Time division multiplexing task scheduling method and system for multi-card computing cluster

Abstract

The invention provides a multi-card computing cluster-oriented time-sharing multiplexing task scheduling method and system, and belongs to the field of multi-card computing clusters, and the method comprises the following steps: firstly, deploying a cluster edge unit, and obtaining computing cluster data in parallel; meanwhile, working characteristic data is constructed according to calculation cluster data and system state data, working cost evaluation data is generated based on a cost utility function, then two-stage coding is carried out, an initial task scheduling scheme set is generated based on a chaos algorithm, and the initial task scheduling scheme set is processed according to a swarm intelligence algorithm; meanwhile, an optimal task scheduling scheme set is generated in combination with a chaos algorithm, then the multi-card computing cluster works based on an optimal task scheduling scheme, actual power consumption cost data and an operation report are obtained, finally, an imitation learning network is constructed, and the optimal task scheduling scheme is rapidly generated through the imitation learning network. According to the task scheduling method, energy consumption and cooling perception can be fused, and the optimal energy consumption can be obtained under the condition that cooling is kept.

Description

Technical Field

[0001] This invention relates to the field of multi-GPU computing cluster technology, and more specifically, to a time-sharing multitasking task scheduling method and system for multi-GPU computing clusters. Background Technology

[0002] Currently, large-scale GPU computing clusters have become a key infrastructure for industrial simulation and large model training. Efficient task scheduling strategies are an important part of releasing cluster computing power and reducing operating costs. Their performance directly determines the utilization efficiency of computing resources. However, with the continuous expansion of cluster size and the increasing complexity and diversity of workloads, the limitations of traditional static or simple rule-based scheduling methods are becoming increasingly apparent when dealing with dynamic and heterogeneous computing environments.

[0003] On the one hand, traditional task scheduling methods, such as those based on resource queue polling or greedy strategies, often lack forward-looking decision-making capabilities. They mostly allocate resources based on the task's immediate resource requests and the current static state of the nodes, and cannot promptly perceive the dynamic effects caused by task execution. This makes it easier for problems such as hot accumulation and high energy consumption to occur within the cluster.

[0004] On the other hand, although there are methods that model scheduling problems as combinatorial optimization and apply swarm intelligence algorithms, these methods often treat computational energy consumption and cooling as separate objectives. In actual operation, the computational load is large, and the cooling problem is easily ignored when searching for the optimal solution.

[0005] In summary, there is an urgent need for a scheduling method that can integrate energy consumption and cooling awareness to achieve optimal energy consumption while maintaining cooling and reducing computational load during processing. Summary of the Invention

[0006] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a time-sharing multiplexing task scheduling method for multi-GPU computing clusters, the method comprising: S100 deploys cluster edge units, acquires computing cluster data in parallel based on cluster edge units, constructs working characteristic data based on computing cluster data and system status data, and generates working cost evaluation data based on cost-utility function. S200, perform two-stage coding and generate an initial task scheduling scheme set based on the chaos algorithm, process the initial task scheduling scheme set according to the swarm intelligence algorithm, and generate an optimized task scheduling scheme in combination with the chaos algorithm. The S300 multi-GPU computing cluster operates based on an optimized task scheduling scheme and acquires actual power consumption cost data and operation reports. S400: During the execution of steps S100 to S300, learning enhancement data is collected, and an imitation learning network is constructed based on the learning enhancement data. The imitation learning network operates based on an imitation strategy and is used to imitate and generate an optimized task scheduling scheme.

[0007] As a further aspect of the present invention, cluster edge units are deployed, and computing cluster data is acquired in parallel based on the cluster edge units. Simultaneously, operational characteristic data is constructed based on the computing cluster data and system status data, and operational cost assessment data is generated based on a cost-utility function, including: The cluster edge unit is deployed for each computing unit in the computing cluster. Based on the cluster edge unit, the performance data and unit task information data of each computing unit are obtained, and the performance data and unit task information data are encapsulated into computing cluster data. Acquire cooling status data and cluster task information data, and encapsulate the cooling status data and cluster task information data into system status data; Data processing is performed on the computing cluster data and the system status data to obtain structured cluster data. Based on the structured cluster data, features are constructed to obtain working feature data. A short-term operating prediction model is configured for each of the cluster edge units. The operating characteristic data is input into the short-term operating prediction model to obtain mixed prediction data, which includes predicted power consumption data, predicted cooling temperature data, and predicted time data. A cost-utility function is constructed, and work cost assessment data is obtained based on the hybrid prediction data.

[0008] As a further aspect of the present invention, data processing is performed on the computing cluster data and the system status data to obtain structured cluster data, and feature construction is performed based on the structured cluster data to obtain working feature data, including: The computing cluster data and the system status data are validated and cleaned for data validity, and the computing cluster data is aggregated to obtain structured cluster data. At the same time, the structured cluster data is standardized. Dynamic load characteristic data is obtained from computing cluster data in structured cluster data; task load characteristic data is obtained from computing cluster data and system status data in structured cluster data; and cross-characteristic data is obtained from computing cluster data in structured cluster data. Work characteristic data is constructed based on dynamic load characteristic data, task load characteristic data, and cross characteristic data.

[0009] As a further aspect of the present invention, two-stage encoding is performed and an initial task scheduling scheme set is generated based on a chaotic algorithm. The initial task scheduling scheme set is then processed according to a swarm intelligence algorithm, and an optimized task scheduling scheme is generated by combining the chaotic algorithm, including: Two-stage encoding is performed on the unit task information data in the computing cluster data and the cluster task information data in the system status data to obtain the initial task scheduling scheme definition and solution space definition. The initial task scheduling scheme definition and solution space definition are imported into the chaotic algorithm. The initial task scheduling scheme set is generated iteratively based on the chaotic algorithm, and the diversity of the initial task scheduling scheme set is verified. The initial task scheduling scheme set is imported into the swarm intelligence algorithm as the initial population. At the same time, the swarm intelligence algorithm calls the cost-utility function and the short-term job prediction model of the cluster edge unit to obtain the fitness of each individual in the initial population. The swarm intelligence algorithm updates the population based on fitness and monitors chaos simultaneously. Based on the chaos monitoring, it triggers a chaos perturbation strategy, which includes chaotic individual perturbation and chaotic swarm perturbation. Obtain an optimized task scheduling solution.

[0010] As a further aspect of the present invention, the initial task scheduling scheme definition and solution space definition are imported into a chaotic algorithm, an initial task scheduling scheme set is iteratively generated based on the chaotic algorithm, and the diversity of the initial task scheduling scheme set is verified, including: Select the Logistic mapping and set the initial values ​​for chaos; Run the Logistic mapping and iteratively generate a chaotic random sequence based on the Logistic mapping. Map the chaotic random sequence to the encoding space to obtain an initial task scheduling scheme set. Obtain the distribution metric and average diversity of the initial task scheduling scheme set, and compare the distribution metric and average diversity with preset conditions.

[0011] As a further aspect of the present invention, the swarm intelligence algorithm performs population updates based on fitness and simultaneously monitors chaos. Based on the chaos monitoring, a chaos perturbation strategy is triggered. This chaos perturbation strategy includes individual chaos perturbation and swarm chaos perturbation, comprising: Define stagnation criteria and monitoring criteria, wherein the stagnation criteria include, but are not limited to, population aggregation criteria, maximum frequency criteria, and diversity criteria; If the stagnation criterion is met, updates will continue; if the stagnation criterion is not met, the chaotic perturbation strategy will be triggered. Apply chaotic individual perturbation to the historical best solution in the currently updated population, and apply chaotic group perturbation to the individuals in the currently updated population.

[0012] As a further aspect of the present invention, the multi-GPU computing cluster operates based on an optimized task scheduling scheme and acquires actual power consumption cost data and operation reports, including: The feasibility of the optimized task scheduling scheme is verified to obtain the verified scheduling scheme; The verification scheduling scheme is distributed to the corresponding computing units, and resource quotas are allocated to the computing units based on the cluster edge units; During task execution, data is collected based on the cluster edge units to obtain actual power consumption cost data, and an operation report is generated based on the cluster edge units.

[0013] As a further aspect of the present invention, during the execution of steps S100 to S300, learning enhancement data is collected, and an imitation learning network is constructed based on the learning enhancement data. The imitation learning network operates based on an imitation strategy and is used to imitate and generate optimized task scheduling schemes, including: Collect state data, action data, and reward data; perform data augmentation on the state data; and obtain learning augmentation data based on the augmented state data, action data, and reward data. An imitation learning network is constructed based on a graph neural network, and graph nodes and edges are defined. An attention mechanism is added to the imitation learning network. The loss function of the imitation learning network includes a main loss function and an auxiliary loss function. The imitation learning network is trained based on learning reinforcement data. Deploy the trained imitation learning network and simultaneously set a selection strategy. Based on the selection strategy, select to use the imitation learning network or steps S100 to S200 to obtain an optimized task scheduling scheme, and then execute step S300.

[0014] As a further aspect of the present invention, a trained imitation learning network is deployed, and a selection strategy is simultaneously set. Based on the selection strategy, either the imitation learning network or steps S100 to S200 are selected to obtain an optimized task scheduling scheme. Step S300 is then executed, including: The selection strategy includes a slow channel and a fast channel. The slow channel refers to obtaining an optimized task scheduling scheme using steps S100 to S200, and the fast channel refers to obtaining an optimized task scheduling scheme using an imitation learning network.

[0015] Furthermore, embodiments of the present invention also provide a time-sharing multiplexing task scheduling system for multi-GPU computing clusters, comprising: Cluster edge unit, which is used to acquire computing cluster data in parallel; The evaluation module is used to construct working characteristic data and generate working cost evaluation data based on computing cluster data and system status data; An optimization module is used to generate an initial task scheduling scheme set and process the initial task scheduling scheme set based on swarm intelligence algorithm and chaos algorithm to generate an optimized task scheduling scheme. The central processing unit is used to allocate tasks based on an optimized task scheduling scheme and to obtain actual power consumption cost data and operation reports; The imitation module is used to construct an imitation learning network, which is used to imitate and generate optimized task scheduling schemes.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, data from the computing cluster is collected through the cluster edge unit. This data, combined with system status data and a short-term prediction model, generates workload assessment data, providing forward-looking data for subsequent scheduling decisions. Next, an initial task scheduling scheme set is generated based on a chaotic algorithm. This initial scheme set is then processed using a swarm intelligence algorithm, and an optimized task scheduling scheme is generated using the chaotic algorithm. A hybrid swarm intelligence algorithm, incorporating both chaotic and efficient methods, is used for scheduling search. During the search, the fusion of energy consumption and cooling control objectives is considered to evaluate strategies, ensuring that the optimized task scheduling scheme achieves both energy consumption and cooling goals. The optimized task scheduling scheme is then deployed for execution. Finally, an imitation learning network is constructed to distill the scheduling decisions from the swarm intelligence and chaotic algorithms. This network enables rapid decision-making, accelerating computation and reducing computational load while maintaining real-time performance. This method provides a task scheduling approach that integrates energy consumption and cooling awareness, achieving optimal energy consumption while maintaining cooling. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall steps of a time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to the present invention.

[0018] Figure 2 This is a flowchart of step S200 of the time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to the present invention.

[0019] Figure 3 This is a schematic diagram of a time-sharing multitasking task scheduling system for multi-GPU computing clusters according to the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating the overall steps of a time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to the present invention. Figure 2 This is a flowchart of step S200 of a time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to the present invention. The following is a detailed description of this time-sharing multiplexing task scheduling method for multi-GPU computing clusters.

[0021] Specifically, a time-sharing multiplexing task scheduling method for multi-GPU computing clusters includes: Step S100: Deploy cluster edge units, acquire computing cluster data in parallel based on cluster edge units, construct working characteristic data based on computing cluster data and system status data, and generate working cost evaluation data based on cost-utility function.

[0022] In this embodiment, step S100 includes: Step S100-1: Deploy the cluster edge unit for each computing unit in the computing cluster, obtain the performance data and unit task information data of each computing unit based on the cluster edge unit, and encapsulate the performance data and the unit task information data into computing cluster data.

[0023] Specifically, the data list to be collected for each cluster edge unit should be clearly defined. The data to be collected in the data list can be divided into performance data and unit task information data for each computing unit. The performance data of each computing unit is represented by the hardware layer data source of each computing unit, which can collect data such as GPU power consumption, GPU utilization, GPU temperature, and unit fan speed. The unit task information data is represented by the task layer data source, which can collect data such as task list, resource requirements of each task, and currently running tasks.

[0024] Understandably, a cluster edge unit refers to a lightweight data acquisition agent deployed on each computing unit. A unified acquisition agent program can be deployed on each computing unit in the form of Docker, etc. After deployment, the list of data to be collected by each cluster edge unit needs to be configured, and then the acquisition frequency of each cluster edge unit can be configured. In the specific implementation process, different acquisition frequencies can be set based on the rate of data change, thereby achieving hierarchical acquisition to balance load and real-time performance. For example, data can be divided into high-frequency and medium-frequency data. High-frequency acquisition can be used to collect GPU power consumption, GPU utilization, and GPU temperature, while medium-frequency acquisition can be used to collect GPU utilization, task lists, etc. Based on the data to be collected, an appropriate corresponding frequency is set. After the acquisition frequency and acquisition data are configured, each cluster edge unit can obtain performance data and unit task information data through the corresponding interface or command tool, and then package and encapsulate the performance data and unit task information data.

[0025] Understandably, for multi-GPU computing clusters, centralized data acquisition methods may be insufficient to meet the real-time requirements of large-scale clusters. By deploying cluster edge units locally on each computing unit, distributed and parallel data acquisition can be achieved, distributing the acquisition pressure to each computing unit and thus enabling faster response.

[0026] Step S100-2: Obtain cooling status data and cluster task information data, and encapsulate the cooling status data and cluster task information data into system status data.

[0027] Specifically, a list of cooling status data and cluster task information data to be collected is defined. The cooling status data represents the current operating status information of the cooling system, and can include data such as data room environment, power consumption and cooling operation data of the rack cooling unit. The cluster task information data is similar to the unit task information data mentioned above, and is used to represent the task list and the list of tasks to be assigned in the central processing unit of the multi-GPU computing cluster, as well as the data such as the resources required by the tasks. After the data is acquired, the cooling status data and cluster task information data can be packaged and packaged.

[0028] Step S100-3: Perform data processing on the computing cluster data and the system status data to obtain structured cluster data, and construct features based on the structured cluster data to obtain working feature data.

[0029] Step S100-3 includes: Step S100-3-1: Perform data validity verification and data cleaning on the computing cluster data and the system status data, and perform data aggregation on the computing cluster data to obtain structured cluster data. At the same time, perform standardization operations on the structured cluster data.

[0030] It is understandable that noise and erroneous data will inevitably exist during the collection of raw data. If erroneous data is retained for decision-making in subsequent steps, it will lead to decision-making errors and cause problems such as scheduling errors. Therefore, after the collection of computing cluster data and system status data is completed, it is necessary to perform data validity verification and data cleaning.

[0031] Specifically, data validity verification can be understood as range verification, mutation filtering verification, and integrity verification. For range verification, it means checking the computing cluster data and system status data within a reasonable physical range. If the computing cluster data and system status data exceed the reasonable physical range, they are cleaned. For example, the operating temperature of a typical GPU is around 0-100 degrees Celsius. If the GPU's operating temperature is 150 degrees Celsius at a certain moment, it is cleaned and short-term interpolation is performed based on historical data. For mutation filtering verification, it means identifying instantaneous spike signals. For example, if the GPU power consumption rises from 250W to 700W and then drops back instantly between two consecutive sampling moments, such data can be cleaned and then interpolated using a simple sliding window average or median. For integrity verification, it means checking whether all the data in the data list that needs to be collected has been collected within the expected time. If a piece of data is missing, it can be filled using the same short-term interpolation strategy based on historical data as range verification.

[0032] Furthermore, once the data validity verification and data cleaning are completed, the computing cluster data can be aggregated. Since the collected multi-GPU computing cluster data is large, the computing cluster data can be aggregated, and several key statistical values ​​or representative values ​​can be aggregated within a preset time window. This will compress the computing cluster data and reduce the pressure on network bandwidth and the central processing unit.

[0033] For example, data aggregation can be performed using methods such as statistical aggregation and derived aggregation. For statistical aggregation, methods such as average value, maximum value, minimum value, and sum can be selected. For example, the average power consumption and average temperature of the GPU in each computing unit within a preset time window can be calculated through the cluster edge unit. For derived aggregation, methods such as temperature gradient, power efficiency, or load balancing can be selected. For example, the maximum temperature difference between GPUs in each computing unit can be calculated through the cluster edge unit to characterize the temperature gradient and reflect the uniformity of heat dissipation. The specific method selected can be based on the specific implementation process or on the subsequent step S100-3-2. Since step S100-3-2 involves feature construction, it needs to be constructed based on the computing cluster data and system status data. The aggregation method can be adaptively selected based on the features that need to be constructed, so that the constructed features can better describe the overall operating status. After the data aggregation is completed, the computing cluster data and system status data can be packaged and encapsulated into structured cluster data.

[0034] Understandably, after acquiring structured cluster data, due to slight delays in acquisition and transmission, related data within the structured cluster data may not be completely synchronized. Therefore, it is necessary to align the time window of the structured cluster data. This can be achieved by defining a short time window and treating all data within that window as being in the same state at the same moment. This ensures that the data used in subsequent steps is acquired at the same point in time or within a very short time interval, avoiding causal inconsistencies caused by time asynchrony. For example, using the temperature 10 seconds later to match the power consumption 10 seconds earlier. Furthermore, since the related data in the structured cluster data have different units and scales—for example, GPU power consumption is measured in watts (typically ranging from 0-500 watts), and GPU temperature is measured in degrees Celsius (typically ranging from 30-100 degrees Celsius)—it is necessary to use relevant methods to normalize the data in the structured cluster data. Common normalization methods such as Z-Score and Min-Max can be applied.

[0035] Step S100-3-2: Obtain dynamic load characteristic data based on computing cluster data in structured cluster data; obtain task load characteristic data based on computing cluster data and system status data in structured cluster data; and obtain cross-characteristic data based on computing cluster data in structured cluster data.

[0036] Among them, dynamic load characteristic data refers to the dynamic behavior and historical inertia of computing units, which is obtained based on performance data in computing cluster data, such as the average power consumption of GPUs in the past 5 minutes and the slope of GPU temperature change in the past 5 minutes. Corresponding to step S100-3-1, if statistical aggregation is used when data aggregation is performed in step S100-3-1, and the average power consumption of GPUs of computing units in each preset time window is selected, the average power consumption of GPUs in the past 5 minutes can be directly obtained based on the average power consumption of GPUs of computing units in the above preset time window. The task load characteristic data and cross characteristic data will not be elaborated on further in the following sections.

[0037] Among them, task load characteristic data represents the current and recent load type and pressure on computing units or overall tasks. It is obtained based on performance data and unit task information data in computing cluster data and cluster task information data in system status data. It can be divided into task combination characteristics and queue pressure characteristics. Task combination characteristics can describe the task load on the current computing unit. Specifically, it encodes the type of task currently running on the GPU, such as AI training, simulation, etc. Different task types have different pressure on the GPU and can be identified from the task tags obtained from the task management system. Queue pressure characteristics can describe the future task load, such as the number of tasks in the waiting queue, which can be pulled from the central task queue of the central processor.

[0038] Among them, cross-feature data represents the interaction of multiple raw data, describing the hidden physical relationships or efficiency indicators within the computing unit. It is obtained based on the cooling status data of the performance data and system status data in the computing cluster data. For example, the difference between the highest and lowest temperatures of the GPU in the computing unit. If this difference is larger, it indicates that the internal heat dissipation and load of the computing unit are more unbalanced. It can also be used to reflect the efficiency of the cooling system. It can be expressed in the form of (highest GPU temperature of computing unit - air intake temperature or liquid inlet temperature of cooling system) / total power consumption of GPU in computing unit. Under the same power consumption, the smaller the temperature rise, the higher the cooling efficiency.

[0039] Step S100-3-3: Construct working characteristic data based on dynamic load characteristic data, task load characteristic data, and cross characteristic data.

[0040] Understandably, combining dynamic load characteristic data, task load characteristic data, and cross-characteristic data to construct working characteristic data can provide a relatively comprehensive description of the current and recent operating status of the multi-GPU computing cluster, CPU, and cooling system. Furthermore, working characteristic data serves as input for subsequent short-term job prediction models, helping these models better understand the physical laws hidden within the data.

[0041] Step S100-4: Configure a short-term operating prediction model for each of the cluster edge units, input the operating characteristic data into the short-term operating prediction model, and obtain mixed prediction data, which includes predicted power consumption data, predicted cooling temperature data, and predicted time data.

[0042] Furthermore, a short-term job prediction model is deployed and configured for each cluster edge unit. The short-term job prediction model is used to predict the GPU power consumption, GPU temperature and the end time of the current task within the next AB minutes based on job characteristic data, that is, predict power consumption data, predict cooling temperature data and predict time data. For the characteristics of GPU power consumption, GPU temperature and task end time prediction, three types of lightweight gradient boosting machine, lightweight temporal convolutional network or LSTM, vector regression model or SARIMAX can be selected. Different cluster edge units can be configured according to the actual situation. Each model is trained and calibrated based on the historical data of the corresponding computing unit.

[0043] Among them, for lightweight gradient boosters, LightGBM and XGBoost can be selected. These models are good at processing tabular data, are relatively flexible in feature engineering, and have high prediction accuracy and fast inference speed. For lightweight temporal convolutional networks or LSTM, if the temporal dependence is strong, TCN or LSTM can be selected, but more training data is required and the training cost is high. As for the final vector regression model or SARIMAX, it is a traditional statistical temporal model, which is more lightweight and suitable for use cases with strong linear relationships.

[0044] Furthermore, after selecting a suitable model for each cluster edge unit, a review window can be set to determine how much historical data the model is based on for prediction. For example, the model may need a window of a preset time within the past 2 minutes, and a corresponding review window can be set. Then, working feature data of the review window length is extracted from the data stream and used as input to the model. In some possible embodiments, in addition to inputting the corresponding working feature data, static attributes of the GPU, such as GPU model, compute node number, and rack location code in the cooling system, can also be obtained as part of the input. This helps the model distinguish the behavior of different compute nodes. It is understandable that changes in power consumption and temperature have inertia, and future states depend on recent historical states. Providing a review window and using it to obtain the data needed by the model is the basis for the model to learn dynamic patterns and make data predictions.

[0045] Furthermore, after the above operations are completed, the predicted power consumption data, predicted cooling temperature data, and predicted time data can be acquired based on the short-term working prediction model. After each cluster edge unit has acquired the working feature data, it can be input into the short-term working prediction model in parallel. Alternatively, the static attributes mentioned above can be input into the working prediction model synchronously and in parallel. Each working prediction model receives its corresponding working feature data and performs fast forward inference, thereby outputting the predicted power consumption data, predicted cooling temperature data, and predicted time data corresponding to each working prediction model. The predicted power consumption data, predicted cooling temperature data, and predicted time data are packaged and encapsulated into hybrid prediction data.

[0046] Step S100-5: Construct a cost-utility function and obtain work cost assessment data based on the hybrid prediction data.

[0047] Specifically, to construct the cost-utility function, the first step is to construct a cooling energy consumption term, which represents the power required for a multi-GPU computing cluster to complete its current task. This is achieved by summing all predicted power consumption data to obtain the total predicted heat load of the entire multi-GPU computing cluster in the future, and then using a corresponding physical simplification model to map the heat load to cooling requirements. For example, a relatively simple cooling energy consumption term can be expressed as: Cooling Energy Consumption = Total Predicted Heat Load / Cooling System Energy Efficiency Ratio. The cooling system energy efficiency ratio can be defined based on the actual cooling system in use. For example, for a liquid-cooled cooling system, specific calculations are required based on factors such as the specific heat capacity, volumetric flow rate, and density of the cooling medium. The energy consumption required by the cooling system can be calculated using the aforementioned cooling energy consumption term.

[0048] Furthermore, after constructing the cooling energy consumption item, a cost function can be built based on it. This cost function can be expressed as the sum of weight a * cooling energy consumption item + weight b * predicted power consumption data, where weight a and weight b are the weight coefficients of the sum of cooling energy consumption item and predicted power consumption data, respectively. Weights a and b are constants set based on actual user needs. If weight a is greater than weight b, computational power consumption is encouraged to take precedence over cooling system power consumption; conversely, if weight b is greater than weight a, cooling system power consumption is encouraged to take precedence over computational power consumption. The construction of the cost function provides a single, comparable value for subsequent optimization algorithms and guides the search direction. Next, constraints can be constructed. Understandably, typical optimization problems require constraints, such as GPU temperature not exceeding a temperature threshold or tasks needing to be completed before a deadline. In some possible embodiments, temperature constraints and time constraints can be selected to set constraints. For each constraint, a function is defined to measure the degree to which it is violated. For example, the temperature constraint can be expressed as penalty coefficient * predicted temperature * temperature penalty, and the time constraint can be expressed as penalty coefficient * predicted time * time penalty. The temperature penalty ranges from MAX(0, predicted highest temperature - preset maximum temperature) to MAX(0, predicted time data - task deadline). The penalty coefficient is a large positive number to ensure that the total cost of any solution that violates the constraint becomes extremely high, thus being eliminated by subsequent optimization algorithms. The final cost-utility function can be simply expressed as cost function + temperature constraint + time constraint. Substituting the mixed prediction data into the cost-utility function yields the work cost assessment data.

[0049] Step S200: Perform two-stage coding and generate an initial task scheduling scheme set based on the chaos algorithm. Process the initial task scheduling scheme set according to the swarm intelligence algorithm, and at the same time generate an optimized task scheduling scheme by combining the chaos algorithm.

[0050] In this embodiment, step S200 includes: Step S200-1: Based on the unit task information data in the computing cluster data and the cluster task information data in the system status data, perform two-stage encoding to obtain the initial task scheduling scheme definition and solution space definition.

[0051] Understandably, the execution of a complete scheduling scheme is determined by two important core decision variables, namely the two phases of the two-phase encoding mentioned in step S200-1. One phase can be represented as task mapping, which describes which computing unit each task waiting to be scheduled should be assigned to, i.e., the mapping relationship from task to computing unit. The other phase can be represented as execution order, which describes the order in which multiple tasks assigned to the same computing unit should be executed, i.e., the mapping relationship from task to sequence position. These somewhat ambiguous scheduling needs to be specified as mathematical variables that can be recognized by the optimization algorithm.

[0052] Specifically, for the task mapping encoding method, an array of length N can be used, where N is the total number of tasks to be scheduled. The array index represents the task ID, and the value at the index represents the GPU number to which the task is assigned. For example, assuming there are 4 tasks (T0, T1, T2, T3) and 3 computing units, the encoding is [2, 0, 2, 1]. This means that task T0 is assigned to computing unit 2, task T1 is assigned to computing unit 0, task T2 is assigned to computing unit 2, and task T3 is assigned to computing unit 1. For the execution order encoding method, an array can also be used, but priority-based encoding is selected. The array index represents the task ID, and the value at the index represents a real number representing the priority. For all tasks assigned to the same computing unit, they are sorted in descending order of priority value. The larger the value, the higher the priority of the task. After the encoding is completed, the initial task scheduling scheme is defined.

[0053] In addition, the solution space needs to be defined. The set of all possible codes constitutes the solution space, and the range of the solution space is the range of the two-stage coding. Then, the hard constraints of the solution space are defined, which are divided into resource constraints and dependency constraints. For resource constraints, a task cannot be assigned to multiple computing units at the same time, and the total number of tasks running on a computing unit cannot exceed its maximum concurrency. For dependency constraints, if there is a dependency relationship between tasks, such as a task must start after another task is completed, it is necessary to check whether the order is legal. Otherwise, a large penalty term is given to ensure that the solution found by the subsequent optimization algorithm is reasonable and feasible. The hard constraints must be satisfied.

[0054] Step S200-2: Import the initial task scheduling scheme definition and solution space definition into the chaotic algorithm, generate the initial task scheduling scheme set iteratively based on the chaotic algorithm, and perform diversity verification on the initial task scheduling scheme set.

[0055] In this embodiment, step S200-2 includes: Step S200-2-1: Select Logistic mapping and set the initial values ​​for chaos.

[0056] Understandably, the Logistic mapping is a chaotic algorithm used to generate random sequences. It is relatively simple and has typical chaotic characteristics. Traditional standard swarm intelligence algorithms may easily get trapped in local optima if the initial population happens to be clustered near a local optimum in the solution space. However, the Logistic mapping can generate deterministic but unpredictable random sequences that are sensitive to initial conditions. The generated sequences are ergodic, random, and regular, and can pass through all states without repetition within a certain range. It is more suitable for generating an initial task scheduling scheme set. Generating an initial task scheduling scheme set based on the Logistic mapping can make it more diverse, thus providing a better starting point for subsequent swarm intelligence optimization algorithms and increasing the probability of finding the global optimum.

[0057] Specifically, when generating random sequences using the Logistic mapping, it is necessary to set the initial value of chaos. Generally, the initial value of chaos should not be a fixed point of the Logistic mapping, and simple rational numbers, such as 0.1, should be avoided to ensure the chaotic nature of the generated chaos.

[0058] Step S200-2-2: Run the Logistic mapping and iteratively generate a chaotic random sequence based on the Logistic mapping. Map the chaotic random sequence to the encoding space to obtain the initial task scheduling scheme set.

[0059] Understandably, step S200-2-2 is used to generate an initial task scheduling scheme set, which will serve as the initial population during subsequent swarm intelligence algorithm optimization. The Logistic mapping is run, and a chaotic random sequence is generated based on the iterative generation of the Logistic mapping. The first X generated chaotic random values ​​need to be discarded to avoid the influence of the initial values. Then, the generated chaotic random sequence can be input into the value encoding space. Values ​​are taken out sequentially from the chaotic random sequence and filled into the task mapping segment and execution order segment according to the above encoding definition, thereby constructing the initial task scheduling scheme set.

[0060] Step S200-2-3: Obtain the distribution metric and average diversity of the initial task scheduling scheme set, and compare the distribution metric and average diversity with the preset conditions.

[0061] Understandably, after generating the initial task scheduling scheme set, its diversity needs to be verified. If the effect is not ideal, it may affect the subsequent swarm intelligence algorithm. This can be achieved by comparing index quantification and pure random initialization. For index quantification, the initial task scheduling scheme set can be compared and verified. For example, the average Euclidean distance between each pair of individual schemes in the initial task scheduling scheme can be calculated. For pure random initialization comparison, the initial task scheduling scheme set can be compared with the average diversity obtained by pure random initialization. Preset conditions include preset comparison conditions, such as the difference threshold of average diversity and the difference threshold of average Euclidean distance. If the preset conditions are not met, it indicates that the effect of generating the initial task scheduling scheme set by chaos is not affected, and the parameters of the chaos algorithm can be adjusted.

[0062] Step S200-3: The initial task scheduling scheme set is imported into the swarm intelligence algorithm as the initial population. At the same time, the swarm intelligence algorithm calls the cost-utility function and the short-term job prediction model of the cluster edge unit to obtain the fitness of each individual in the initial population.

[0063] In some possible embodiments, the swarm intelligence algorithm can be selected from swarm intelligence algorithms such as PSO or GA. After obtaining the initial task scheduling scheme, the initial task scheduling scheme can be used as the initial population input to the swarm intelligence algorithm. Each individual in the initial population is a task scheduling scheme in the initial task scheduling scheme set. At this time, the short-term job prediction model in the cluster edge unit is called to obtain the mixed prediction data of each task scheduling scheme. At the same time, the cost-utility function is used as the fitness function of the swarm intelligence algorithm, and the mixed prediction data of each task scheduling scheme is input into the fitness function to obtain the fitness of each individual in the initial population.

[0064] Step S200-4: The swarm intelligence algorithm updates the population based on fitness and performs chaos monitoring at the same time. Based on the chaos monitoring, a chaos perturbation strategy is triggered, which includes chaotic individual perturbation and chaotic swarm perturbation.

[0065] Understandably, population update is a fundamental step in swarm intelligence algorithms. Taking the PSO and GA algorithms as examples, they respectively employ individual and group update strategies, and crossover and mutation update strategies. In the PSO algorithm, each individual updates its search direction and new scheduling scheme based on its own historical best position and the group's historical best position. The GA algorithm, on the other hand, uses crossover and mutation operations to exchange some task allocation information among individuals or randomly change the allocation of a certain task in order to search for new regions. In each update, a new fitness is generated, and the few individuals with the highest fitness in each generation are retained for the next update.

[0066] In this embodiment, step S200-4 includes: Step S200-4-1: Define stagnation criteria and monitoring criteria. The stagnation criteria include, but are not limited to, population aggregation criteria, maximum frequency criteria, and diversity criteria.

[0067] Understandably, the stagnation criterion is the trigger condition for the chaotic perturbation strategy. The maximum number of updates criterion is that when the maximum number of updates is reached, the improvement in the fitness value of the historical best solution is less than a preset threshold. The population clustering criterion is to calculate the average distance from all individuals to the population best solution. If it is too small, it indicates that the population is clustered around a certain point. The diversity criterion is to calculate the average distance between all individuals in the population. When this distance is less than a preset threshold, it indicates that the population is too assimilated. As for the monitoring criterion, it can be set in a similar way as after each update or after every two updates.

[0068] Step S200-4-2: If the stagnation criterion is met, the update continues; if the stagnation criterion is not met, the chaotic perturbation strategy is triggered.

[0069] Step S200-4-3: Perform chaotic individual perturbation on the historical best solution in the currently updated population, and perform chaotic group perturbation on the individuals in the currently updated population.

[0070] It is understandable to perturb the current best historical solution found in the population, because the current best solution may be located in a local optimum. Perturbing it may move it away from the local optimum, thereby discovering a better region. For the current population, a certain proportion of individuals in the population are randomly selected for perturbation. When applying the perturbation, it is necessary to preserve elite individuals from being affected by the perturbation. As for the perturbation intensity, a smaller intensity can be used initially. If it is still impossible to break out of the stagnation, the perturbation intensity can be gradually increased.

[0071] Furthermore, a perturbation vector is generated. The perturbation vector is a chaotic sequence of the same length as the individual's code. Based on the chaotic sequence, different parts of the individual's code are perturbed. For example, for the task mapping segment, the chaotic sequence value can be used as a probability to randomly change the task allocation. For example, for the execution order segment, additive perturbation can be used to map the chaotic value to the interval between -1 and 1, realizing perturbation in the positive and negative directions, thereby adjusting the execution order.

[0072] Understandably, compared to ordinary random perturbation, chaotic perturbation, due to its inherent ergodicity and irregularity, can move individuals to unexplored regions in the solution space that differ significantly from the current region, increasing the probability of escaping local optima. The perturbed individuals can be integrated into the population to form the next generation of individuals, or the newly generated individuals can replace individuals with low fitness in the population. Alternatively, the next generation of individuals can be formed by combining the retained elite individuals with the new individuals generated by the perturbation. Once generation is complete, the chaotic perturbation strategy is stopped.

[0073] Step S200-5: Obtain the optimized task scheduling scheme.

[0074] Specifically, the above steps are executed repeatedly, and a termination condition is set. The termination condition can be a quality condition, an iteration condition, a time condition, etc. For example, for the quality condition, the process can stop when a solution with sufficiently high fitness is found, or when the preset maximum number of iterations is reached, and the optimal solution is output, which is the optimized task scheduling scheme to ensure the scheme with the lowest total energy consumption cost.

[0075] In step S300, the multi-GPU computing cluster operates based on an optimized task scheduling scheme and obtains actual power consumption cost data and operation reports.

[0076] In this embodiment, step S300 includes: Step S300-1: Perform a feasibility verification on the optimized task scheduling scheme and obtain the verified scheduling scheme.

[0077] Specifically, before the central processing unit issues the optimized task scheduling scheme, it is necessary to perform a final verification of the optimized task scheduling scheme. If verification is not performed, invalid or erroneous scheduling instructions may occur. During verification, conditions such as resource conflicts can be checked. Taking resource conflicts as an example, before actually issuing instructions, the optimized task scheduling scheme is checked again to see if there are any obvious resource conflicts. For example, it is checked whether two tasks that require 16GB of video memory are simultaneously assigned to a computing unit with only 16GB of video memory. If such a situation occurs, step S200 can be re-executed, or a simple temporary handling strategy can be adopted, such as assigning the task to an empty computing unit. After verification, the optimized task scheduling scheme can be converted into a readable and executable specific command sequence, that is, the verification scheduling scheme.

[0078] Step S300-2: The verification scheduling scheme is distributed to the corresponding computing unit, and resource quotas are allocated to the computing unit based on the cluster edge unit.

[0079] Specifically, after obtaining the verification scheduling scheme, the central processing unit can distribute each task in the scheme to the corresponding computing unit. While the computing unit starts the task, the cluster edge unit can use underlying quota technologies such as cgroups and NVIDIA MIG to set resource usage limits for each task, such as power consumption limits, computing power limits, and video memory limits, as a safety net.

[0080] Understandably, prediction and optimization algorithms cannot be completely accurate. During actual operation, a task may suddenly generate a computational load that exceeds expectations due to data characteristics, causing power consumption and temperature to soar. In this case, resource quotas are needed to force the impact of the task to be limited within a safe range to prevent abnormal fluctuations of a single task from affecting the entire multi-GPU computing cluster.

[0081] Step S300-3: During task execution, data is collected based on the cluster edge unit to obtain actual power consumption cost data, and an operation report is generated based on the cluster edge unit.

[0082] Understandably, during task execution, the cluster edge units continuously collect data from the computing units and generate corresponding operation reports based on the collected data through the central processing unit. These reports may include information such as temperature, actual power consumption, and completion time.

[0083] In step S400, during the execution of steps S100 to S300, learning enhancement data is collected, and an imitation learning network is constructed based on the learning enhancement data. The imitation learning network operates based on an imitation strategy and is used to imitate and generate an optimized task scheduling scheme.

[0084] In this embodiment, step S400 includes: Step S400-1: Obtain state data, action data, and reward data; perform data augmentation on the state data; and obtain learning augmentation data based on the augmented state data, action data, and reward data.

[0085] Understandably, in the initial stage of system operation, relevant data from steps S100 to S300 are continuously recorded, and the collected data is combined into state data, action data, and reward data. The state data comes from the computing cluster data, system state data, and work cost assessment data from step S100, such as GPU temperature, GPU power consumption, task list, cooling system energy consumption, and predicted power consumption. The action data comes from the optimized task scheduling scheme from step S200, and the reward data comes from the difference between the actual power cost data and the predicted power consumption. After collecting the state data, action data, and reward data, the state data needs to be augmented to enhance the robustness of the subsequent construction and training of the imitation learning network. Specifically, the existing state data is slightly perturbed. For example, Gaussian noise that conforms to its historical noise model is added to GPU temperature and GPU power consumption to generate new states that are similar but slightly different. The action data and reward data are left unchanged. This combination generates learning augmentation data.

[0086] Step S400-2: Construct an imitation learning network based on a graph neural network and define graph nodes and edges. Add an attention mechanism to the imitation learning network. The loss function of the imitation learning network includes a main loss function and an auxiliary loss function. Train the imitation learning network based on learning reinforcement data.

[0087] Specifically, graph neural networks can capture the topology of multi-GPU computing clusters more clearly and understand common problems such as local heat accumulation. Using graph neural networks as the basic architecture of imitation learning networks, the entire multi-GPU computing cluster is naturally modeled as a graph, where the computing units are graph nodes, and node features include GPU power consumption, GPU temperature, etc. Edges can represent physical connections between GPUs, such as through the same switch or thermal impact relationships. An attention mechanism is added to the imitation learning network, that is, self-attention or cross-attention layers are introduced into the imitation learning network, so that the network can dynamically focus on the key GPUs or tasks in the current state, such as the GPU with the highest temperature, the task with the largest amount of computation, etc., thereby making better decisions. As for the input and output layers of the imitation learning network, the input layer is used to input state data, while the output layer is designed with a multi-head structure to output different parts of the solution.

[0088] Furthermore, for the loss function of imitation learning networks, the most direct method is behavior cloning, which can use supervised learning loss. The main loss function is usually a weighted sum, for example, total loss = x * classification loss + y * regression loss. The classification loss can be such as cross-entropy, used for GPU task allocation, while the mean squared error can be used for task priority. An auxiliary loss function is added on the basis of the main loss function, thereby introducing reward prediction as an auxiliary task, so that the imitation learning network can not only imitate actions, but also try to predict the reward value of the action, which helps the imitation learning network understand the quality of the action, such as the total energy consumption of the overall scheme and the predicted temperature, thus improving the generalization ability. The loss function of the imitation learning network is obtained by a hybrid combination of the auxiliary loss function and the main loss function to handle the fitting of discrete and continuous parts in the scheduling scheme.

[0089] Furthermore, after defining the architecture and loss function of the imitation learning network, it can be trained. The training dataset is the augmented learning data, which is divided into training, validation, and test sets. The imitation learning network is trained multiple times using the training set. In each training round, the imitation learning network receives a batch of state data, action data, and reward data, and obtains the predicted loss. Then, the weights of the imitation learning network are updated through backpropagation. After each training cycle, the performance of the imitation learning network is evaluated using the validation set, and the loss function is monitored. Training ends when a preset condition is met, such as an early stopping mechanism, which stops training after the maximum number of iterations is reached, thus deriving the trained imitation learning network.

[0090] Step S400-3: Deploy the trained imitation learning network and simultaneously set a selection strategy. Based on the selection strategy, select to use the imitation learning network or steps S100 to S200 to obtain an optimized task scheduling scheme, and then execute step S300.

[0091] Specifically, the trained imitation learning network is deployed to the central processing unit (CPU). The CPU continuously receives status data collected by the cluster edge units and selects either the slow or fast channel based on the selection strategy.

[0092] The selection strategy includes a slow channel and a fast channel. The slow channel refers to obtaining an optimized task scheduling scheme using steps S100 to S200, and the fast channel refers to obtaining an optimized task scheduling scheme using an imitation learning network.

[0093] Understandably, in most cases, such as 90% directly entering the fast track and 10% entering the slow track, the imitation learning network quickly generates an optimized task scheduling scheme to make faster scheduling decisions. The inclusion of the slow track is a consideration for robustness and the accuracy of the final scheduling scheme, as high-risk or complex and unfamiliar scenarios may arise. In such scenarios, the imitation learning network may not be able to make good decisions, necessitating the entry into the slow track. The chaotic algorithm and swarm intelligence algorithm in steps S100 to S200 then make decisions on the scheduling scheme to obtain a better one. The conditions for triggering the slow track can be set to high uncertainty scenarios, rare scenarios, performance degradation scenarios, etc. When the imitation learning network outputs the scheduling scheme, it can also output its confidence level. If the confidence level is below a certain threshold, for example, if the imitation learning network hesitates about which computing unit to assign a task to, the slow track can be triggered. The fast channel makes the final decision through chaotic algorithms and swarm intelligence algorithms. For rare scenarios, the average of the current state data and all state data in all reinforcement learning data can be calculated when the imitation learning network is working. If the current state differs significantly from historical experience, it is a relatively unfamiliar rare scenario, which can be handled by the slow channel. In addition, for performance degradation scenarios, the energy consumption of the scheduling scheme generated by the imitation learning network can be continuously compared with the energy consumption of the schemes generated by the chaotic algorithm and swarm intelligence algorithm recently. If the imitation learning network is consistently lower than the chaotic algorithm and swarm intelligence algorithm, the calling frequency of steps S100 to S200 can be increased. A feedback optimization process can also be added to the imitation training network to update it. For example, the DAgger algorithm framework can be used to update and optimize the imitation training network. Since update optimization is a commonly used technique, it will not be elaborated on further here.

[0094] In this embodiment, Figure 3 The diagram illustrates a time-sharing multiplexing task scheduling system for multi-GPU computing clusters, which can realize the ideas of this application, according to some embodiments of this application.

[0095] Specifically, a time-sharing multiplexing task scheduling system for multi-GPU computing clusters includes: Cluster edge unit, which is used to acquire computing cluster data in parallel; The evaluation module is used to construct working characteristic data and generate working cost evaluation data based on computing cluster data and system status data; An optimization module is used to generate an initial task scheduling scheme set and process the initial task scheduling scheme set based on swarm intelligence algorithm and chaos algorithm to generate an optimized task scheduling scheme. The central processing unit is used to allocate tasks based on an optimized task scheduling scheme and to obtain actual power consumption cost data and operation reports; The imitation module is used to construct an imitation learning network, which is used to imitate and generate optimized task scheduling schemes.

[0096] The specific usage and function of this embodiment are explained below: First, data from the computing cluster is collected through the cluster edge unit. This data, combined with system status data and a short-term prediction model, generates workload assessment data, providing forward-looking data for subsequent scheduling decisions. Next, an initial task scheduling scheme set is generated based on a chaotic algorithm. This initial scheme set is then processed using a swarm intelligence algorithm, and an optimized task scheduling scheme is generated using the chaotic algorithm. A hybrid swarm intelligence algorithm, incorporating both chaotic and efficient methods, is used for scheduling search. During the search, a strategy evaluation is performed, considering both energy consumption and cooling control objectives to ensure the optimized task scheduling scheme achieves the dual goals of energy consumption and cooling. The optimized task scheduling scheme is then deployed for execution. Finally, an imitation learning network is constructed to distill the scheduling decisions from the swarm intelligence and chaotic algorithms. This network enables rapid decision-making, accelerating computation and reducing computational load while ensuring real-time performance. This method provides a task scheduling approach that integrates energy consumption and cooling awareness to achieve optimal energy consumption while maintaining cooling.

[0097] In this embodiment, an electronic device is also provided, including: At least one processor; and at least one memory communicatively connected to the processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform the method proposed in Embodiment 1 of the present invention.

[0098] The following is a detailed introduction to the various components of the electronic device: In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement Embodiment 1 of this invention, such as one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs).

[0099] The processor can perform various functions of an electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0100] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0101] The memory can be a real-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; this embodiment of the invention does not specifically limit this.

[0102] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via limited means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0103] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0104] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0105] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A time-sharing multiplexing task scheduling method for multi-GPU computing clusters, characterized in that, It includes the following steps: S100 deploys cluster edge units, acquires computing cluster data in parallel based on cluster edge units, constructs working characteristic data based on computing cluster data and system status data, and generates working cost evaluation data based on cost-utility function. S200, perform two-stage coding and generate an initial task scheduling scheme set based on the chaos algorithm, process the initial task scheduling scheme set according to the swarm intelligence algorithm, and generate an optimized task scheduling scheme in combination with the chaos algorithm. The S300 multi-GPU computing cluster operates based on an optimized task scheduling scheme and acquires actual power consumption cost data and operation reports. S400: During the execution of steps S100 to S300, learning enhancement data is collected, and an imitation learning network is constructed based on the learning enhancement data. The imitation learning network operates based on an imitation strategy and is used to imitate and generate an optimized task scheduling scheme.

2. The time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 1, characterized in that, Deploy cluster edge units, acquire computing cluster data in parallel based on the cluster edge units, and construct operational characteristic data based on the computing cluster data and system status data, and generate operational cost assessment data based on the cost-utility function, including: The cluster edge unit is deployed for each computing unit in the computing cluster. Based on the cluster edge unit, the performance data and unit task information data of each computing unit are obtained, and the performance data and unit task information data are encapsulated into computing cluster data. Acquire cooling status data and cluster task information data, and encapsulate the cooling status data and cluster task information data into system status data; Data processing is performed on the computing cluster data and the system status data to obtain structured cluster data. Based on the structured cluster data, features are constructed to obtain working feature data. A short-term operating prediction model is configured for each of the cluster edge units. The operating characteristic data is input into the short-term operating prediction model to obtain mixed prediction data, which includes predicted power consumption data, predicted cooling temperature data, and predicted time data. A cost-utility function is constructed, and work cost assessment data is obtained based on the hybrid prediction data.

3. The time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 2, characterized in that, Data processing is performed on the computing cluster data and the system status data to obtain structured cluster data. Based on the structured cluster data, feature construction is performed to obtain operational feature data, including: The computing cluster data and the system status data are validated and cleaned for data validity, and the computing cluster data is aggregated to obtain structured cluster data. At the same time, the structured cluster data is standardized. Dynamic load characteristic data is obtained from computing cluster data in structured cluster data; task load characteristic data is obtained from computing cluster data and system status data in structured cluster data; and cross-characteristic data is obtained from computing cluster data in structured cluster data. Work characteristic data is constructed based on dynamic load characteristic data, task load characteristic data, and cross characteristic data.

4. The time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 1, characterized in that, Two-stage coding is performed, and an initial task scheduling scheme set is generated based on a chaotic algorithm. The initial task scheduling scheme set is then processed using a swarm intelligence algorithm, and an optimized task scheduling scheme is generated by combining the chaotic algorithm, including: Two-stage encoding is performed on the unit task information data in the computing cluster data and the cluster task information data in the system status data to obtain the initial task scheduling scheme definition and solution space definition. The initial task scheduling scheme definition and solution space definition are imported into the chaotic algorithm. The initial task scheduling scheme set is generated iteratively based on the chaotic algorithm, and the diversity of the initial task scheduling scheme set is verified. The initial task scheduling scheme set is imported into the swarm intelligence algorithm as the initial population. At the same time, the swarm intelligence algorithm calls the cost-utility function and the short-term job prediction model of the cluster edge unit to obtain the fitness of each individual in the initial population. The swarm intelligence algorithm updates the population based on fitness and monitors chaos simultaneously. Based on the chaos monitoring, it triggers a chaos perturbation strategy, which includes chaotic individual perturbation and chaotic swarm perturbation. Obtain an optimized task scheduling solution.

5. A time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 4, characterized in that, The initial task scheduling scheme definition and solution space definition are imported into the chaotic algorithm. An initial task scheduling scheme set is iteratively generated based on the chaotic algorithm, and the diversity of the initial task scheduling scheme set is verified, including: Select the Logistic mapping and set the initial values ​​for chaos; Run the Logistic mapping and iteratively generate a chaotic random sequence based on the Logistic mapping. Map the chaotic random sequence to the encoding space to obtain an initial task scheduling scheme set. Obtain the distribution metric and average diversity of the initial task scheduling scheme set, and compare the distribution metric and average diversity with preset conditions.

6. The time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 4, characterized in that, The swarm intelligence algorithm updates the population based on fitness and simultaneously monitors chaos. Based on this chaos monitoring, a chaos perturbation strategy is triggered. This strategy includes both individual chaos perturbation and swarm chaos perturbation, including: Define stagnation criteria and monitoring criteria, wherein the stagnation criteria include, but are not limited to, population aggregation criteria, maximum frequency criteria, and diversity criteria; If the stagnation criterion is met, updates will continue; if the stagnation criterion is not met, the chaotic perturbation strategy will be triggered. Apply chaotic individual perturbation to the historical best solution in the currently updated population, and apply chaotic group perturbation to the individuals in the currently updated population.

7. The time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 1, characterized in that, The multi-GPU computing cluster operates based on an optimized task scheduling scheme and acquires actual power consumption cost data and operation reports, including: The feasibility of the optimized task scheduling scheme is verified to obtain the verified scheduling scheme; The verification scheduling scheme is distributed to the corresponding computing units, and resource quotas are allocated to the computing units based on the cluster edge units; During task execution, data is collected based on the cluster edge units to obtain actual power consumption cost data, and an operation report is generated based on the cluster edge units.

8. The time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 1, characterized in that, During steps S100 to S300, learning enhancement data is collected, and an imitation learning network is constructed based on the learning enhancement data. The imitation learning network operates based on an imitation strategy and is used to imitate and generate optimized task scheduling schemes, including: Collect state data, action data, and reward data; perform data augmentation on the state data; and obtain learning augmentation data based on the augmented state data, action data, and reward data. An imitation learning network is constructed based on a graph neural network, and graph nodes and edges are defined. An attention mechanism is added to the imitation learning network. The loss function of the imitation learning network includes a main loss function and an auxiliary loss function. The imitation learning network is trained based on learning reinforcement data. Deploy the trained imitation learning network and simultaneously set a selection strategy. Based on the selection strategy, select to use the imitation learning network or steps S100 to S200 to obtain an optimized task scheduling scheme, and then execute step S300.

9. A time-sharing multiplexing task scheduling method for multi-GPU computing clusters according to claim 8, characterized in that, Deploy the trained imitation learning network and simultaneously set a selection strategy. Based on the selection strategy, select to use the imitation learning network or steps S100 to S200 to obtain an optimized task scheduling scheme. Execute step S300, including: The selection strategy includes a slow channel and a fast channel. The slow channel refers to obtaining an optimized task scheduling scheme using steps S100 to S200, and the fast channel refers to obtaining an optimized task scheduling scheme using an imitation learning network.

10. A time-sharing multiplexing task scheduling system for multi-GPU computing clusters, used to implement the method described in any one of claims 1 to 9, characterized in that, include: Cluster edge unit, which is used to acquire computing cluster data in parallel; The evaluation module is used to construct working characteristic data and generate working cost evaluation data based on computing cluster data and system status data; An optimization module is used to generate an initial task scheduling scheme set and process the initial task scheduling scheme set based on swarm intelligence algorithm and chaos algorithm to generate an optimized task scheduling scheme. The central processing unit is used to allocate tasks based on an optimized task scheduling scheme and to obtain actual power consumption cost data and operation reports; The imitation module is used to construct an imitation learning network, which is used to imitate and generate optimized task scheduling schemes.

Images