Heterogeneous computing cluster deployment method and collaborative scheduling system

Abstract

The invention relates to the technical field of computers, provides a heterogeneous computing cluster deployment method and a collaborative scheduling system, realizes full-process automation and intelligent management from resource evaluation, task scheduling to dynamic optimization by accurately sensing the performance and task characteristics of computing units and the running state of a heterogeneous computing cluster, and improves the scheduling efficiency compared with a traditional scheduling scheme. According to the scheduling system, the heterogeneous computing cluster resource utilization rate, the task execution efficiency and the system stability are remarkably improved, the energy consumption cost is effectively reduced, the scheduling system adapts to the cluster environment with diversified task loads and dynamic changes, and an efficient and reliable collaborative scheduling solution is provided for a large-scale heterogeneous computing scene.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a heterogeneous computing cluster deployment method and a collaborative scheduling system. Background Art

[0002] With the rapid development of fields such as artificial intelligence and big data processing, heterogeneous computing clusters have become key infrastructure for supporting large-scale and complex computing tasks by integrating the advantages of multiple computing units such as CPU, GPU and FPGA. Existing heterogeneous computing clusters have shown great potential in improving computing performance, but there are still many problems that need to be solved in terms of resource scheduling and management.

[0003] From the perspective of resource assessment, traditional heterogeneous computing clusters often use a single indicator or a simple weighted method to evaluate the performance of computing units. For example, computing power is measured only based on CPU main frequency or memory capacity. This evaluation method ignores the multidimensional characteristics of computing units and cannot dynamically adapt to changes in hardware performance over time and load. When faced with complex task scenarios, the scheduling strategy formulated based on such inaccurate assessments can easily lead to an imbalance in the allocation of computing unit resources. Some units may be overloaded and experience performance bottlenecks, while other units may be idle, resulting in resource waste.

[0004] In the task scheduling phase, existing scheduling strategies often focus on a single optimization objective, such as simply pursuing task execution speed or reducing energy costs. These strategies lack consideration of the deep adaptation relationship between task characteristics and resource status. Simultaneously, simple rule-driven scheduling approaches struggle to cope with diverse task loads and dynamically changing cluster environments. They fail to effectively balance multiple key objectives, such as task latency, resource balance, and energy costs, resulting in low overall cluster efficiency. Most strategies fail to address how to achieve task scheduling and intelligent deployment through resource and status awareness. Summary of the Invention

[0005] To address the deficiencies of the existing technology, this application provides a heterogeneous computing cluster deployment method and a collaborative scheduling system.

[0006] In a first aspect, the present application provides a heterogeneous computing cluster collaborative scheduling system, which includes: obtaining the instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption of each computing unit to construct a hardware capability vector for each computing unit, and predicting the performance of each computing unit through a bidirectional long short-term memory network, and connecting the performance of each computing unit through a graph convolutional network to determine the resource vector of the heterogeneous computing cluster;

[0007] By collecting statistics on task sets and extracting their task features, we can obtain the task vectors of heterogeneous computing clusters. We can then integrate the directional similarity and numerical matching between the task vectors and resource vectors to generate an adaptation matrix. Finally, we can generate a scheduling strategy for the heterogeneous computing cluster based on multi-objective optimization.

[0008] Simulate the spatiotemporal efficiency distribution of different scheduling strategies, identify risk calculation units with resource competition based on the spatiotemporal efficiency distribution, elastically shard the risk calculation units, generate a sharding instruction set, which includes sharding granularity, isolation boundaries, and communication routing tables, and output a deployment energy efficiency report. At the same time, monitor the operating status of different scheduling strategies in real time, including the hardware temperature curve and network packet loss rate. Make a comprehensive judgment based on the operating status and deployment energy efficiency report to update the scheduling strategy.

[0009] As an optional implementation, the logic for determining the resource vector of the heterogeneous computing cluster includes:

[0010] Configure a multi-head attention mechanism to build a bidirectional long short-term memory network, and train the bidirectional long short-term memory network using historical capability vectors to predict the performance of each computing unit;

[0011] Through the graph convolutional network, each computing unit is regarded as a node, and the physical connection relationship between computing units is regarded as the edge between nodes to build a cluster resource graph. The node attributes represent the performance of each computing unit.

[0012] The weights of the edges in the cluster resource graph are determined based on the physical distance between computing units and the correlation of hardware capability vectors. The cluster resource graph is iteratively updated through a graph convolutional network to determine the resource vectors of the heterogeneous computing cluster.

[0013] Monitor the actual performance of each computing unit, record the difference between the predicted performance and the actual performance of each computing unit through the error matrix, and calculate the average error of the error matrix. Based on the average error of the error matrix, trigger the parameter adjustment of the bidirectional long short-term memory network and update the resource vector of the heterogeneous computing cluster.

[0014] As an optional implementation manner, the sub-logic for constructing the hardware capability vector of each computing unit includes:

[0015] A heterogeneous computing cluster includes multiple computing units. The instruction throughput, floating-point peak value, memory bandwidth, and thermal design power consumption of each computing unit are obtained and standardized.

[0016] Assign weights to instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption based on task type, and construct a hardware capability vector for each computing unit by weighted summation.

[0017] An update cycle is set, and the fluctuation range of instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption is calculated within the update cycle. The distribution of weight values is updated according to the fluctuation range, and the hardware capability vector of each computing unit is reconstructed.

[0018] As an optional implementation, the generation logic of the scheduling policy includes:

[0019] Taking task delay, resource balance and energy consumption cost as optimization targets, normalizing the optimization targets and determining the target weights of the optimization targets according to the task type and system status to form an objective function;

[0020] Based on the genetic algorithm, each chromosome represents a scheduling scheme, each gene represents a task, and the gene value represents the computing unit assigned to each task. The fitness value of each chromosome is calculated according to the objective function, and the initial scheduling strategy is generated through selection, crossover and mutation iteration based on the fitness matrix.

[0021] By exchanging the computing units assigned to the two tasks, adjacent candidate scheduling strategies are generated, and the scheduling strategies are optimized using the simulated annealing algorithm;

[0022] During the execution of the scheduling strategy, the load status of the heterogeneous computing cluster is monitored in real time, and the scheduling strategy is dynamically selected based on the load status.

[0023] As an optional implementation manner, the generation sub-logic of the adaptation matrix includes:

[0024] Count the task sets and extract their task features, including computational intensity, data dependency, time sensitivity, and communication topology. These features are then standardized and converted into task vectors for heterogeneous computing clusters.

[0025] Assign dimension weights to task vectors and resource vectors based on task types, and calculate the directional similarity and numerical matching between task vectors and resource vectors based on the dimension weights.

[0026] The direction similarity and numerical matching degree are weighted and fused to obtain the adaptation score of each task-resource pair. The adaptation matrix is generated based on the adaptation scores of all task-resource pairs.

[0027] The adaptation matrix is dynamically optimized through reinforcement learning to update the adaptation matrix.

[0028] As an optional implementation, the generation logic of the sharding instruction set includes:

[0029] Based on the resource usage trend of each risk calculation unit, the resource pressure gradient of each risk calculation unit is calculated and risk calculation units with high resource pressure gradients are identified to prioritize elastic sharding and determine the sharding granularity. At the same time, data flow analysis is performed on the task sets running on the risk calculation units to identify the dependencies between tasks and dynamically adjust the sharding granularity based on the dependencies between tasks.

[0030] A resource dependency conflict graph is constructed using tasks as nodes and resource competition intensity as edges. The edge weights are determined based on the shared resource ratio and resource competition probability between tasks. The minimum cut set is then found in the resource dependency conflict graph to identify the isolation boundary of the shard and set the resource threshold for the isolation boundary.

[0031] Analyze the data transmission requirements between tasks in the shards, including transmission volume, frequency, and timeliness. Evaluate routing costs based on the cluster resource map and the load status of the heterogeneous computing cluster. Combine these requirements with routing costs to plan inter-shard communication paths. These paths must include at least one primary path and two backup paths, forming a communication routing table.

[0032] The sharding granularity, isolation boundary and communication routing table are structured and encoded, and the sharding instruction set is generated with the identifier of the risk calculation unit as the index. The feasibility of the sharding instruction set is verified to update the sharding instruction set.

[0033] As an optional implementation, the identification sub-logic of the risk calculation unit includes:

[0034] Simulate the deployment of scheduling policies, record the cluster resource graph during the deployment of each scheduling policy, and predict the resource usage trend of each computing unit. Combine the cluster resource graph and resource usage trend to form the spatiotemporal performance distribution of each scheduling policy.

[0035] Compare the local density of each node in the cluster resource graph to identify spatial anomaly nodes, and analyze the resource usage trend of each computing unit to identify temporal anomaly nodes. Mark computing units that meet both spatial and temporal anomaly node requirements as potential contention points.

[0036] Based on historical failure data, the resource contention probability of the cluster resource graph is learned, and multiple simulations are performed on each potential contention point to determine the resource contention intensity of each potential contention point.

[0037] The competition points are comprehensively judged based on the resource competition intensity and resource usage trend of each potential competition point to identify the risk calculation units where resource competition exists.

[0038] As an optional implementation, the updating logic of the scheduling policy includes:

[0039] Parse and extract the characteristics of task delay distribution and resource usage distribution in the deployment energy efficiency report to obtain the standard deviation of the average task delay and resource utilization rate. Simultaneously process and extract the characteristics of the hardware temperature curve and network packet loss rate in the running state to obtain the proportion of high-temperature nodes and the proportion of high-packet-loss routes.

[0040] Determine the indicator weights for average task delay, standard deviation of resource utilization, proportion of high-temperature nodes, and proportion of high-packet-loss routes based on task type. Determine the indicator score through weighted summation. Determine whether to update the scheduling strategy based on the indicator score.

[0041] When the scheduling policy needs to be updated, the adjustment direction of the scheduling policy is determined by combining the average task latency, the standard deviation of resource utilization, the proportion of high-temperature nodes, and the proportion of routes with high packet loss. The adjustment direction is prioritized according to the task type.

[0042] Adapt the local adjustment mechanism and the global optimization mechanism according to the adjustment direction to update the scheduling strategy, simulate the updated scheduling strategy, and simultaneously determine the indicator score of the updated scheduling strategy within the monitoring period to determine whether to update the scheduling strategy again.

[0043] As an optional implementation, the local adjustment mechanism includes reselecting computing units and regenerating shard instruction sets; the global optimization mechanism includes adjusting the crossover rate and mutation rate of the genetic algorithm and then regenerating the initial scheduling strategy.

[0044] In a second aspect, the present application provides a method for deploying a heterogeneous computing cluster, the method comprising: obtaining the instruction throughput, floating point peak, memory bandwidth, and thermal design power consumption of each computing unit to construct a hardware capability vector for each computing unit;

[0045] The efficiency of each computing unit is predicted through a bidirectional long short-term memory network, and the efficiency of each computing unit is connected through a graph convolutional network to determine the resource vector of the heterogeneous computing cluster;

[0046] Count the task sets and extract the task features of the task sets to obtain the task vectors of the heterogeneous computing cluster;

[0047] The directional similarity and numerical matching of task vectors and resource vectors are integrated to generate an adaptation matrix, and a scheduling strategy for heterogeneous computing clusters is generated based on multi-objective optimization.

[0048] Simulate the spatiotemporal efficiency distribution of different deployment scheduling strategies and identify computing units with resource competition risks based on the spatiotemporal efficiency distribution.

[0049] Perform elastic sharding on the risk computing unit, generate a sharding instruction set that includes sharding granularity, isolation boundaries, and communication routing tables, and output a deployment energy efficiency report;

[0050] Real-time monitoring of the operating status of different scheduling strategies, including hardware temperature curves and network packet loss rates, is used to update scheduling strategies based on comprehensive judgments based on the operating status and deployment energy efficiency reports.

[0051] Compared with the existing technology, the beneficial effects of this application are: by accurately sensing the performance, task characteristics and operating status of computing units and heterogeneous computing clusters, it realizes full-process automation and intelligent management from resource evaluation, task scheduling to dynamic optimization. Compared with traditional scheduling schemes, this scheduling system significantly improves the resource utilization, task execution efficiency and system stability of heterogeneous computing clusters, effectively reduces energy consumption costs, adapts to diversified task loads and dynamically changing cluster environments, and provides efficient and reliable collaborative scheduling solutions for large-scale heterogeneous computing scenarios.

[0052] By obtaining hardware performance parameters such as instruction throughput and floating-point peak, the hardware capability vector of each computing unit is constructed. Combined with bidirectional long short-term memory networks and graph convolutional network technologies, the performance of computing units can be accurately predicted and the resource vector of heterogeneous computing clusters can be determined. This process can comprehensively and dynamically evaluate the performance of computing units, taking into account the multi-dimensional characteristics of hardware capabilities and performance change trends. Compared with traditional single-indicator evaluation methods, it can more accurately reflect the actual processing capabilities of computing units in different task scenarios. The resource vector of the heterogeneous computing cluster generated in this way provides accurate resource status information for subsequent task scheduling, enabling scheduling strategies to be formulated based on real and reliable resource data, avoiding unreasonable scheduling problems caused by resource misjudgment, and ensuring the accuracy of task and resource matching from the source.

[0053] Starting from the task characteristics of the task set, the task vector of the heterogeneous computing cluster is generated. The directional similarity and numerical matching of the task vector and the resource vector are integrated to generate an adaptation matrix, and a scheduling strategy is generated based on multi-objective optimization, so as to deeply explore the adaptation relationship between task characteristics and resource status. By quantifying the degree of matching between tasks and resources in direction and value, the scheduling strategy generated based on the task-resource matching relationship fully considers multiple factors such as task delay, resource balance and energy consumption cost, and realizes comprehensive optimization of multiple objectives. Compared with traditional single-objective or simple rule-driven scheduling strategies, it can more reasonably allocate tasks under different task types and system states, effectively balance resource utilization and task execution efficiency, reduce resource waste, and improve the overall operation efficiency of heterogeneous computing clusters.

[0054] By simulating the spatiotemporal efficiency distribution of scheduling strategies, identifying risky computing units for resource competition, and performing elastic sharding and generating sharding instruction sets, and combining deployment of energy efficiency reports with operation status monitoring to update scheduling strategies, it is possible to proactively discover potential resource competition risks during cluster operation. Through elastic sharding technology, risk computing units are dynamically optimized to effectively isolate resource-competing tasks and ensure stable task execution. The operation status is monitored in real time and scheduling strategies are updated based on energy efficiency reports, enabling the scheduling system to quickly respond to environmental changes in heterogeneous computing clusters, adjust scheduling plans in a timely manner, and ensure that the scheduling system is always in an efficient operating state. This dynamic and adaptive elastic deployment mechanism enhances the scheduling system's adaptability to complex and changing computing environments, improves the fault tolerance and stability of heterogeneous computing clusters, and reduces the task failure rate caused by resource competition or environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:

[0056] Figure 1 A system flow chart of a heterogeneous computing cluster collaborative scheduling system provided in an embodiment of the present application;

[0057] Figure 2 A sub-logic diagram for generating the adaptation matrix of the heterogeneous computing cluster collaborative scheduling system provided in an embodiment of the present application;

[0058] Figure 3 A logic diagram for generating a sharding instruction set for a heterogeneous computing cluster collaborative scheduling system provided in an embodiment of the present application;

[0059] Figure 4 This is a flow chart of the method for deploying a heterogeneous computing cluster provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0061] Example 1

[0062] like Figure 1As shown, a system flow chart of a heterogeneous computing cluster collaborative scheduling system is provided for an embodiment of the present application. The system includes a resource construction module, a heterogeneous scheduling module and a flexible deployment module.

[0063] The resource construction module is used to obtain the instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption of each computing unit to construct the hardware capability vector of each computing unit, and predict the performance of each computing unit through a bidirectional long short-term memory network. The performance of each computing unit is connected through a graph convolutional network to determine the resource vector of the heterogeneous computing cluster.

[0064] Furthermore, the sub-logic for constructing the hardware capability vector of each computing unit includes:

[0065] A heterogeneous computing cluster includes multiple computing units. The instruction throughput, floating-point peak value, memory bandwidth, and thermal design power consumption of each computing unit are obtained and standardized.

[0066] Assign weights to instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption based on task type, and construct a hardware capability vector for each computing unit by weighted summation.

[0067] An update cycle is set, and the fluctuation range of instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption is calculated within the update cycle. The distribution of weight values is updated according to the fluctuation range, and the hardware capability vector of each computing unit is reconstructed.

[0068] Since the hardware configuration and performance of each computing unit in a heterogeneous computing cluster vary significantly, for example, the capabilities of devices such as CPU, GPU and FPGA in instruction processing and data calculation are different, in order to achieve accurate scheduling, it is necessary to fully obtain the core performance data of each computing unit and eliminate the interference of dimensional differences on subsequent analysis; through the dedicated monitoring interface of the scheduling system, the original data of instruction throughput, floating-point peak, memory bandwidth and thermal design power consumption of each computing unit are obtained in real time. After obtaining the original data, the scheduling system needs to perform standardized conversion on each data. Specifically, for numerical parameters such as instruction throughput, floating-point peak and memory bandwidth, These numerical parameters are mapped to the [0,1] interval using the maximum and minimum normalization formulas. The thermal design power consumption needs to be combined with the safe operating temperature threshold of the equipment and converted into a risk coefficient, which is also normalized to the [0,1] interval, where 0 indicates complete safety and 1 indicates reaching the risk threshold. By fully acquiring core performance data, the scheduling system can fully and accurately grasp the hardware characteristics of each computing unit, providing a reliable basis for subsequent resource evaluation and scheduling decisions. Standardization effectively eliminates the differences in dimensions and numerical ranges between different data types, making the performance data of different computing units comparable and avoiding evaluation biases caused by different data characteristics.

[0069] Different types of tasks have different requirements for the performance indicators of the computing unit. For example, deep learning training tasks have higher requirements for floating-point peak and memory bandwidth, while data processing tasks are more dependent on instruction throughput. Therefore, it is necessary to weightedly fuse the acquired data according to the task type to construct a hardware capability vector that is more in line with the actual task requirements; instruction throughput, floating-point peak, memory bandwidth and thermal design power consumption are used as basic elements, and the weights of each basic element are dynamically allocated according to the differences in hardware capability requirements of different task types. For example, for graphics rendering tasks, the weight values of floating-point peak and memory bandwidth will be increased accordingly, so that the fused vector can highlight the advantages and capabilities of the computing unit in processing graphics rendering tasks. For data processing tasks, the weight value of instruction throughput will be higher. Through weighted summation, the four standardized basic elements are fused and calculated, and finally a value that comprehensively reflects the hardware capability of the computing unit under the task type is obtained, and it is encoded into a hardware capability vector containing four data.

[0070] This feature fusion method based on task characteristics enables the hardware capability vector to more accurately reflect the adaptability of the computing unit to a specific task. Based on the hardware capability vector, the scheduling system can more reasonably allocate tasks to the most suitable computing unit, thereby improving task execution efficiency and resource utilization. The constructed hardware capability vector becomes the basis for subsequent monitoring of computing unit performance fluctuations and updating the vector. At the same time, the hardware capability vectors of multiple computing units also provide core data support for determining the resource vectors of heterogeneous computing clusters, which helps to evaluate the resource status and processing capabilities of the cluster as a whole.

[0071] During actual operation, the performance of a computing unit fluctuates due to various factors, including workload changes, hardware aging, and ambient temperature. To ensure that the hardware capability vector accurately reflects the actual performance of the computing unit, a dynamic update mechanism is required. The scheduling system sets a fixed update cycle. During each update cycle, it continuously monitors the computing unit's instruction throughput, floating-point peak value, memory bandwidth, and thermal design power consumption. The scheduling system calculates the fluctuation range of these metrics within the update cycle and compares them with preset fluctuation thresholds to determine whether the computing unit's performance has changed significantly. When the fluctuation range of a certain metric exceeds the fluctuation threshold, the scheduling system automatically triggers a weight adjustment procedure. At this time, the scheduling system reevaluates and modifies the weight previously assigned to the task type based on the specific fluctuation range of the metric. It then recalculates and updates the hardware capability vector based on the new weight. If the memory bandwidth of a computing unit fluctuates significantly over a period of time, the scheduling system analyzes the cause of the memory bandwidth fluctuation. If it is determined to be caused by frequent data exchange tasks, the scheduling system appropriately increases the weight of memory bandwidth for this task type and regenerates the hardware capability vector.

[0072] The dynamic update mechanism can track the performance changes of computing units in real time and adjust the hardware capability vector in a timely manner to ensure its accuracy and timeliness. This enables the scheduling system to always make decisions based on the actual performance of the computing units when assigning tasks, avoiding inefficient task execution or waste of resources due to performance misjudgment. The updated hardware capability vector provides more accurate data for predicting the performance of computing units, helping the bidirectional long short-term memory network to more accurately predict the future performance of computing units, thereby providing more reliable basic data for determining the resource vectors of heterogeneous computing clusters.

[0073] Specifically, the logic for determining the resource vector of a heterogeneous computing cluster includes:

[0074] Configure a multi-head attention mechanism to build a bidirectional long short-term memory network, and train the bidirectional long short-term memory network using historical capability vectors to predict the performance of each computing unit;

[0075] Through the graph convolutional network, each computing unit is regarded as a node, and the physical connection relationship between computing units is regarded as the edge between nodes to build a cluster resource graph. The node attributes represent the performance of each computing unit.

[0076] The weights of the edges in the cluster resource graph are determined based on the physical distance between computing units and the correlation of hardware capability vectors. The cluster resource graph is iteratively updated through a graph convolutional network to determine the resource vectors of the heterogeneous computing cluster.

[0077] Monitor the actual performance of each computing unit, record the difference between the predicted performance and the actual performance of each computing unit through the error matrix, and calculate the average error of the error matrix. Based on the average error of the error matrix, trigger the parameter adjustment of the bidirectional long short-term memory network and update the resource vector of the heterogeneous computing cluster.

[0078] In order to achieve efficient scheduling of heterogeneous computing cluster resources, it is necessary to understand in advance the performance of each computing unit in the future in order to reasonably allocate tasks. The performance of computing units is affected by many factors and has certain dynamic change characteristics. Therefore, it is necessary to use historical data (that is, historical capability vectors) for prediction; construct a bidirectional long short-term memory network, that is, a long short-term memory network based on bidirectional attention. First, the hardware capability vector (historical capability vector) of each computing unit in the past period of time is collected, and the bidirectional long short-term memory network is trained in combination with historical operation data including task execution status and resource utilization rate in the past period of time. When processing historical capability vectors and historical operation data, the bidirectional long short-term memory network not only analyzes the data from front to back in chronological order to capture the positive change trend of the data, but also analyzes the data again from back to front to explore the hidden reverse dependencies in the data.

[0079] At the same time, the bidirectional long short-term memory network introduces a multi-head attention mechanism, which enables the bidirectional long short-term memory network to allocate attention resources in a targeted manner according to the importance of different data, so as to more comprehensively and accurately capture the laws and characteristics of the hardware capabilities of the computing unit changing over time. By learning and training the historical capability vector, the bidirectional long short-term memory network can predict the future performance of each computing unit. For example, for a computing unit that is performing a complex computing task, the bidirectional long short-term memory network predicts whether it can continue to process tasks stably in the next period of time, or whether there will be a problem, based on the changes in its hardware capability vector and historical operation data in the past hour. Performance degradation; computing unit efficiency prediction based on historical data enables the scheduling system to grasp the performance trend of computing units in advance, provide forward-looking decision-making basis for task allocation, avoid task delays or failures caused by insufficient computing unit performance during task execution, improve the reliability of task execution and the overall operating efficiency of heterogeneous computing clusters, and accurate computing unit efficiency prediction results provide key node attribute information for the subsequent construction of cluster resource graphs. The predicted efficiency of each computing unit will be reflected in the cluster resource graph as a node attribute, which helps to build a graph structure that can truly reflect the status of cluster resources as a whole, laying the foundation for determining the cluster resource vector.

[0080] There are complex physical connections and logical collaboration relationships between the computing units in a heterogeneous computing cluster. The performance information of a single computing unit cannot fully reflect the overall resource status and collaborative processing capabilities of the heterogeneous computing cluster. Therefore, it is necessary to model the relationship between the computing units and determine the resource vector of the heterogeneous computing cluster through graph structure aggregation to achieve an overall evaluation of the resources of the heterogeneous computing cluster; the scheduling system abstracts the entire heterogeneous computing cluster into a graph structure consisting of nodes and edges, where each computing unit corresponds to a node in the cluster resource graph, and the node attribute is the previously predicted performance of the computing unit. The edges between nodes are constructed based on the physical connection relationship between the computing units, which includes the network topology and data transmission links.

[0081] In the process of building edges, the scheduling system will assign corresponding weights to each edge according to the physical distance between computing units and the correlation of hardware capability vectors. The weight of the edge reflects the closeness of the connection between the two computing units. The physical distance directly affects the data transmission delay and communication overhead, and is an important factor in building a cluster resource map. The physical distance is the physical location information of all computing units directly obtained through the cluster management system, which can accurately capture the communication overhead differences between computing units, provide a basis for building a more realistic cluster resource map, and help generate better scheduling strategies; at the same time, the correlation of hardware capability vectors between computing units reflects the performance differences of computing units in processing similar tasks. Combined with physical distance, it can more comprehensively evaluate the collaboration potential between computing units and solve the cosine similarity of the two hardware capability vectors between computing units to evaluate the method of the two hardware capability vectors. The directional consistency is calculated, and the Euclidean distance of the two hardware capability vectors between the computing units is solved at the same time to evaluate the numerical proximity of the two hardware capability vectors. The directional consistency and numerical proximity are weightedly fused to obtain the hardware capability vector correlation score. For example, for two GPU nodes, if their floating-point peak and memory bandwidth are highly similar, the correlation score is high, indicating that they have similar performance when processing graphics rendering or deep learning tasks; and then the physical distance and hardware capability vector correlation are weightedly fused to form the final weight of each edge, so that the cluster resource graph can more accurately reflect the collaboration potential between computing units and improve the representation ability of resource vectors. The weight of the fused edge is used as the input parameter of the graph convolutional network, which directly affects the iterative update process of the graph convolutional network on the cluster resource graph, and affects the quality of the resource vector of the heterogeneous computing cluster.

[0082] The scheduling system then iteratively updates this cluster resource graph multiple times using a graph convolutional network. During each update, each node receives information from neighboring nodes. Through information exchange and aggregation, the graph convolutional network learns the mutual influence and synergy between computing units. After three rounds of information propagation and aggregation, the scheduling system concatenates the relevant features of all nodes and removes redundant information through dimensionality reduction. Ultimately, it forms a cluster resource vector that represents the resource status and synergy of the entire heterogeneous computing cluster. Through graph structure aggregation, the scheduling system can comprehensively and intuitively display the relationships and overall resource status of computing units in the heterogeneous computing cluster. The resource vector of the heterogeneous computing cluster not only contains the performance information of each computing unit but also reflects the synergy between computing units. This provides a more accurate and comprehensive basis for the scheduling system to perform global resource scheduling, helping to improve the resource utilization and task processing capabilities of the heterogeneous computing cluster. The determined cluster resource vector becomes an important basis for subsequent matching with the task vector and generating scheduling strategies. The heterogeneous scheduling module can more rationally allocate tasks based on the cluster resource status reflected by the cluster resource vector and the characteristics of the task vector, achieving efficient matching of resources and tasks.

[0083] The predicted results of the performance of computing units will be affected by many factors and have certain errors. In order to ensure the accuracy of the resource vectors of heterogeneous computing clusters, it is necessary to establish a prediction result feedback mechanism to calibrate the resource vectors according to the difference between the actual performance and the predicted performance; the scheduling system continuously monitors the actual performance of each computing unit, and records the difference between the predicted performance and the actual performance of each computing unit through the error matrix. At the same time, the scheduling system calculates the average error of the error matrix to evaluate the accuracy of the overall prediction results. When the average error is greater than the preset error standard, it means that there is a large deviation in the previous prediction. At this time, the scheduling system will trigger two-way long-term and short-term Parameter adjustment of the memory network. During the parameter adjustment process, the scheduling system adjusts the parameters of the bidirectional long short-term memory network according to the error distribution and trend reflected by the error matrix to optimize the prediction ability of the bidirectional long short-term memory network. After the adjustment is completed, the scheduling system recalculates and updates the resource vector of the heterogeneous computing cluster based on the updated bidirectional long short-term memory network. If it is found that the predicted performance of multiple computing units is generally higher than the actual performance, the scheduling system will adjust the parameters of the bidirectional long short-term memory network so that it can more accurately reflect the actual performance of the computing units in subsequent predictions, and then update the resource vector of the heterogeneous computing cluster to make it more in line with the actual situation.

[0084] The prediction result feedback and vector calibration mechanism can promptly detect and correct errors in the performance prediction of computing units, ensuring the accuracy and reliability of the resource vectors of the heterogeneous computing cluster. This enables the scheduling system to make more reasonable and accurate decisions when scheduling tasks based on the cluster resource vectors, thereby improving the overall operating efficiency and resource utilization of the heterogeneous computing cluster. The calibrated resource vectors of the heterogeneous computing cluster provide the heterogeneous scheduling module with more accurate resource status information, which helps to improve the accuracy of task and resource matching and generate more optimized scheduling strategies. At the same time, the accurate resource vectors of the heterogeneous computing cluster also provide reliable data support for the elastic deployment module to identify risky computing units and update scheduling strategies, promoting the efficient operation of the entire heterogeneous computing cluster collaborative scheduling system.

[0085] The heterogeneous scheduling module is used to count the task sets and extract the task features of the task sets to obtain the task vectors of the heterogeneous computing clusters. It integrates the directional similarity and numerical matching of the task vectors and resource vectors to generate an adaptation matrix, and generates a scheduling strategy for the heterogeneous computing clusters based on multi-objective optimization.

[0086] Furthermore, if Figure 2 As shown, the generation logic of the adaptation matrix includes:

[0087] Count the task sets and extract their task features, including computational intensity, data dependency, time sensitivity, and communication topology. These features are then standardized and converted into task vectors for heterogeneous computing clusters.

[0088] Assign dimension weights to task vectors and resource vectors based on task types, and calculate the directional similarity and numerical matching between task vectors and resource vectors based on the dimension weights.

[0089] The direction similarity and numerical matching degree are weighted and fused to obtain the adaptation score of each task-resource pair. The adaptation matrix is generated based on the adaptation scores of all task-resource pairs.

[0090] The adaptation matrix is dynamically optimized through reinforcement learning to update the adaptation matrix.

[0091] The tasks in heterogeneous computing clusters have diverse characteristics. For example, computationally intensive tasks require high floating-point computing capabilities, while data-intensive tasks rely on high memory bandwidth. To achieve accurate scheduling, these unstructured task characteristics need to be converted into computable vector representations. The scheduling system uses a multi-dimensional feature extractor to analyze the task set. The computational intensity is quantified by calculating the proportion of floating-point operations, and the data dependency is quantified by analyzing the number of triggering times of data flows between tasks. Time sensitivity is determined by the compression ratio of the task deadline. The communication topology refers to the current network load obtained in real time. The extracted task features are analyzed by Z-score. Standardization processing eliminates dimensional differences and then maps them to a four-dimensional vector space to form a task vector. For example, a real-time video encoding task will be mapped to a vector representation with high computational intensity, medium data dependency, high time sensitivity, and a specific communication topology. This converts complex task characteristics into standardized task vectors, enabling the scheduling system to mathematically quantify task requirements and provide a unified metric for subsequent matching with resource vectors, avoiding scheduling deviations caused by subjective judgment. The standardized task vector provides clear input for dimensional weight allocation, ensuring the accuracy of subsequent directional similarity and numerical matching calculations, which directly affects the quality of the adaptation matrix.

[0092] Different types of tasks have different demands on resources. A single matching indicator cannot fully evaluate the adaptability of tasks and resources. For example, scientific computing tasks pay more attention to floating-point peak values, while Web service tasks are sensitive to instruction throughput. The dimension weights of task vectors and resource vectors are dynamically allocated according to different task types. For example, for computationally intensive tasks, the dimension weights of floating-point peak values and memory bandwidth will be significantly increased. For real-time tasks, the dimension weight of time sensitivity increases. When calculating directional similarity, the cosine similarity algorithm is used to evaluate the consistency of the direction between the task vector and the resource vector, reflecting the degree of matching between the two in the direction. The numerical matching degree is calculated by the Euclidean distance between the task vector and the resource vector. The degree of similarity between the numerical values of each dimension of the resource vector. For example, if an AI training task vector and a GPU resource vector are highly matched in terms of directional consistency and numerical matching in the floating-point peak and memory bandwidth dimensions, it indicates that the task is suitable for running on the GPU. Through dynamic weight allocation and multi-dimensional matching, the scheduling system can intelligently adjust the matching strategy according to task characteristics, improving the adaptation accuracy of tasks and resources, avoiding resource waste or performance bottlenecks caused by traditional single-indicator matching. The calculation results of directional similarity and numerical matching provide basic data for the generation of adaptation scores, directly determining the initial adaptation score of each task resource pair in the adaptation matrix, and affecting the generation of subsequent scheduling strategies.

[0093] The cluster operating environment changes dynamically, and a static adaptation matrix cannot respond in real time to emergencies such as hardware performance fluctuations and network congestion. Therefore, an adaptive mechanism is needed to continuously optimize the adaptation matrix. The scheduling system uses a deep Q-network to construct a reinforcement learning agent. The agent takes the current adaptation matrix, historical scheduling results, and resource usage status as state inputs. The agent performs actions to adjust task-resource matching relationships and obtains reward signals consisting of task completion time and resource utilization. To improve training efficiency, an experience replay mechanism is used to store historical state-action-reward triplets, and a target network stabilization training process is introduced. For example, when a computing unit is detected to have degraded performance, the agent adjusts the allocation of related tasks and gradually learns a more optimal matching strategy. The reinforcement learning mechanism enables the adaptation matrix to be automatically optimized based on the real-time status of the heterogeneous computing cluster, improving the scheduling system's adaptability to dynamic environments, reducing the need for manual intervention, and extending the effective lifecycle of scheduling strategies. The optimized adaptation matrix provides more accurate task-resource matching relationships for scheduling strategy generation, enabling optimization algorithms such as genetic algorithms to search within a more optimal initial solution space, thereby improving the quality and efficiency of scheduling strategy generation.

[0094] Specifically, the generation logic of the scheduling policy includes:

[0095] Taking task delay, resource balance and energy consumption cost as optimization targets, normalizing the optimization targets and determining the target weights of the optimization targets according to the task type and system status to form an objective function;

[0096] Based on the genetic algorithm, each chromosome represents a scheduling scheme, each gene represents a task, and the gene value represents the computing unit assigned to each task. The fitness value of each chromosome is calculated according to the objective function, and the initial scheduling strategy is generated through selection, crossover and mutation iteration based on the fitness matrix.

[0097] By exchanging the computing units assigned to the two tasks, adjacent candidate scheduling strategies are generated, and the scheduling strategies are optimized using the simulated annealing algorithm;

[0098] During the execution of the scheduling strategy, the load status of the heterogeneous computing cluster is monitored in real time, and the scheduling strategy is dynamically selected based on the load status.

[0099] A single optimization objective cannot balance the performance, efficiency, and cost of heterogeneous computing clusters. For example, pursuing low latency alone will lead to overallocation of resources, while focusing solely on energy consumption will sacrifice task completion time. The scheduling system uses task latency, resource balance, and energy cost as core optimization objectives. Through Min-Max normalization, each objective value is mapped to the interval [0, 1] to eliminate dimensionality effects. The objective weight of each optimization objective is adjusted according to the task type and system status. For real-time tasks, the objective weight of task latency is significantly increased, while for background batch tasks, the objective weight of energy cost is increased. For example, during periods of power shortage, the scheduling system automatically increases the objective weight of energy cost and prioritizes scheduling solutions with high energy efficiency. Multi-objective optimization can flexibly adjust the focus of optimization objectives according to actual needs, achieving comprehensive resource utilization in heterogeneous computing clusters and avoiding system performance imbalance caused by excessive pursuit of a single objective. The normalized objective function provides a clear fitness evaluation standard for the genetic algorithm, guiding it to find the optimal solution in the search space and directly affecting the quality of the initial scheduling strategy.

[0100] Genetic algorithms have advantages in global search, but are prone to falling into local optimality. Simulated annealing algorithms can accept inferior solutions with a certain probability, which helps to escape local optimality. The combination of the two can give full play to their respective advantages; the scheduling system uses chromosome encoding to represent the scheduling plan. Each gene corresponds to a task, and the gene value is the computing unit assigned to the task. The fitness value of each chromosome is calculated according to the objective function. The higher the fitness value, the better the plan. The initial population is generated through the task resource matching relationship determined in the adaptation matrix. In genetic operations, the crossover operator uses two-point crossover, and the mutation operator is implemented by randomly redistributing tasks. The initial scheduling strategy is generated through continuous iteration of selection, crossover and mutation, while retaining a certain proportion of high-quality plans to ensure that the overall quality of the population does not decline.

[0101] During the simulated annealing process, the current optimal solution is used as the starting point, and adjacent solutions, that is, adjacent candidate scheduling strategies, are generated by exchanging the computing units assigned to the two tasks. A probabilistic acceptance mechanism is adopted, which not only accepts better candidate scheduling strategies, but also accepts slightly worse candidate scheduling strategies with a certain probability, so as to avoid the algorithm from falling into the local optimal solution. As the optimization process proceeds, the probability of accepting the degraded strategy is gradually reduced. At the same time, the temperature parameter decays exponentially with the number of iterations, and the control algorithm gradually shifts from global search to local fine search. The collaborative optimization mechanism combines the global search capability of the genetic algorithm and the local escape capability of the simulated annealing algorithm, which improves the probability of the genetic algorithm finding the global optimal solution and shortens the convergence time. The generated scheduling strategy performs better in terms of multi-objective balance. The optimized scheduling strategy provides multiple candidate solutions for the dynamic selection mechanism, so that the scheduling system can quickly switch to the most appropriate scheduling strategy according to the real-time load status, thereby improving the response speed and stability of the scheduling system.

[0102] Cluster loads fluctuate dynamically, and a single scheduling strategy cannot adapt to all scenarios. For example, resource balance needs to be prioritized during peak hours, while energy optimization can be prioritized during off-peak hours. The scheduling system monitors the load status of each computing unit in real time, including metrics such as CPU utilization, memory usage, and network bandwidth. It constructs a load feature vector and uses K-means clustering to classify the load status into different modes, such as light load, medium load, and heavy load. It pre-calculates and stores the optimal scheduling strategy for each load mode, forming a scheduling strategy library. When load status changes are detected, the corresponding scheduling strategy is quickly selected from the scheduling strategy library through pattern matching. For example, if the load of a computing unit is continuously greater than 80%, the scheduling system automatically switches to a scheduling strategy with higher resource balance to avoid performance degradation caused by single-point overload. Dynamic load awareness enables the scheduling system to adapt to the state changes of heterogeneous computing clusters in real time, improving the relevance and effectiveness of scheduling strategies and reducing task delays and resource waste caused by load fluctuations. The dynamically selected scheduling strategy provides a more suitable initial solution for the elastic deployment module, reduces the risk of resource competition, improves the accuracy of shard instruction set generation, and promotes the coordinated optimization of the entire scheduling system.

[0103] The elastic deployment module is used to simulate the spatiotemporal efficiency distribution of deploying different scheduling strategies, identify risk calculation units with resource competition based on the spatiotemporal efficiency distribution, elastically shard the risk calculation units, generate a sharding instruction set, which includes sharding granularity, isolation boundaries, and communication routing tables, and output a deployment energy efficiency report. At the same time, it monitors the operating status of different scheduling strategies in real time, including the hardware temperature curve and network packet loss rate. The scheduling strategy is updated based on a comprehensive judgment based on the operating status and deployment energy efficiency report.

[0104] Furthermore, the identification sub-logic of the risk calculation unit includes:

[0105] Simulate the deployment of scheduling policies, record the cluster resource graph during the deployment of each scheduling policy, and predict the resource usage trend of each computing unit. Combine the cluster resource graph and resource usage trend to form the spatiotemporal performance distribution of each scheduling policy.

[0106] Compare the local density of each node in the cluster resource graph to identify spatial anomaly nodes, and analyze the resource usage trend of each computing unit to identify temporal anomaly nodes. Mark computing units that meet both spatial and temporal anomaly node requirements as potential contention points.

[0107] Based on historical failure data, the resource contention probability of the cluster resource graph is learned, and multiple simulations are performed on each potential contention point to determine the resource contention intensity of each potential contention point.

[0108] The competition points are comprehensively judged based on the resource competition intensity and resource usage trend of each potential competition point to identify the risk calculation units where resource competition exists.

[0109] Traditional resource monitoring only focuses on instantaneous states and cannot detect potential resource competition risks in advance. By determining the spatiotemporal efficiency distribution, resource usage trends can be predicted and forward-looking scheduling can be achieved. The scheduling system uses a discrete event simulation engine to simulate and deploy each scheduling strategy. During the simulation process, the topological changes of the cluster resource graph within each time window, including information such as node connection relationships and resource usage, are recorded. At the same time, the historical capability vector is used to predict the resource usage trend of each computing unit. The cluster resource graph and resource usage trend are combined to form a spatiotemporal efficiency distribution matrix. The spatiotemporal efficiency distribution can intuitively display the dynamic changes in resource usage, helping the scheduling system to identify resource hotspots and competition areas in advance, providing a reliable basis for risk warning. The generated spatiotemporal efficiency distribution provides a data basis for the identification of spatial anomaly nodes and temporal anomaly nodes, which directly affects the positioning accuracy of potential competition points.

[0110] Single-dimensional anomaly detection is prone to miss risk points in complex scenarios. For example, although the resource utilization rate of a computing unit is not greater than the utilization rate threshold, the difference in resource utilization with its adjacent nodes is too large, which will still cause competition. The scheduling system adopts a dual spatial and temporal anomaly detection mechanism to calculate the local density of each node in the cluster resource graph in the spatial dimension, and identify spatial abnormal nodes by comparing the resource utilization rate with that of adjacent nodes. The local density of each node refers to the absolute value of the difference in resource utilization between the node and the adjacent nodes at the same time point. The absolute value of the difference in resource utilization between the node and all adjacent nodes is summed up, and the resulting sum is the local density of the node. If the local density of a node is significantly greater than or less than the local density of its neighboring nodes, the node is considered to be abnormal. If there is a spatial anomaly at a point, it is a spatial anomaly node. In the time dimension, a time series analysis is performed on the resource usage trend of each computing unit, and the resource usage mutation points are identified through a sliding window algorithm and marked as time anomaly nodes. Only nodes that meet both spatial and temporal anomaly conditions will be marked as potential competition points. For example, if the resource utilization rate of a GPU node increases sharply in a short period of time and its local density is significantly different from that of the adjacent CPU nodes, it will be identified as a potential competition point. The multi-dimensional anomaly detection mechanism improves the accuracy of risk identification and reduces the false alarm rate through cross-validation. It can more accurately locate the resource competition risk points that really need attention. The accurately marked potential competition points provide clear target objects for subsequent resource competition intensity assessment, which helps to concentrate computing resources for in-depth analysis.

[0111] Based on historical fault data, the resource competition pattern of the cluster resource graph is learned, and multiple simulations are performed on each potential competition point to determine the resource competition intensity of each potential competition point;

[0112] The competition points are comprehensively judged based on the resource competition intensity and resource usage trend of each potential competition point to identify the risk calculation units where resource competition exists.

[0113] Identifying potential contention points alone is not sufficient to support decision-making; the likelihood and impact of contention must also be assessed in order to determine priorities. The resource contention pattern of the cluster resource graph is learned based on historical fault data, and Monte Carlo simulation is performed on each potential contention point. Multiple sets of simulation scenarios are generated by randomly perturbing input parameters, and the frequency and intensity of resource contention in each set of scenarios are counted. For example, for a storage node marked as a potential contention point, the scheduling system will simulate I / O conflicts under different data access modes, assess the probability of resource contention and the impact on system performance, and comprehensively consider resource usage trends and resource contention intensity to perform risk classification on each potential contention point and determine the final risk calculation unit. The quantitative assessment mechanism provides a clear priority ranking for the risk calculation unit, enabling the scheduling system to adopt different response strategies based on the risk level and optimize resource utilization efficiency. The classification results of the risk calculation unit directly determine the triggering order and optimization intensity of elastic sharding, providing key input for the generation of the sharding instruction set.

[0114] Specifically, if Figure 3 As shown, the generation logic of the sharding instruction set includes:

[0115] Based on the resource usage trend of each risk calculation unit, the resource pressure gradient of each risk calculation unit is calculated and risk calculation units with high resource pressure gradients are identified to prioritize elastic sharding and determine the sharding granularity. At the same time, data flow analysis is performed on the task sets running on the risk calculation units to identify the dependencies between tasks and dynamically adjust the sharding granularity based on the dependencies between tasks.

[0116] A resource dependency conflict graph is constructed using tasks as nodes and resource competition intensity as edges. The edge weights are determined based on the shared resource ratio and resource competition probability between tasks. The minimum cut set is then found in the resource dependency conflict graph to identify the isolation boundary of the shard and set the resource threshold for the isolation boundary.

[0117] Analyze the data transmission requirements between tasks in the shards, including transmission volume, frequency, and timeliness. Evaluate routing costs based on the cluster resource map and the load status of the heterogeneous computing cluster. Combine these requirements with routing costs to plan inter-shard communication paths. These paths must include at least one primary path and two backup paths, forming a communication routing table.

[0118] The sharding granularity, isolation boundary and communication routing table are structured and encoded, and the sharding instruction set is generated with the identifier of the risk calculation unit as the index. The feasibility of the sharding instruction set is verified to update the sharding instruction set.

[0119] Risk calculation units have different resource pressures and different task dependencies. Fixed sharding granularity cannot meet the needs of efficient resource allocation and task collaboration, and needs to be dynamically adjusted to balance resource utilization and task execution efficiency. The scheduling system monitors the resource usage trends of risk calculation units in real time, such as the curves of CPU usage and memory usage over time. By calculating the ratio of the difference in resource usage between adjacent time points to the time interval, the resource pressure gradient is obtained. For example, if the CPU usage of a risk calculation unit rises from 30% to 70% within 10 minutes, and the time interval is 10 minutes, then its resource pressure gradient is relatively high. Risk calculation units with a resource pressure gradient greater than the set gradient threshold (for example, the resource usage change per minute is greater than 5%) are marked as priority sharding objects.

[0120] At the same time, data flow analysis is performed on the task set running on the risk computing unit. By tracing the input and output relationship of the data between tasks, a task dependency graph is constructed. If the output data of task A is the input data of task B, and the two are closely related, then it is determined that there is a strong dependency between task A and task B; if the tasks only share a small amount of resources, then it is a weak dependency. For task groups with strong dependencies, a coarser sharding granularity is used to divide them into the same shard to reduce data transmission overhead. For tasks with weak dependencies or no dependencies, a finer sharding granularity is used to achieve resource isolation; this method can quickly locate computing units that urgently need sharding, dynamically adjust the granularity according to task dependencies, avoid task execution delays or resource waste due to improper sharding, and improve the overall performance of heterogeneous computing clusters. The determined sharding granularity provides the basis for the subsequent construction of resource dependency conflict graphs and identification of isolation boundaries, which affects the overall architecture of the sharding solution.

[0121] To avoid resource competition between tasks in a shard, it is necessary to clearly define isolation boundaries and set resource thresholds to ensure stable operation of each shard and prevent excessive resource use. Taking tasks as nodes and the previously determined resource competition intensity as edges, a resource dependency conflict graph is constructed. By analyzing task configuration files and runtime resource application records, the types (such as CPU cores, memory areas, and storage bandwidth, etc.) and quantities of shared resources between tasks are counted, and the proportion of shared resources is calculated. Combined with the frequency of competition among tasks in the same resource usage scenario in historical data, the probability of resource competition is determined. The shared resource ratio and the probability of resource competition are weighted and integrated to obtain the weight of the edge. The higher the weight, the greater the possibility of resource competition between tasks.

[0122] Using the minimum cut set algorithm in graph theory, we search for a partitioning scheme in the resource dependency conflict graph that can reasonably divide tasks into different shards. These dividing lines are the isolation boundaries of the shards. At the same time, based on the total resources of the risk calculation unit and the task requirements, we set a resource threshold for each isolation boundary. For example, if the total memory of a risk calculation unit is 128GB, based on the expected memory usage of the task, we set the memory resource threshold for a certain isolation boundary to 32GB. When the memory usage of the shard on one side of the boundary approaches the threshold, a resource warning is triggered. The precise setting of isolation boundaries and resource thresholds effectively isolates resource-competing tasks, ensures the independent and stable operation of each shard, and improves system reliability and resource utilization. The determined isolation boundaries and resource thresholds provide constraints for communication path planning and affect the generation of communication routing tables.

[0123] There is a need for data interaction between tasks in the shards. The network environment and cluster load change dynamically, so efficient and reliable communication paths need to be planned to ensure timely and low-cost data transmission. The scheduling system analyzes the properties and operating status of the tasks in the shards to determine the data transmission requirements. For big data processing tasks, the transmission volume is large. For real-time monitoring tasks, high requirements are placed on transmission frequency and timeliness. Based on the connection relationship and resource status of the computing units in the cluster resource graph, as well as the current load status of the heterogeneous computing cluster (such as the traffic occupancy of each link), the routing costs of different communication paths are evaluated, including factors such as transmission delay, bandwidth consumption, and packet loss rate.

[0124] A path planning algorithm, such as the Dijkstra algorithm combined with a load balancing strategy, is used to plan at least one main path and two backup paths for communication between each shard. The main path prioritizes links with low transmission delay and sufficient bandwidth, and the backup path is enabled when the main path fails or is congested. For example, the main path selects a directly connected link with a bandwidth of 10Gbps, backup path 1 selects a link that passes through an intermediate node but has a bandwidth of 8Gbps, and backup path 2 selects another link with a bandwidth of 6Gbps. A communication routing table containing path information, link parameters, and switching conditions is generated; the multi-path communication routing table ensures the reliability and efficiency of data transmission, reduces the risk of task failure due to network failure or uneven load, and improves the stability of the scheduling system. The communication routing table, shard granularity, and isolation boundary together constitute the core content of the shard instruction set, providing data support for subsequent encoding and verification.

[0125] The sharding scheme is converted into a computer-readable instruction set to facilitate execution by the scheduling system. Verification ensures the feasibility of the instruction set and avoids problems after deployment. The determined sharding granularity, isolation boundary, and communication routing table are structured and encoded in JSON format. The unique identifier of the risk calculation unit is used as the index to construct a sharding instruction set containing the parameters of each shard. The sharding instruction set is simulated and executed in a sandbox environment. During the simulation, task loads and resource status similar to those in actual conditions are injected, and indicators such as task execution time, resource utilization, and data transmission success rate are monitored. If the task execution time exceeds the expected time threshold, or resource utilization fluctuates abnormally, or data transmission fails frequently, it is determined that there is a problem with the instruction set. The previous steps are returned to adjust the sharding scheme, regenerate the instruction set, and verify it until it passes verification. The structured encoding makes the sharding instruction set easy to store, transmit, and execute. The strict verification process ensures the feasibility and effectiveness of the instruction set, reducing deployment risks. The verified sharding instruction set provides a specific execution plan for flexible deployment and guides the system to optimize and adjust the risk calculation unit.

[0126] Specifically, the update logic of the scheduling policy includes:

[0127] Parse and extract the characteristics of task delay distribution and resource usage distribution in the deployment energy efficiency report to obtain the standard deviation of the average task delay and resource utilization rate. Simultaneously process and extract the characteristics of the hardware temperature curve and network packet loss rate in the running state to obtain the proportion of high-temperature nodes and the proportion of high-packet-loss routes.

[0128] Determine the indicator weights for average task delay, standard deviation of resource utilization, proportion of high-temperature nodes, and proportion of high-packet-loss routes based on task type. Determine the indicator score through weighted summation. Determine whether to update the scheduling strategy based on the indicator score.

[0129] When the scheduling policy needs to be updated, the adjustment direction of the scheduling policy is determined by combining the average task latency, the standard deviation of resource utilization, the proportion of high-temperature nodes, and the proportion of routes with high packet loss. The adjustment direction is prioritized according to the task type.

[0130] Adapt the local adjustment mechanism and the global optimization mechanism according to the adjustment direction to update the scheduling strategy, simulate the updated scheduling strategy, and simultaneously determine the indicator score of the updated scheduling strategy within the monitoring period to determine whether to update the scheduling strategy again.

[0131] The local adjustment mechanism includes the reselection of computing units and the regeneration of shard instruction sets; the global optimization mechanism includes adjusting the crossover rate and mutation rate of the genetic algorithm and then regenerating the initial scheduling strategy.

[0132] Deployment energy efficiency reports and operating status data contain rich information. Extracting key indicator features can comprehensively evaluate the execution effect of scheduling strategies and provide a basis for updating decisions. From the deployment energy efficiency reports, the scheduling system analyzes the task delay distribution and calculates the average task delay by taking the average value of the difference between the task start time and completion time. The system also analyzes the resource usage distribution and calculates the standard deviation of resource usage to reflect the degree of balance in resource allocation. For example, the CPU usage of 10 computing units is calculated, and its standard deviation is calculated using a formula. The larger the standard deviation, the more unbalanced the resource allocation.

[0133] From the operating status data, the hardware temperature curve is processed, the number of high-temperature nodes (for example, temperatures exceeding 80°C) is counted, and the proportion of high-temperature nodes is calculated; the network packet loss rate is extracted, the number of high-packet loss routes (packet loss rate exceeding 5%) is determined, and the proportion of high-packet loss routes is calculated; the extracted key indicator features can intuitively reflect the performance of the scheduling strategy in task execution efficiency, resource allocation, hardware stability and network transmission, and help discover potential problems. These indicator features are the basic data for calculating indicator scores and determining whether to update the scheduling strategy.

[0134] Different task types place varying emphasis on various indicators. By calculating indicator scores and comparing them with indicator thresholds, the current scheduling strategy can be judged to meet requirements and decide whether to trigger an update. Based on task types (such as real-time tasks, batch tasks, and compute-intensive tasks), the scheduling system determines indicator weights for average task latency, standard deviation of resource utilization, proportion of high-temperature nodes, and proportion of routes with high packet loss. For example, for real-time tasks, the average task latency is weighted at 0.5, the standard deviation of resource utilization is weighted at 0.2, the proportion of high-temperature nodes is weighted at 0.1, and the proportion of routes with high packet loss is weighted at 0.2. The extracted indicator features are normalized to a value between 0 and 1, and then weighted summed according to the indicator weights to obtain an indicator score. A threshold is set, such as a score below 0.6. When the calculated indicator score is less than the threshold, a scheduling strategy update is determined; otherwise, the current strategy is maintained. This approach, by combining task type requirements with scheduling strategy evaluation, makes scheduling strategy updates more targeted and reasonable. After determining whether to update the scheduling strategy, if an update is necessary, the system proceeds to adjust the direction and optimize the strategy.

[0135] Different indicator anomalies correspond to different scheduling strategy issues. Clarifying the adjustment direction and sorting it can make the scheduling strategy update more efficient and give priority to solving key problems; establish a mapping relationship between indicator anomalies and adjustment directions. If the average task delay is too high, it corresponds to adjustment directions such as unreasonable computing unit allocation and improper task scheduling order; if the standard deviation of resource utilization is too large, it points to the direction in which the load balancing strategy needs to be optimized; if the proportion of high-temperature nodes is high, the adjustment direction is related to hardware overheating caused by the heat dissipation solution or resource allocation; if the proportion of high packet loss routes is high, it involves the adjustment direction of network configuration or task data transmission path; prioritize the adjustment directions according to the task type. For real-time tasks, the average task delay is too high and it is the primary problem. Prioritize the adjustment of computing unit allocation and task scheduling order; for batch tasks, the standard deviation of resource utilization is too large, which will lead to resource waste. Prioritize the load balancing strategy; accurately locate and sort the adjustment directions so that the scheduling strategy update focuses on key issues, improves update efficiency, and quickly improves system performance. The determined adjustment direction and priority guide the selection of appropriate local adjustment mechanism or global optimization mechanism for scheduling strategy update.

[0136] Depending on the adjustment direction, the scheduling strategy is updated using local adjustment or global optimization mechanisms, and simulation verification is used to ensure that the new strategy is effective to avoid blind deployment. If the adjustment direction points to a local problem, a local adjustment mechanism is enabled, such as computing unit reselection. The scheduling system evaluates the execution efficiency and resource usage of the current task in each computing unit, and migrates some tasks to computing units with lower load and greater adaptability. The sharding instruction set is regenerated, and the sharding instruction set generation logic steps are re-executed based on the new task allocation and resource status to adjust the sharding scheme.

[0137] If the adjustment direction involves global optimization, adjust the crossover rate and mutation rate of the genetic algorithm (for example, adjust the crossover rate from 0.6 to 0.7, and the mutation rate from 0.01 to 0.02), regenerate the initial scheduling strategy, and iteratively generate a new scheduling strategy through the selection, crossover, and mutation operations of the genetic algorithm; run the updated scheduling strategy in a simulation environment, simulate the actual task load and the status of the heterogeneous computing cluster, monitor indicators such as the average task delay, the standard deviation of resource utilization, the proportion of high-temperature nodes, and the proportion of high-packet loss routes, calculate the indicator score of the updated scheduling strategy, and if the indicator score improves and achieves the expected effect, apply the new scheduling strategy to the actual cluster; if it does not meet the standard, return to the adjustment mechanism and re-optimize the strategy until the requirements are met; update the strategy by mechanism and strictly simulate and verify it to ensure that the new scheduling strategy can effectively solve the problem, improve system performance, and reduce the risks brought by the scheduling strategy update. Apply the updated scheduling strategy to the heterogeneous computing cluster to achieve dynamic optimization of the scheduling system and continuously improve the collaborative scheduling efficiency of the heterogeneous computing cluster.

[0138] The local adjustment mechanism includes the reselection of computing units and the regeneration of sharded instruction sets. In some cases, only adjusting the computing unit allocation of some tasks can significantly improve the scheduling effect, without the need for global re-optimization. Taking into account factors such as the current load status, hardware capability vector, and task characteristics, the scheduling system evaluates the execution adaptability of each task on different computing units. When a computing unit is detected as a bottleneck, computing units with high adaptability and low load are prioritized for task migration. For example, when a CPU node's resource utilization is found to be too high, the scheduling system will evaluate the adaptability of tasks running on that node on other CPU or GPU nodes and migrate tasks suitable for running on the GPU to the GPU node. Pre-copy and gradual switching technologies are used during the migration process to ensure task continuity and reduce the impact of interruptions caused by migration. The intelligent reselection mechanism of computing units can quickly respond to local resource anomalies and improve the overall scheduling effect by fine-tuning task allocation. It is efficient and lightweight. After task migration, the scheduling system will reevaluate the rationality of the sharded instruction set and, if necessary, trigger the regeneration of the sharded instruction set to adapt to the new task allocation pattern.

[0139] When computing units are reselected or cluster status changes, the original shard instruction set becomes inapplicable and requires dynamic adjustment to maintain optimal performance. The scheduling system monitors the shard status in real time. When it detects that a shard's resource utilization rate consistently exceeds the utilization threshold or communication latency increases significantly, it triggers a shard instruction set regeneration process. It reanalyzes the dependencies and resource contention patterns between tasks in the shard and adjusts the shard granularity, isolation boundaries, and communication routing tables based on the latest cluster resource map and load status. If inter-task communication in a shard suddenly increases, the scheduling system evaluates whether isolation boundaries need to be adjusted to reduce communication overhead or communication paths need to be optimized to improve transmission efficiency. After feasibility verification, the regenerated shard instruction set is gradually deployed in production environments. The dynamic shard instruction set regeneration mechanism enables the system to quickly adapt to cluster status changes, maintain the optimality of sharding strategies, and improve the flexibility and robustness of the scheduling system. The updated shard instruction set provides a more refined execution plan for the scheduling strategy, helping to further optimize resource utilization and task execution efficiency, while also providing a foundation for the next round of strategy evaluation and adjustment.

[0140] The global optimization mechanism includes adjusting the crossover rate and mutation rate of the genetic algorithm and then regenerating the initial scheduling strategy. Fixed genetic algorithm parameters (such as crossover rate and mutation rate) cannot maintain high efficiency at both the early and late stages of the search. Adaptive adjustment of parameters can balance global search and local optimization capabilities. The genetic algorithm parameters are dynamically adjusted according to the current number of iterations and search progress. In the early stage of the search, a higher crossover rate and mutation rate are used to promote population diversity and expand the search range. As the iteration proceeds, the crossover rate and mutation rate are gradually reduced to improve the algorithm's local search capability and accelerate convergence. For example, in the first 10 generations of the algorithm, the crossover rate is set to The crossover rate is set to 0.8 and the mutation rate is set to 0.1; as the number of iterations increases, the crossover rate decreases linearly to 0.6 and the mutation rate decreases to 0.05. At the same time, a population diversity evaluation mechanism is introduced. When insufficient population diversity is detected, the mutation rate is automatically increased to prevent the algorithm from falling into a local optimum. The genetic algorithm parameter adaptive adjustment mechanism can intelligently adjust the algorithm behavior according to the search progress, improve the algorithm's search efficiency and optimization quality, and make the generated scheduling strategy closer to the global optimal solution. The optimized genetic algorithm generates a better initial scheduling strategy, which provides a better starting point for the simulated annealing algorithm and helps to further improve the quality of the scheduling strategy.

[0141] Traditional simulated annealing algorithms can get stuck in local optima and converge slowly when solving complex optimization problems. Enhanced optimization mechanisms can improve algorithm performance. The scheduling system employs multiple strategies to enhance the simulated annealing algorithm. First, a more efficient neighborhood generation mechanism is designed. In addition to swapping the allocation units of two tasks, operations such as task group migration and resource reallocation are introduced to expand the search space. Second, the temperature decay function is improved, employing a nonlinear decay strategy that rapidly reduces the temperature in the early stages of the search to accelerate convergence and then slowly reduces the temperature in the later stages to refine the search for the optimal solution. Furthermore, a memory mechanism is introduced to record historical optimal solutions. When the algorithm gets stuck in a local optimum, a better solution is selected from the memory with a certain probability as a new starting point. For example, if the algorithm fails to find a better solution after 10 consecutive iterations, the scheduling system randomly selects a better solution from the memory and restarts the search. The enhanced simulated annealing algorithm can more effectively escape local optima, increasing the probability of finding the global optimal solution. It also accelerates convergence and shortens the time it takes to generate scheduling policies. The resulting high-quality scheduling policies provide a better initial solution for flexible deployment, reducing the workload of subsequent risk calculation unit identification and sharding optimization, and improving the efficiency of the entire scheduling system.

[0142] Example 2

[0143] like Figure 4 As shown, a flow chart of a method for deploying a heterogeneous computing cluster is provided for an embodiment of the present application, and the method includes:

[0144] Obtain the instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption of each computing unit to construct the hardware capability vector of each computing unit;

[0145] The efficiency of each computing unit is predicted through a bidirectional long short-term memory network, and the efficiency of each computing unit is connected through a graph convolutional network to determine the resource vector of the heterogeneous computing cluster;

[0146] Count the task sets and extract the task features of the task sets to obtain the task vectors of the heterogeneous computing cluster;

[0147] The directional similarity and numerical matching of task vectors and resource vectors are integrated to generate an adaptation matrix, and a scheduling strategy for heterogeneous computing clusters is generated based on multi-objective optimization.

[0148] Simulate the spatiotemporal efficiency distribution of different deployment scheduling strategies and identify computing units with resource competition risks based on the spatiotemporal efficiency distribution.

[0149] Perform elastic sharding on the risk computing unit, generate a sharding instruction set that includes sharding granularity, isolation boundaries, and communication routing tables, and output a deployment energy efficiency report;

[0150] Real-time monitoring of the operating status of different scheduling strategies, including hardware temperature curves and network packet loss rates, is used to update scheduling strategies based on comprehensive judgments based on the operating status and deployment energy efficiency reports.

[0151] Since the principle of solving the problem by the method in the embodiment of the present application is similar to that of the system described above in the embodiment of the present application, the implementation of the method refers to the implementation of the system, and the repeated parts will not be repeated.

Claims

1. Heterogeneous computing cluster collaborative scheduling system, characterized by: include: Obtain the instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption of each computing unit to construct a hardware capability vector for each computing unit. The performance of each computing unit is predicted using a bidirectional long short-term memory network. A graph convolutional network is used to connect the performance of each computing unit to determine the resource vector of the heterogeneous computing cluster. By collecting statistics on task sets and extracting their task features, we can obtain the task vectors of heterogeneous computing clusters. We can then integrate the directional similarity and numerical matching between the task vectors and resource vectors to generate an adaptation matrix. Finally, we can generate a scheduling strategy for the heterogeneous computing cluster based on multi-objective optimization. Simulate the spatiotemporal efficiency distribution of different scheduling strategies, identify risk calculation units with resource competition based on the spatiotemporal efficiency distribution, elastically shard the risk calculation units, generate a sharding instruction set, which includes sharding granularity, isolation boundaries, and communication routing tables, and output a deployment energy efficiency report. At the same time, monitor the operating status of different scheduling strategies in real time, including the hardware temperature curve and network packet loss rate. Make a comprehensive judgment based on the operating status and deployment energy efficiency report to update the scheduling strategy.

2. The heterogeneous computing cluster collaborative scheduling system according to claim 1, characterized in that: The logic for determining the resource vector of the heterogeneous computing cluster includes: Configure a multi-head attention mechanism to build a bidirectional long short-term memory network, and train the bidirectional long short-term memory network using historical capability vectors to predict the performance of each computing unit; Through the graph convolutional network, each computing unit is regarded as a node, and the physical connection relationship between computing units is regarded as the edge between nodes to build a cluster resource graph. The node attributes represent the performance of each computing unit. The weights of the edges in the cluster resource graph are determined based on the physical distance between computing units and the correlation of hardware capability vectors. The cluster resource graph is iteratively updated through a graph convolutional network to determine the resource vectors of the heterogeneous computing cluster. Monitor the actual performance of each computing unit, record the difference between the predicted performance and the actual performance of each computing unit through the error matrix, and calculate the average error of the error matrix. Based on the average error of the error matrix, trigger the parameter adjustment of the bidirectional long short-term memory network and update the resource vector of the heterogeneous computing cluster.

3. The heterogeneous computing cluster collaborative scheduling system according to claim 2, characterized in that: The sub-logic for constructing the hardware capability vector of each computing unit includes: A heterogeneous computing cluster includes multiple computing units. The instruction throughput, floating-point peak value, memory bandwidth, and thermal design power consumption of each computing unit are obtained and standardized. Assign weights to instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption based on task type, and construct a hardware capability vector for each computing unit by weighted summation. An update cycle is set, and the fluctuation range of instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption is calculated within the update cycle. The distribution of weight values is updated according to the fluctuation range, and the hardware capability vector of each computing unit is reconstructed.

4. The heterogeneous computing cluster collaborative scheduling system according to claim 3, characterized in that: The generation logic of the scheduling strategy includes: Taking task delay, resource balance and energy consumption cost as optimization targets, normalizing the optimization targets and determining the target weights of the optimization targets according to the task type and system status to form an objective function; Based on the genetic algorithm, each chromosome represents a scheduling scheme, each gene represents a task, and the gene value represents the computing unit assigned to each task. The fitness value of each chromosome is calculated according to the objective function, and the initial scheduling strategy is generated through selection, crossover and mutation iteration based on the fitness matrix. By exchanging the computing units assigned to the two tasks, adjacent candidate scheduling strategies are generated, and the scheduling strategies are optimized using the simulated annealing algorithm; During the execution of the scheduling strategy, the load status of the heterogeneous computing cluster is monitored in real time, and the scheduling strategy is dynamically selected based on the load status.

5. The heterogeneous computing cluster collaborative scheduling system according to claim 4, characterized in that: The generation sub-logic of the adaptation matrix includes: Count the task sets and extract their task features, including computational intensity, data dependency, time sensitivity, and communication topology. These features are then standardized and converted into task vectors for heterogeneous computing clusters. Assign dimension weights to task vectors and resource vectors based on task types, and calculate the directional similarity and numerical matching between task vectors and resource vectors based on the dimension weights. The direction similarity and numerical matching degree are weighted and fused to obtain the adaptation score of each task-resource pair. The adaptation matrix is generated based on the adaptation scores of all task-resource pairs. The adaptation matrix is dynamically optimized through reinforcement learning to update the adaptation matrix.

6. The heterogeneous computing cluster collaborative scheduling system according to claim 5, characterized in that: The generation logic of the sharding instruction set includes: Based on the resource usage trend of each risk calculation unit, the resource pressure gradient of each risk calculation unit is calculated and risk calculation units with high resource pressure gradients are identified to prioritize elastic sharding and determine the sharding granularity. At the same time, data flow analysis is performed on the task sets running on the risk calculation units to identify the dependencies between tasks and dynamically adjust the sharding granularity based on the dependencies between tasks. A resource dependency conflict graph is constructed using tasks as nodes and resource competition intensity as edges. The edge weights are determined based on the shared resource ratio and resource competition probability between tasks. The minimum cut set is then found in the resource dependency conflict graph to identify the isolation boundary of the shard and set the resource threshold for the isolation boundary. Analyze the data transmission requirements between tasks in the shards, including transmission volume, frequency, and timeliness. Evaluate routing costs based on the cluster resource map and the load status of the heterogeneous computing cluster. Combine these requirements with routing costs to plan inter-shard communication paths. These paths must include at least one primary path and two backup paths, forming a communication routing table. The sharding granularity, isolation boundary and communication routing table are structured and encoded, and the sharding instruction set is generated with the identifier of the risk calculation unit as the index. The feasibility of the sharding instruction set is verified to update the sharding instruction set.

7. The heterogeneous computing cluster collaborative scheduling system according to claim 6, characterized in that: The identification sub-logic of the risk calculation unit includes: Simulate the deployment of scheduling policies, record the cluster resource graph during the deployment of each scheduling policy, and predict the resource usage trend of each computing unit. Combine the cluster resource graph and resource usage trend to form the spatiotemporal performance distribution of each scheduling policy. Compare the local density of each node in the cluster resource graph to identify spatial anomaly nodes, and analyze the resource usage trend of each computing unit to identify temporal anomaly nodes. Mark computing units that meet both spatial and temporal anomaly node requirements as potential contention points. Based on historical failure data, the resource contention probability of the cluster resource graph is learned, and multiple simulations are performed on each potential contention point to determine the resource contention intensity of each potential contention point. The competition points are comprehensively judged based on the resource competition intensity and resource usage trend of each potential competition point to identify the risk calculation units where resource competition exists.

8. The heterogeneous computing cluster collaborative scheduling system according to claim 7, characterized in that: The update logic of the scheduling strategy includes: Parse and extract the characteristics of task delay distribution and resource usage distribution in the deployment energy efficiency report to obtain the standard deviation of the average task delay and resource utilization rate. Simultaneously process and extract the characteristics of the hardware temperature curve and network packet loss rate in the running state to obtain the proportion of high-temperature nodes and the proportion of high-packet-loss routes. Determine the indicator weights for average task delay, standard deviation of resource utilization, proportion of high-temperature nodes, and proportion of high-packet-loss routes based on task type. Determine the indicator score through weighted summation. Determine whether to update the scheduling strategy based on the indicator score. When the scheduling policy needs to be updated, the adjustment direction of the scheduling policy is determined by combining the average task latency, the standard deviation of resource utilization, the proportion of high-temperature nodes, and the proportion of routes with high packet loss. The adjustment direction is prioritized according to the task type. Adapt the local adjustment mechanism and the global optimization mechanism according to the adjustment direction to update the scheduling strategy, simulate the updated scheduling strategy, and simultaneously determine the indicator score of the updated scheduling strategy within the monitoring period to determine whether to update the scheduling strategy again.

9. The heterogeneous computing cluster collaborative scheduling system according to claim 8, characterized in that: The local adjustment mechanism includes reselection of computing units and regeneration of shard instruction sets; the global optimization mechanism includes regeneration of the initial scheduling strategy after adjusting the crossover rate and mutation rate of the genetic algorithm.

10. A method for deploying a heterogeneous computing cluster, implemented based on the heterogeneous computing cluster collaborative scheduling system according to any one of claims 1 to 9, characterized in that: include: Obtain the instruction throughput, floating-point peak, memory bandwidth, and thermal design power consumption of each computing unit to construct the hardware capability vector of each computing unit; The efficiency of each computing unit is predicted through a bidirectional long short-term memory network, and the efficiency of each computing unit is connected through a graph convolutional network to determine the resource vector of the heterogeneous computing cluster; Count the task sets and extract the task features of the task sets to obtain the task vectors of the heterogeneous computing cluster; The directional similarity and numerical matching of task vectors and resource vectors are integrated to generate an adaptation matrix, and a scheduling strategy for heterogeneous computing clusters is generated based on multi-objective optimization. Simulate the spatiotemporal efficiency distribution of different deployment scheduling strategies and identify computing units with resource competition risks based on the spatiotemporal efficiency distribution. Perform elastic sharding on the risk computing unit, generate a sharding instruction set that includes sharding granularity, isolation boundaries, and communication routing tables, and output a deployment energy efficiency report; Real-time monitoring of the operating status of different scheduling strategies, including hardware temperature curves and network packet loss rates, is used to update scheduling strategies based on comprehensive judgments based on the operating status and deployment energy efficiency reports.

Images