoring model, and generate a scheduling decision to prioritize scheduling to the AMCU if the adaptability score exceeds a preset threshold. When executing the scheduling decision, if the target AMCU node fails, the scheduling is performed with a lower priority.

[0034] Specifically, the dynamic scheduling and execution module receives and parses verified computational tasks, extracting feature vectors from the computational graph structure of the task, including the proportion of matrix operations, numerical precision requirements, computational intensity, and data locality dimensions, to form a unique feature fingerprint for the computational task. The module evaluates the feature fingerprint using a pre-defined AMCU suitability scoring model, which establishes scoring rules based on the performance advantages of AMCU in matrix operations, convolution computation, and quantized inference. When the suitability score exceeds a preset threshold, the module prioritizes scheduling the computational task to AMCU for execution, fully leveraging AMCU's high efficiency and low power consumption characteristics in specific computing scenarios. Before scheduling, the module determines whether the target AMCU node has hardware failures, network anomalies, or excessive load. If a failure is detected, a lower priority scheduling strategy is implemented, transferring the task to a less desirable GPU or CPU node for execution.

[0035] The fault tolerance mechanism module is used to determine whether there is a fault in the status of heterogeneous computing power resource nodes. If a fault is found, a rescheduling process is triggered.

[0036] Specifically, the fault tolerance mechanism module continuously monitors the operational status of heterogeneous computing power resource nodes, identifying node failures through heartbeat detection, performance metric anomaly analysis, and task execution timeout judgment. When a node failure is detected, the module immediately triggers a rescheduling process. The rescheduling process is as follows: the fault tolerance mechanism module reports detailed information about the node failure event to iRegNode for secondary verification, including the failure type, scope of impact, and estimated recovery time. After iRegNode completes the failure verification, it marks the relevant node as faulty or isolated and synchronizes the updated information to the global resource directory through the consortium consensus network, ensuring that other system components can obtain the latest resource status in a timely manner.

[0037] During the rescheduling process, the fault tolerance mechanism module queries the computing task allocation records to accurately identify all computing tasks running on the faulty node. These affected faulty tasks are removed from the current execution queue and added to the waiting queue. The scheduling process is re-executed, and a new scheduling decision is generated based on the updated resource status. The computing tasks are allocated to available standby nodes. Depending on the task's checkpoint mechanism and state preservation, the system chooses to resume execution from the most recent breakpoint or start again from the beginning, minimizing the impact of the fault on task completion. The entire rescheduling process generates audit logs and sends them to the consortium consensus network for evidence storage via iRegNode.

[0038] In a preferred embodiment, when the system receives a deep learning inference task, the platform access module performs compliance verification via iRegNode, and the dynamic scheduling and execution module extracts the task feature fingerprint, identifying that the task contains 85% INT8 precision matrix multiplication operations. The AMCU adaptability scoring model calculates a high adaptability score of 88, based on which the system generates an AMCU priority scheduling strategy. In actual execution, compared to traditional GPU solutions, this scheduling decision reduces energy consumption to 1 / 8 of the original and increases throughput by 5 times, while real-time verification via iRegNode ensures the security and compliance of data processing.

[0039] After the entire scheduling process is completed, the fault tolerance mechanism module generates detailed audit logs containing task execution trajectories, resource usage, performance metrics, and anomaly handling records. These logs are reliably marked using iRegNode's digital signature and timestamp services, and then sent to the consortium consensus network for distributed storage, forming an immutable execution record that provides reliable data support for subsequent system optimization, fault analysis, and compliance review.

[0040] Example 2 Please see Figure 2 This embodiment provides a method for dynamic scheduling of heterogeneous computing resources based on multi-objective optimization. The method utilizes a scheduling system deployed in a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance. The scheduling system enables intelligent allocation and management of heterogeneous computing resources, specifically including the following steps: S1. The platform access module calls the service node registration smart contract interface provided by the platform to register the scheduling system as a computing power scheduling service node.

[0041] Specifically, the platform access module calls the service node registration smart contract interface provided by the platform to register the scheduling system as a computing power scheduling service node. The system first submits a registration request to the platform, including the scheduling system's identity identifier, service capability description, and security authentication information. After verification by the smart contract, it obtains the service node identity recognized by the platform and establishes a trusted connection channel with the platform.

[0042] S2. Obtain the computing tasks audited by iRegNode in the platform, call the iRegNode interface to perform compliance verification on the computing tasks, and send the computing tasks that pass the verification to the dynamic scheduling and execution module.

[0043] Specifically, the system retrieves computational tasks audited by iRegNode on the platform, calls the iRegNode interface to perform compliance verification on the computational tasks, and sends the verified computational tasks to the dynamic scheduling and execution module. After receiving deep learning model training tasks submitted by enterprise users, the system first verifies the legality of the task's data source, the rationality of its computational resource requirements, and its security and privacy requirements through the iRegNode interface to ensure that the task complies with the platform's governance standards. Tasks that pass verification are marked as schedulable and forwarded to subsequent modules.

[0044] S3. The data acquisition and perception module collects the status data of heterogeneous computing power resource nodes in the global resource directory published by the platform.

[0045] Specifically, the data acquisition and sensing module collects status data of heterogeneous computing power resource nodes from the global resource catalog published by the platform. It monitors various heterogeneous computing power resources, including CPU, GPU, FPGA, and AMCU computing units, in real time, collecting key status indicators such as processor utilization, memory usage, network bandwidth, and energy consumption levels for each node, and constructs a dynamic global resource catalog.

[0046] S4. The dynamic scheduling and execution module parses the verified computational tasks, extracts feature vectors to form feature fingerprints, and evaluates the feature fingerprints using a preset AMCU adaptability scoring model. If the adaptability score exceeds a preset threshold, a scheduling decision is generated to prioritize scheduling to the AMCU. When executing the scheduling decision, if the target AMCU node fails, a lower priority scheduling is performed.

[0047] S5, the multi-objective modeling and optimization module, obtains a multi-objective optimization function based on a deep reinforcement learning model, and generates an initial scheduling decision by dynamically adjusting the objective weights.

[0048] Specifically, the dynamic scheduling and execution module parses the verified computational tasks, extracts feature vectors to form feature fingerprints, and evaluates these fingerprints using a preset AMCU suitability scoring model. If the suitability score exceeds a preset threshold, a scheduling decision is generated to prioritize scheduling to the AMCU. When executing the scheduling decision, if the target AMCU node fails, a lower priority scheduling is performed. The system extracts feature vectors from the computational graph structure of the computational task, including multiple dimensions such as the proportion of matrix operations, numerical precision requirements, computational intensity, and data locality, to form the feature fingerprint of the computational task.

[0049] Matrix operation percentage can be understood as the proportion of matrix multiplication, convolution, and other operations to the total number of operations, with a quantization value range of [0, 1]. Numerical precision requirement can be understood as the most frequently occurring numerical type in the task computation graph, with enumerated values ​​such as {INT8, FP16, FP32, FP64}. Computational intensity can be understood as the ratio of the total number of operations to the amount of data movement, assessing whether the task is computationally intensive or memory-intensive. Data locality can be understood as an index derived from the analysis of the continuity and reusability of data access patterns, with a range of [0, 1], where a higher value indicates better data locality.

[0050] For example, the model pre-defines a scoring function (or lookup table) and a weighting coefficient for each feature fingerprint. This can be dynamically adjusted through training on historical data or platform strategies.

[0051] The weight of matrix operation percentage (M_density) is set to 40%. If M_density ≥ 0.7, it is scored 40 points; if 0.5 ≤ M_density < 0.7, it is scored 25 points; if 0.3 ≤ M_density < 0.5, it is scored 10 points; otherwise, it is scored 0 points.

[0052] The weight of numerical precision requirements (INT8 / FP16) is set at 30%, with INT8 / FP16 receiving 30 points, FP32 receiving 15 points, and FP64 or others receiving 0 points.

[0053] The weight of the intensity (C_intensity) is set to 20%. If C_intensity > 10, it is scored 20 points; if 5 ≤ C_intensity ≤ 10, it is scored 12 points; if C_intensity < 5, it is scored 0 points.

[0054] The weight of data locality (L_locality) is set to 10%. If L_locality > 0.8, it scores 10 points; if 0.5 ≤ L_locality ≤ 0.8, it scores 8 points; and if L_locality < 0.5, it scores 0 points.

[0055] The scores of each dimension are multiplied by their corresponding weights and summed to obtain a comprehensive adaptability score (Score_AMCU) of 0-100. The calculated Score_AMCU is compared with a preset scheduling threshold, which is 75 in this embodiment. If Score_AMCU ≥ 75, an AMCU-priority scheduling decision is generated and the task enters the AMCU-priority scheduling path; if Score_AMCU < 75, the task enters the standard heterogeneous resource scheduling process.

[0056] S6. The platform access module calls the iRegNode interface to determine whether the initial scheduling decision is compliant. If not, it returns to step S5 for correction.

[0057] Specifically, the multi-objective modeling and optimization module derives a multi-objective optimization function based on a deep reinforcement learning model. It generates initial scheduling decisions by dynamically adjusting the objective weights. The multi-objective optimization function includes constraints for maximizing system throughput, maximizing resource utilization, maximizing energy efficiency, and minimizing average task latency. The reward function R of the deep reinforcement learning model is defined by the formula R = w1·Perf + w2·Util + w3·Eneg + w4·Comp, where Perf represents the normalized performance score, Util represents resource utilization, Eneg represents the reciprocal of the energy efficiency score, Comp represents the compliance score, and w1, w2, w3, and w4 represent weight coefficients.

[0058] S7, the dynamic scheduling and execution module, makes the final compliant scheduling decision and allocates computing tasks to matching heterogeneous computing resource nodes.

[0059] Specifically, the platform access module calls the iRegNode interface to determine whether the initial scheduling decision is compliant. If not, it returns to step S5 for correction. The system submits the initial scheduling decision to iRegNode for compliance verification, checking whether the scheduling scheme violates the platform's resource usage policy, data security requirements, and service level agreement. When it is detected that the load of a certain AMCU node in the scheduling decision exceeds the platform's 85% limit, iRegNode returns a non-compliant result, and the system re-executes step S5 to adjust the weight coefficients and re-optimize the scheduling scheme.

[0060] The dynamic scheduling and execution module makes the final compliant scheduling decision, allocating computing tasks to matching heterogeneous computing resource nodes. The system allocates the matrix operation part of the deep learning inference task to AMCU nodes for parallel execution, and the data preprocessing part to GPU processing, realizing collaborative computing of heterogeneous resources.

[0061] The system also includes a fault detection and handling mechanism to determine whether there are faults in the heterogeneous computing power resource nodes. If a fault is found, a rescheduling process is triggered. The rescheduling process is as follows: the node fault event is reported to iRegNode for inspection and confirmation; corresponding measures are taken according to the fault type; iRegNode marks the confirmed faulty node as faulty or isolated; the global resource catalog is updated through the consortium consensus network; the computing task allocation record is queried to identify the faulty computing task; the faulty computing task is added to the scheduling queue; the scheduling process is re-executed to generate a new scheduling decision; the computing task is allocated to a standby node and execution is resumed from the breakpoint or start point; an audit log is generated and sent to the consortium consensus network through iRegNode for evidence storage.

[0062] When the fault type is determined to be a hardware failure, the system immediately triggers a computation task rescheduling process, migrating inference tasks running on the failed AMCU node to a backup GPU node for continued execution. When the fault type is determined to be a performance anomaly, the system performs performance degradation processing, optimizing computational resource allocation and maintaining service availability by reducing the number of concurrent tasks and adjusting computational precision. When the fault type is determined to be a compliance violation, the system triggers a compliance circuit breaker mechanism. When the fault type is determined to be a network interruption, the system performs network switching and reconnection, activating a backup network channel to ensure the continuity of task execution.

[0063] The system utilizes the scoring results output by the fault risk scoring model to implement a compliant circuit breaker mechanism. When the risk result is determined to be high-risk, the system immediately suspends scheduling services, marks the faulty node as isolated, stops assigning new tasks to the node, and migrates existing tasks to other nodes. When the risk result is determined to be medium-risk, the system reduces the load and monitors recovery, reduces the task allocation weight of the node by 50%, and continuously monitors its performance indicators. When the risk result is determined to be low-risk, the system records the faulty node information and continuously monitors the status information of heterogeneous computing power resource nodes, preventing potential faults by increasing the monitoring frequency and setting early warning thresholds. The fault risk scoring model is implemented based on the Exponentially Weighted Moving Average (EWMA) algorithm, predicting the fault risk level by analyzing the changing trends of historical performance data of nodes.

[0064] In a preferred embodiment, the fault risk scoring model can also employ the isolated forest algorithm, which constructs a random forest to identify abnormal node behavior patterns, thereby improving the accuracy and real-time performance of fault prediction.

[0065] Example 3 To verify the effectiveness of the foregoing embodiments of the present invention, experiments and performance evaluations were conducted. Benchmark tests were performed in a simulated platform environment. A cluster test environment with 32 heterogeneous nodes was constructed, including 4 AMCU prototype machines, 8 GPU servers, and 20 CPU servers. The platform and the scheduling system of the present invention were deployed, with an enhanced Kubernetes scheduler as the baseline. Matrix-intensive tasks and synthetic scientific computing tasks from publicly available benchmark datasets such as MLPerf were used to create a 72-hour mixed load stream.

[0066] Experimental results demonstrate that this invention, through precise AMCU adaptive scheduling, increases AMCU node utilization from approximately 41.2% to 86.7% compared to traditional scheduling, driving the overall weighted utilization of the system to over 85%. For the same batch of matrix-intensive tasks, the total energy efficiency ratio (throughput / total energy consumption) was measured on both AMCU and GPU clusters. The results show that the energy efficiency on AMCU is 3 to 5 times that on GPU. Utilizing the platform's global view, cross-domain task scheduling decision latency is less than 100 milliseconds. Through integration with iRegNode and the consensus network, node fault detection and task migration time are reduced to the second level, achieving a system availability of 99.9%.

[0067] Example 4 In this embodiment, a quantum chemistry simulation task is audited by the platform iRegNode and then forwarded to the scheduling system. The scheduling system extracts a feature fingerprint showing that matrix operations account for 78% of the task, and the AMCU compatibility score is 82. The multi-objective modeling and optimization module, considering performance and cross-platform synergy, generates a hybrid scheduling strategy: the matrix operation portion is scheduled to the AMCU node, and the quantum computing portion is scheduled to the QPU node. During execution, the fault tolerance mechanism module detects an anomaly in the QPU decoherence time (increased risk score), proactively issues a warning, and triggers a task migration to classical simulation mode, ensuring task continuity. All scheduling and fault tolerance operations are audited and documented through the platform protocol.

[0068] In summary, the beneficial effects of all embodiments of the present invention are as follows: by accessing a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance and applying the scheduling method of the present invention, the overall resource utilization rate of the system is increased from approximately 60% to over 85% compared to traditional schedulers; for matrix-intensive tasks, by scheduling to AMCU units with higher energy efficiency, a 3-5 times energy efficiency improvement compared to pure GPU solutions is achieved; utilizing the platform's global resource view, the latency of cross-domain task scheduling decisions is reduced to less than 100 milliseconds; through the fault tolerance mechanism linked with iRegNode within the platform, the detection time of node failures and task migration is shortened to the second level, and the system availability reaches 99.9%; all scheduling decisions are verified in real time by iRegNode, preventing the execution of illegal tasks from the source.

Claims

1. A heterogeneous computing resource dynamic scheduling system based on multi-objective optimization, wherein the system is deployed in a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance, characterized in that... include: The platform access module is used to interact with the consortium consensus network and iRegNode data in the platform; The system calls the smart contract registration interface provided by the platform to register the scheduling system as a computing power scheduling service node; obtains computing tasks that have passed the iRegNode audit, calls the iRegNode interface to perform compliance verification on the computing tasks, and sends the verified computing tasks to the dynamic scheduling and execution module. The data acquisition and sensing module is used to collect status data of heterogeneous computing power resource nodes in the global resource directory published by the platform; The multi-objective modeling and optimization module is used to obtain multi-objective optimization functions based on deep reinforcement learning models and generate initial scheduling decisions by dynamically adjusting the objective weights. The dynamic scheduling and execution module is used to parse the verified computing tasks, extract feature vectors to form feature fingerprints, evaluate the feature fingerprints using a preset AMCU adaptability scoring model, and generate a scheduling decision to prioritize scheduling to the AMCU if the adaptability score exceeds a preset threshold. When executing the scheduling decision, if the target AMCU node fails, the scheduling is performed with a lower priority.

2. The scheduling system according to claim 1, characterized in that, It also includes a fault tolerance mechanism module, which is used to determine whether there is a fault in the status of heterogeneous computing power resource nodes. If a fault is found, a rescheduling process is triggered. The rescheduling process is as follows: The node failure event is reported to iRegNode for inspection and confirmation. iRegNode marks the confirmed faulty node as faulty or isolated and updates the global resource catalog through the consortium consensus network. Query the computing task allocation record, identify the faulty computing task, add the faulty computing task to the scheduling queue, re-execute the scheduling process, generate a new scheduling decision, allocate the computing task to the standby node, and resume execution from the breakpoint or the starting point. An audit log is generated and sent to the consortium consensus network via iRegNode for evidence storage.

3. A dynamic scheduling method for heterogeneous computing resources based on multi-objective optimization, wherein the method is implemented using a scheduling system deployed in a digital ecosystem platform consortium based on multi-party consensus and embedded intelligent governance, characterized in that... Includes the following steps: S1. The platform access module calls the smart contract registration interface provided by the platform to register the scheduling system as a computing power scheduling service node. S2. Obtain the computing tasks audited by iRegNode in the platform, call the iRegNode interface to perform compliance verification on the computing tasks, and send the computing tasks that pass the verification to the dynamic scheduling and execution module. S3. The data acquisition and sensing module collects the status data of heterogeneous computing power resource nodes in the global resource directory published by the platform. S4. The dynamic scheduling and execution module parses the verified computational tasks, extracts feature vectors to form feature fingerprints, and evaluates the feature fingerprints using a preset AMCU adaptability scoring model. If the adaptability score exceeds a preset threshold, a scheduling decision is generated to prioritize scheduling to the AMCU. When executing the scheduling decision, if the target AMCU node fails, a lower priority scheduling is performed. S5. The multi-objective modeling and optimization module obtains a multi-objective optimization function based on a deep reinforcement learning model, and generates an initial scheduling decision by dynamically adjusting the objective weights. S6. The platform access module calls the iRegNode interface to determine whether the initial scheduling decision is compliant. If not, it returns to step S5 for correction. S7, the dynamic scheduling and execution module, makes the final compliant scheduling decision and allocates computing tasks to matching heterogeneous computing resource nodes.

4. The scheduling method according to claim 3, characterized in that, The method further includes: determining whether the heterogeneous computing power resource node has a fault; if a fault is found, a rescheduling process is triggered, wherein the rescheduling process is as follows: The node failure event is reported to iRegNode for inspection and confirmation. According to the failure type, the corresponding measures are executed. iRegNode marks the confirmed faulty node as faulty or isolated and updates the global resource catalog through the consortium consensus network. Query the computing task allocation record, identify the faulty computing task, add the faulty computing task to the scheduling queue, re-execute the scheduling process, generate a new scheduling decision, allocate the computing task to the standby node, and resume execution from the breakpoint or the starting point. An audit log is generated and sent to the consortium consensus network via iRegNode for evidence storage.

5. The scheduling method according to claim 4, characterized in that, If the fault type is determined to be a hardware fault, the computing task rescheduling process is immediately triggered. If the fault type is determined to be a performance anomaly, then performance degradation processing is performed to optimize the allocation of computing resources; If the fault type is determined to be a compliance violation, the compliance circuit breaker mechanism will be triggered. If the fault type is determined to be a network interruption, then network switching and reconnection are performed.

6. The scheduling method according to claim 5, characterized in that, The compliant circuit breaker mechanism is executed using the scoring results output by the fault risk scoring model, wherein: If the risk assessment results indicate a high risk, the scheduling service will be immediately suspended and the faulty node will be marked as isolated. If the risk assessment indicates a medium risk, reduce the load and monitor recovery. If the risk assessment result is low risk, the fault node information is recorded and the status information of heterogeneous computing power resource nodes is continuously monitored.

7. The scheduling method according to claim 6, characterized in that, The fault risk scoring model can be based on exponentially weighted moving average (EWMA) or the isolated forest algorithm.

8. The scheduling method according to claim 3, characterized in that, In step S4, feature vectors including the proportion of matrix operation, numerical precision requirements, computational intensity, and data locality dimensions are extracted from the computation graph structure of the computation task to form the feature fingerprint of the computation task. The numerical precision requirement refers to the numerical type that appears most frequently in the computation graph, the computation intensity is the ratio of the total number of operands to the amount of data movement, and the data locality is an index derived from the analysis of the continuity and reusability of data access patterns.

9. The scheduling method according to claim 3, characterized in that, The multi-objective optimization function described in step S5 includes at least the objective constraints of maximizing system throughput, maximizing resource utilization, maximizing energy efficiency, and minimizing average task latency.

10. The scheduling method according to claim 3, characterized in that, The reward function R of the deep reinforcement learning model described in step S5 is defined by the following formula: R = w1·Perf + w2·Util + w3·Eneg + w4·Comp; Where Perf represents the normalized performance score, Util represents the resource utilization rate, Eneg represents the reciprocal of the energy efficiency score, and Comp represents the compliance score; w1, w2, w3, and w4 represent weighting coefficients.

Images