Hybrid heterogeneous computing power-oriented computing power and power collaboration task scheduling system and hybrid heterogeneous computing power-oriented computing power and power collaboration task scheduling method

Abstract

The invention discloses a mixed heterogeneous computing power-oriented computing power collaborative task scheduling method and system, relates to the technical field of computing power collaboration, is deployed in an intelligent computing center management server, and comprises a data acquisition module, an energy consumption characteristic modeling module, a task schedulability analysis module, a collaborative scheduling decision module and a scheduling instruction execution module. The method comprises the steps of collecting task, computing power and electric power data through a multi-source interface, constructing a refined energy consumption prediction model, quantifying a task schedulable potential generation vector, dynamically selecting a DRL or GA algorithm based on a task scene, solving a constraint optimization problem with the minimum total operation cost, outputting an optimal scheduling strategy, translating the optimal scheduling strategy into an underlying platform instruction for execution, and forming closed-loop optimization. The method breaks through IT and OT system barriers, realizes cooperation of computing power and electric power, reduces operation cost, improves heterogeneous computing power energy efficiency and new energy consumption rate, enhances scheduling adaptability and robustness, and is suitable for low-carbon intelligent operation of an intelligent computing center.

Description

Technical Field

[0001] This invention relates to the field of computing-powered collaborative technology, and in particular to a computing-powered collaborative task scheduling system and method for hybrid heterogeneous computing power. Background Technology

[0002] Driven by both the national strategy of "Eastern Data, Western Computing" and the technological wave of large-scale artificial intelligence models, intelligent computing centers have experienced explosive growth. However, behind this rapid development, the operation of intelligent computing centers faces four core challenges:

[0003] First, the high energy consumption and high cost are prominent issues. Computationally intensive tasks such as training large AI models have led to a significant increase in the energy consumption of intelligent computing centers, with electricity costs accounting for a very high proportion of total operating expenditures (OPEX). At the same time, the power usage efficiency (PUE) of intelligent computing centers remains high, and the resulting carbon emission pressure has become the primary bottleneck restricting their sustainable operation.

[0004] Second, the integration of new energy sources is difficult. Intelligent computing centers are often encouraged to integrate green energy sources such as wind power and photovoltaics in their vicinity. However, the output of new energy sources is highly volatile and intermittent. For example, photovoltaics only generate power during the day, while the computing load of intelligent computing centers (especially real-time inference tasks) is often present around the clock. This results in a serious mismatch between the "peak of green electricity supply" and the "peak of computing demand," making it difficult to efficiently utilize low-cost green energy.

[0005] Third, heterogeneous computing power has energy efficiency blind spots. Modern intelligent computing centers (especially domestically produced ones) generally adopt a hybrid heterogeneous computing power architecture, covering GPUs, NPUs, CPUs and other hardware from different manufacturers and generations. The performance-to-power ratio of these hardware components when performing the same AI task varies significantly. For example, some NPUs perform well in inference tasks but poorly in training tasks. Existing scheduling methods cannot fully leverage the energy efficiency advantages of various hardware components.

[0006] Fourth, there is a contradiction between the diversity of AI tasks and the rigidity of scheduling. The AI ​​tasks undertaken by the intelligent computing center are diverse, and the business requirements are significantly different: large model training tasks have high power consumption and long processing time, but are not sensitive to latency and support the "flexible" characteristics of interruption and time shift; while real-time inference tasks have low power consumption and short processing time, but have extremely high latency requirements and are "rigid" tasks that cannot be interrupted. The existing scheduling system is difficult to accurately adapt to these differentiated requirements.

[0007] Existing technologies have significant limitations, with the core issue being the fundamental flaw of a "separation between IT and OT" in intelligent computing center scheduling systems. From the perspective of IT schedulers, the core goals of mainstream systems such as Kubernetes and Slurm are to improve resource utilization (e.g., increase GPU memory usage) or ensure task fairness (e.g., by using a fair-share algorithm). They can identify task requirements and hardware resources, but they are completely "blind" to electricity-related information. They are unaware of real-time electricity price fluctuations and cannot judge the energy consumption differences of the same task on different hardware, thus failing to achieve optimizations aimed at reducing costs or carbon emissions.

[0008] From the perspective of OT management systems, power management systems such as BMS and EMS primarily focus on ensuring power supply security or achieving coarse-grained peak shaving. They can sense total power consumption and electricity price changes, but they are completely unaware of the characteristics of AI tasks: when electricity prices are at their peak, these systems can at most issue a coarse instruction like "reduce total power consumption by 10%", unable to distinguish between interruptible flexible tasks and uninterruptible rigid tasks. This "one-size-fits-all" control approach is highly susceptible to mistakenly interrupting high-priority inference tasks, triggering a cascading drop in Quality of Service (QoS).

[0009] In summary, IT systems understand tasks but not electricity, while OT systems understand electricity but not tasks, creating a significant "semantic gap." Therefore, there is an urgent need for an integrated collaborative optimization method that can simultaneously consider the characteristics of AI tasks, the performance of heterogeneous hardware, and real-time power data to solve the aforementioned technical challenges and help intelligent computing centers achieve low-carbon, cost-reduced, and intelligent operation. Summary of the Invention

[0010] The purpose of this invention is to provide a computing and power collaborative task scheduling system and method for hybrid heterogeneous computing power. This method and system can identify the energy consumption characteristics of different AI tasks on hybrid heterogeneous computing power, and combine real-time electricity prices and new energy output to perform collaborative scheduling of computing power tasks and power resources, so as to achieve cost reduction, efficiency improvement and low-carbon operation of intelligent computing centers.

[0011] To achieve the above objectives, this invention provides a computing-power collaborative task scheduling system for hybrid heterogeneous computing power, deployed on the management server of an intelligent computing center. The system includes a data acquisition module, an energy consumption characteristic modeling module, a task schedulability analysis module, a collaborative scheduling decision module, and a scheduling instruction execution module. The scheduling instruction execution module is communicatively connected to the resource management platform of the intelligent computing center. Each module sequentially and collaboratively implements computing-power collaborative scheduling. The data acquisition module, as the system perception layer, aggregates heterogeneous data from multiple sources, including tasks, computing power, and electricity, in real time and periodically through preset interfaces and protocols. After preprocessing, the data is stored in a time-series database. The energy consumption characteristic modeling module connects to the data acquisition module, extracts historical and real-time data from the time series database to construct a feature-label dataset, and uses the gradient boosting decision tree algorithm to train the energy consumption prediction model. The energy consumption prediction model is periodically updated or retraining is triggered when a new task / hardware type is added. The task schedulability analysis module connects to the data acquisition module and has a built-in task attribute rule library. It queries the rule library according to task type and business requirements and dynamically generates a schedulability vector that includes time elasticity, interruptibility, and service quality adjustability. The collaborative scheduling decision module, as the core of the system, receives power information, energy consumption prediction models, tasks to be scheduled, and schedulability vectors. It has a built-in hybrid decision routing function and dynamically selects deep reinforcement learning or genetic algorithms based on the number of tasks and the task arrival rate. Under the premise of meeting the constraints, it solves the constrained optimization problem with the goal of minimizing the total operating cost and generates the optimal scheduling strategy. The scheduling instruction execution module connects the collaborative scheduling decision module and the intelligent computing center resource management platform. It translates the optimal scheduling strategy into API calls or command-line instructions that the underlying platform can recognize, executes task deployment, startup, pause and resumption operations, and provides feedback on the actual execution status of the tasks.

[0012] Preferably, the preset interface and protocol of the data acquisition module include: For the task layer: connect to the task scheduling queue and resource management platform of the intelligent computing center via API; For the computing hardware layer: obtain the physical state and power consumption of the hardware through out-of-band management interfaces and in-band monitoring tools; For the power layer: connect to the power monitoring system or smart meters of the intelligent computing center via standard industrial protocols.

[0013] Preferably, the features extracted by the energy consumption characteristic modeling module include task type, task parameters, hardware model, hardware load and construction interaction features, and the extracted tag is the real-time power consumption of the hardware. The model training process includes data partitioning, hyperparameter tuning, and model training. The training objective is to minimize the root mean square error or mean absolute percentage error of the validation set. The model is stored in a standard format or deployed as a microservice API. The formula for calculating the root mean square error is:

[0014] in, To determine the number of samples in the validation set, To predict power consumption for the model, This represents the actual power consumption for data acquisition.

[0015] Preferably, the schedulability vector of the task schedulability analysis module is represented as follows:

[0016] in: , indicating task The maximum time window during which execution can be postponed. This is the latest deadline for completing the task. For task submission time, The estimated shortest execution time for the task; This represents a Boolean value, obtained by querying the rule base. It takes the value 1 or 0, where 1 indicates interruptible and 0 indicates non-interruptible. Task type; , represents a floating-point value, obtained by querying the rule base, with a value range of 0 to 1, where 1 indicates fully adjustable and 0 indicates strict service quality.

[0017] Preferably, the hybrid decision-making routing logic of the collaborative scheduling decision module is implemented through a routing controller, specifically as follows: If the number of tasks to be scheduled And the task completion rate The scenario is determined to be a static batch task, and a genetic algorithm is selected. The genetic algorithm uses chromosome encoding to achieve global optimization search, and its core is used in static batch task scenarios. The implementation process is as follows: (1) Chromosome encoding: Each chromosome corresponds to a complete scheduling scheme. The length of the chromosome is the same as the number of tasks to be scheduled, N. Each gene is a tuple. , k For hardware indexing, For the task Startup time; decode chromosome π to obtain x for all tasks. i,k,t (π) and T finish,i (π); (2) Constraint check: Check whether scheme π violates resource capacity constraints and delay / interruption constraints. If the constraints are violated, then let C total (π)=∞; if no constraints are violated, then calculate C according to the objective function. total (π) = C elec (π) +C penalty (π); The selection of genetic operators during the calculation process is as follows: The selection operator employs either Tournament Selection or Roulette Wheel Selection to screen for highly fit parent individuals; the crossover operator uses single-point crossover or uniform crossover, where parent individuals exchange some task hardware allocation and start-up time information to generate offspring; the mutation operator mutates the offspring, for example, by randomly selecting tasks, changing their hardware allocation, or modifying the schedulability vector. Time flexibility Allowed window offset of its startup time ; (3) Calculate fitness, fitness function ,in To prevent positive numbers from being divided by zero; (4) Termination condition: The algorithm terminates after the preset number of iterations G or the fitness function value converges, and the scheduling strategy corresponding to the chromosome with the highest fitness is output as the optimal strategy. ; If the number of tasks to be scheduled or task arrival rate The scenario was identified as a large-scale dynamic task flow scenario, and a deep reinforcement learning algorithm was selected. The deep reinforcement learning algorithm uses Markov decision process modeling and is primarily used for rapid decision-making in large-scale dynamic task flow scenarios, specifically including: (1) State modeling, state s t It consists of an electricity price vector for the next K time steps, a cluster hardware state matrix, and a task queue state matrix. The electricity price vector contains the predicted electricity price for each time step. Each row of the cluster hardware state matrix contains the real-time hardware load, hardware model, and idle / busy status. Each row of the task queue state matrix contains the task type and schedulability vector. Task submission time ; (2) Feature extraction: 1D CNN or LSTM is used to extract the temporal features of the electricity price vector and the task queue. MLP is used to process the cluster hardware state matrix. The features are concatenated and used as the input of the deep reinforcement learning agent. (3) Definition of action: Action For composite discrete actions, it means that the task will be performed at time t. Assigned to hardware The action space size is the product of the number of tasks to be scheduled N and the total number of hardware M, N×M; (4) Algorithm selection: The agent is trained using either the PPO or A2C algorithm; (5) Reward Design: The reward is represented as follows:

[0018] in, It is the instantaneous power cost consumed by the cluster during time t. The timeout penalty cost for a task starting at time t; (6) Training process: The agent continuously interacts with the environment in the simulation environment or the digital twin environment of the real system, and learns the optimal scheduling strategy by maximizing the cumulative reward. ; in This is a threshold for the number of tasks. This represents the total number of tasks in the queue to be scheduled. This represents the number of new tasks added per unit of time.

[0019] Preferably, the solution process for the collaborative scheduling decision module is as follows: The decision variable is defined as follows: π is a scheduling strategy. Let i be a binary decision variable, where i is the task index and k is the hardware index. If the task... Under policy π, at time t, the hardware is... If it runs on, then Otherwise, it is 0; The objective function of the total operating cost minimization optimization problem is:

[0020]

[0021]

[0022] in, Let T be the electricity cost, M be the total duration of the scheduling cycle, M be the total number of hardware units in the heterogeneous computing resource pool, and N be the total number of tasks to be scheduled. For the task In hardware Expected power consumption at time t. Let be the electricity price at time t. For time step; As a penalty cost for delay, For the task i The penalty weight, For the task i Actual completion time under strategy π.

[0023] Preferably, the constraints in the solution process of the collaborative scheduling decision module include: Task completion constraints: Ensure that the task execution time is no less than the estimated minimum execution time; Resource capacity constraints: ; For the task For hardware Resource consumption For hardware Total resource capacity; Delay constraints: The stringency of the constraints is determined by the time elasticity in the schedulability vector. Decide, This is a flexible constraint; Interruptible constraints: ,but exist arrive The values ​​must be consecutively 1, and no breaks are allowed.

[0024] Preferably, the scheduling strategy translation logic of the scheduling instruction execution module includes: For the Kubernetes platform: Translate tasks into Pod / Job YAML files, defining container images, startup commands, and parameters; translate hardware into node affinity or node selectors; translate time into immediate creation or delayed queues; translate pause for interruptible tasks into adjusting the number of Job replicas or deleting Pods; translate resume into restoring the number of replicas or recreating the Pod and resuming from the checkpoint. For the Slurm platform: tasks are translated into .sbatch batch scripts, including execution commands and parameters; hardware is translated into #SBATCH hardware constraint instructions or partition specification instructions; time is translated into immediate execution or future execution; and pause of interruptible tasks is translated into scontrol hold.<job_id> The correct translation is scontrol release.<job_id> .

[0025] A computing-electricity collaborative task scheduling method for hybrid heterogeneous computing power includes the following steps: Step 1: Real-time perception of multi-dimensional status. Task information, computing power status and power information are acquired through the data acquisition module, and stored in the time series database after timestamp alignment and cleaning preprocessing. Step 2: Build and update the energy consumption characteristic model. Extract time series database data through the energy consumption characteristic modeling module, complete feature engineering, model training and model maintenance, and output a high-precision energy consumption prediction model. Step 3: Quantify the schedulable potential of tasks. Perform rule matching and vector calculation on each task to be scheduled through the task schedulability analysis module to generate a schedulability vector. Step 4: Solve the computational-electricity collaborative scheduling optimization problem, including algorithm routing, algorithm selection, and execution. During algorithm selection, a hybrid decision-making routing function is implemented through a routing controller, and either a deep reinforcement learning algorithm or a genetic algorithm is selected as the solution algorithm. During execution, the energy consumption prediction model is called to obtain the expected power consumption, and the electricity cost is calculated in combination with the electricity price. Under the constraints of task completion, resource capacity, latency, and interruptibility, the selected solution algorithm is used to solve the optimization problem that minimizes the total operating cost and generate the optimal scheduling strategy. Step 5: Issue and execute scheduling instructions. The scheduling instruction execution module translates the optimal scheduling strategy into underlying platform instructions and executes them. Step 6: Status feedback and closed loop. The scheduling instruction execution module monitors the actual execution status of the task and feeds it back to the data acquisition module, providing a basis for subsequent data perception and model updates.

[0026] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: 1. Reduce the operating costs of the intelligent computing center: By accurately matching tasks and heterogeneous hardware through a refined energy consumption model, and combining real-time electricity prices and renewable energy output, the task execution time can be rationally planned in the time dimension, and the allocation of hardware resources can be optimized in the spatial dimension, thereby reducing ineffective energy consumption and electricity expenditure and achieving effective reduction in operating costs.

[0027] 2. Improve the energy efficiency of hybrid heterogeneous computing power: Deeply adapt to the energy consumption characteristics of hybrid heterogeneous computing power (especially domestic computing power) to achieve optimal matching between tasks and hardware, avoid energy waste caused by resource mismatch, and improve the overall energy utilization efficiency and return on investment of computing power clusters.

[0028] 3. Enhance the capacity for renewable energy consumption: By quantifying the flexibility of AI task load, the computing power load of the intelligent computing center can be flexibly adjusted, and the task execution volume can be proactively increased during peak renewable energy output periods, thereby promoting the efficient consumption of green energy and contributing to the achievement of dual-carbon goals.

[0029] 4. Achieve synergistic optimization of computing power and power: Break down the barriers between IT scheduling systems and OT power systems, and establish linkage links between AI tasks, heterogeneous hardware and power systems at the data layer, model layer and decision layer, so as to realize the transformation from passive response to proactive collaboration and provide technical support for the low-carbon and intelligent operation of intelligent computing centers.

[0030] 5. Enhance scheduling adaptability and robustness: Through the hybrid decision-making routing function, the optimal solution algorithm is dynamically selected according to the task scenario, taking into account both global optimization capability and fast real-time response capability, so that the system can maintain efficient and stable operation under complex and ever-changing workloads, and improve the universality of scheduling and decision quality.

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the structure of a computing-powered collaborative task scheduling system for hybrid heterogeneous computing power according to Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating Embodiment 2 of the task scheduling system for a computing power system oriented towards hybrid heterogeneous computing power according to the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1 like Figure 1 As shown, a computing-power collaborative task scheduling system for hybrid heterogeneous computing power is deployed on the management server of an intelligent computing center. It includes a data acquisition module, an energy consumption characteristic modeling module, a task schedulability analysis module, a collaborative scheduling decision module, and a scheduling instruction execution module. The scheduling instruction execution module is communicatively connected to the resource management platform of the intelligent computing center. Each module works in sequence to collaboratively implement computing-power collaborative scheduling. I. Data Acquisition Module Configuration As the system perception layer, it aggregates heterogeneous data from multiple sources, including tasks, computing power, and electricity, in real time and periodically through preset interfaces and protocols, providing a unified data foundation for subsequent energy consumption modeling and scheduling decisions.

[0037] 1. Preset interfaces and protocols include: Task layer: Connects to the task scheduling queue (such as RabbitMQ, Kafka) and resource management platform (such as Kubernetes, Slurm) of the intelligent computing center via API (such as REST API).

[0038] Computing hardware layer: Obtain the physical status and power consumption of the hardware through out-of-band management interfaces (such as IPMI - Intelligent Platform Management Interface) and in-band monitoring tools (such as NVIDIA's DCGM - Data Center GPU Manager, or Huawei's CANN NPU monitoring tool).

[0039] Power layer: Connects to the power monitoring system (SCADA / EMS) or smart meters of the intelligent computing center via standard industrial protocols (such as Modbus, OPC-UA).

[0040] 2. Data collection: Intelligent computing task information includes task identifier, task type (large model training, model fine-tuning, batch inference, real-time inference), task parameters (model name, dataset size, batch size, computational precision), and business requirements (task submission time, latest completion deadline, priority).

[0041] Heterogeneous computing power resource pool status: includes hardware identifier, hardware model (GPU A100, NPU 910B, CPU X86, etc.), static resources (total video memory, total FLOPS) and dynamic status (real-time load, real-time power consumption, core temperature, idle / busy status).

[0042] Electricity information includes time-of-use electricity prices in real time or for the next 24 hours, forecasted output curves of new energy sources (wind / solar), and grid load factor or demand response signals.

[0043] 3. Data Processing: The collected raw data is aligned with timestamps, cleaned and preprocessed, and then stored in a time-series database (such as InfluxDB or Prometheus) for use by other modules.

[0044] II. Energy Consumption Characteristic Modeling Module Data Extraction: Historical and real-time task execution data and power consumption data are extracted from the time-series database to construct a feature-label dataset. A refined energy consumption prediction model P is then built. est This model is used to predict a specific AI task J. i In a specific heterogeneous computing unit H k Expected real-time power P during operation k,actual (t).

[0045] Feature engineering: For task type J type,i Hardware model H type,k Perform one-hot encoding; consider task parameters such as dataset size, batch size, and hardware load. k (t) Numerical features are used; optional interactive features of task type and hardware model can be constructed.

[0046] Model training: Gradient Boosting Decision Tree (GDBT, such as XGBoost or LightGBM) is preferred. GDBT models can effectively handle high-dimensional, sparse features with complex nonlinear relationships (such as the relationship between task parameters and hardware power consumption), are robust, can automatically perform feature interaction, and have high prediction accuracy. Alternative multivariate nonlinear regression models (such as multinomial regression) are suitable for scenarios with relatively simple energy consumption curves and limited computational resources. The trained model f is as follows:

[0047]

[0048] The training process is as follows: 1. Data partitioning: The historical feature-label dataset is divided into training, validation, and test sets in a 7:2:1 ratio; 2. Hyperparameter tuning: Use the validation set to tune the hyperparameters of the GDBT model, for example, by using grid search or Bayesian optimization to determine the optimal learning rate, maximum tree depth, and subsample ratio. 3. Model Training: Train model f on the training set. The goal is to minimize the root mean square error (RMSE) or mean absolute percentage error (MAPE) on the validation set to ensure the model's generalization ability. The formula for calculating the root mean square error is:

[0049]

[0050] in, To determine the number of samples in the validation set, To predict power consumption for the model, This represents the actual power consumption for data acquisition.

[0051] 4. Periodic Updates: The model is configured to be updated daily or weekly, or to trigger retraining when a new hardware type or new task type is detected; The trained model is saved in PMML or ONNX standard format, or deployed as a REST API microservice, which can be called in real time by the collaborative scheduling decision module during optimization to obtain the expected power consumption P for any task-hardware combination. est .

[0052] III. Task Scheduling Analysis Module Rule base configuration: Includes a built-in configurable task attribute rule base in JSON or XML format, predefining interruptibility and QoS adjustability benchmarks for each task type. An example rule base is shown below: [ {"J_type": "Large Model Training", "base_s_interrupt": 1, "base_s_qos": 1}, {"J_type": "Model fine-tuning", "base_s_interrupt": 1, "base_s_qos": 0.5}, {"J_type": "Batch Inference", "base_s_interrupt": 0, "base_s_qos": 0}, {"J_type": "Real-time Inference", "base_s_interrupt": 0, "base_s_qos": 0} ] The module determines the task type J based on the data acquisition module. type,i and business needs J req,i (Business requirements include T) arrival,i and T deadline,i ), query the rule base, and automatically quantify each task J to be scheduled. i The schedulable potential for each task Dynamically generate schedulability vectors that include time elasticity, interruptibility, and service quality adjustability. , represented as:

[0053] Time flexibility , indicating task The maximum time window (in seconds or minutes) that can be postponed for execution. This is the latest deadline for completing the task. For task submission time, The estimated minimum execution time for the task can be provided by the energy consumption characteristic modeling module or historical data statistics; Interruptibility This represents a Boolean value, obtained by querying the rule base. It takes the value 1 or 0, where 1 indicates interruptible and 0 indicates non-interruptible. Task type; Service quality adjustability This is a floating-point value indicating whether the task allows the scheduler to sacrifice some performance for energy saving. It is obtained by querying the rule base, and its value ranges from 0 to 1, where 1 indicates fully adjustable and 0 indicates strict quality of service.

[0054] The generated schedulability vector It is immediately sent to the collaborative scheduling decision module as a key constraint input during its optimization solution.

[0055] IV. Collaborative Scheduling Decision Module As the core of the system, it connects the data acquisition module, energy consumption characteristic modeling module, and task schedulability analysis module. It has a built-in routing controller that dynamically selects either a deep reinforcement learning algorithm or a genetic algorithm based on the number of tasks and their arrival rate to solve a constrained optimization problem with the objective of minimizing total operating cost, generating the optimal scheduling strategy. This refers to a series of decisions regarding task allocation and start time.

[0056] 1. The module's inputs are: power information E(t), energy consumption prediction model Pest(t), and the queue of tasks to be scheduled {J}. i} and its corresponding schedulability vector {S} i}

[0057] 2. The mathematical description of this optimization problem is: The decision variable is defined as follows: π is a scheduling strategy. Let i be a binary decision variable, where i is the task index and k is the hardware index. If the task... Under policy π, at time t, the hardware is... If it runs on, then Otherwise, it is 0; The objective function of the total operating cost minimization optimization problem is:

[0058]

[0059]

[0060] in, Let T be the electricity cost, M be the total duration of the scheduling cycle, M be the total number of hardware units in the heterogeneous computing resource pool, and N be the total number of tasks to be scheduled. For the task In hardware Expected power consumption at time t. Let be the electricity price at time t. For time step; As a penalty cost for delay, For the task i The penalty weight, For the task i Actual completion time under strategy π.

[0061] 3. The constraints are: Task completion constraints: Ensure that the task execution time is no less than the estimated minimum execution time; Resource capacity constraints: ; For the task For hardware Resource consumption For hardware Total resource capacity; Delay constraints: The stringency of the constraints is determined by the time elasticity in the schedulability vector. Decide, This is a flexible constraint; Interruptible constraints: ,but exist arrive The values ​​must be consecutively 1, and no breaks are allowed.

[0062] 4. The routing controller performs hybrid decision-making routing functions. Before executing a specific decision-making algorithm, it determines the current queue of tasks to be scheduled {J}. i The module leverages its real-time characteristics to automatically select the optimal solution algorithm. At the start of the decision period t, the module obtains the number of tasks to be scheduled. and task completion rate , This is a threshold for the number of tasks. This represents the total number of tasks in the queue to be scheduled. This represents the number of new tasks added per unit of time.

[0063] If the number of tasks to be scheduled And the task completion rate This scenario is identified as a static batch task scenario, and a genetic algorithm is chosen. In such scenarios, the number of tasks N is relatively small, and no new tasks are being added. Genetic algorithms excel at performing global searches in these (NP-hard) combinatorial optimization problems to find near-optimal solutions. Deep reinforcement learning algorithms employ Markov decision processes for modeling, and are primarily used for rapid decision-making in large-scale dynamic task flow scenarios. Specifically, they include:

[0064] (1) State modeling, state s t It consists of an electricity price vector for the next K time steps, a cluster hardware state matrix, and a task queue state matrix. The electricity price vector contains the predicted electricity price for each time step. Each row of the cluster hardware state matrix contains the real-time hardware load, hardware model, and idle / busy status. Each row of the task queue state matrix contains the task type and schedulability vector. Task submission time ; (2) Feature extraction: 1D CNN or LSTM is used to extract the temporal features of the electricity price vector and the task queue. MLP is used to process the cluster hardware state matrix. The features are concatenated and used as the input of the deep reinforcement learning agent. (3) Definition of action: Action For composite discrete actions, it means that the task will be performed at time t. Assigned to hardware The action space size is the product of the number of tasks to be scheduled N and the total number of hardware M, N×M; (4) Algorithm selection: The agent is trained using either the PPO or A2C algorithm; (5) Reward Design: The reward is represented as follows:

[0065] in, It is the instantaneous power cost consumed by the cluster during time t. The timeout penalty cost for a task starting at time t; (6) Training process: The agent continuously interacts with the environment in the simulation environment or the digital twin environment of the real system, and learns the optimal scheduling strategy by maximizing the cumulative reward. .

[0066] If the number of tasks to be scheduled or task arrival rate This scenario is identified as a large-scale dynamic task flow scenario, and a deep reinforcement learning (DRL) algorithm is selected. In such scenarios, the number of tasks is large or tasks arrive consecutively. After pre-training, the DRL agent can provide faster real-time decision responses (ms-level) and adapt to dynamically changing environments. The genetic algorithm uses chromosome encoding to achieve global optimization search, and its core is used in static batch task scenarios, specifically including:

[0067] (1) Chromosome encoding: Each chromosome corresponds to a complete scheduling scheme. The length of the chromosome is the same as the number of tasks to be scheduled, N. Each gene is a tuple. , k For hardware indexing, For the task Startup time; decode chromosome π to obtain x for all tasks. i,k,t (π) and T finish,i (π); (2) Constraint check: Check whether scheme π violates resource capacity constraints and delay / interruption constraints. If the constraints are violated, then let C total (π) = ∞; if no constraints are violated, then calculate C according to the objective function. total (π) = C elec (π)+C penalty (π); The selection of genetic operators during the calculation process is as follows: The selection operator employs either Tournament Selection or Roulette Wheel Selection to screen for highly fit parent individuals; the crossover operator uses single-point crossover or uniform crossover, where parent individuals exchange some task hardware allocation and start-up time information to generate offspring; the mutation operator mutates the offspring, for example, by randomly selecting tasks, changing their hardware allocation, or modifying the schedulability vector. Time flexibility Allowed window offset of its startup time ; (3) Calculate fitness, fitness function , where ϵ is a positive number to prevent division by zero; (4) Termination Condition: After determining to use the GA algorithm, the system enters an iterative loop. In each iteration, the series of operations "calculate fitness → selection → crossover → mutation" are executed sequentially until the algorithm iterates to the preset generation G or the fitness function value converges and terminates. The scheduling strategy corresponding to the chromosome with the highest fitness is output as the optimal strategy. ;

[0068] V. Scheduling Instruction Execution Module Connecting the collaborative scheduling decision module and the intelligent computing center resource management platform will implement the optimal scheduling strategy. Translate the data into API calls or command-line instructions that the underlying platform can recognize, perform task deployment, startup, pause, and resumption operations, and provide feedback on the actual execution status of the task.

[0069] 1. For the Kubernetes platform: (1) Task J i Translation: The module is based on J i Information (such as J) type,i and J params,i It loads or dynamically generates a YAML description file for a Kubernetes Job or Pod from a pre-defined template library. This YAML defines the required container image, startup command, and arguments.

[0070] (2) Hardware H k Translation: The module will H k This is translated as node affinity or node selector in Kubernetes.

[0071] (3) Time T start,i translate: If T start,i "Execute immediately" means the module immediately executes the corresponding command through the K8sAPI Client.

[0072] If T start,i At some future time, the module will place the YAML task into an internal delay queue (DelayQueue) and set a timer to execute the task at time T. start,i Submissions are triggered automatically at any time.

[0073] (4) Pause / Resume (for Si.s) interrupt =1 task): Pause: The decision module issues a pause command.

[0074] Recovery: The decision module issues a recovery command.

[0075] 2. For the Slurm platform (1) Task J i Translation: The module dynamically generates a Slurm batch script (.sbatch file). The script content includes J i The execution command.

[0076] (2) Hardware H k Translation: The module adds hardware constraints in the #SBATCH directive of the .sbatch script to ensure that tasks are assigned to the correct heterogeneous computing partitions.

[0077] (3) Time T start,i translate: If T start,i It executes immediately; the module immediately calls sbatch.<script_name> .sbatch.

[0078] If T start,i This refers to a future time; the module will append a time parameter when it is called: batch--begin=<ISO_time_format><script_name> .sbatch. Slurm's scheduler will automatically put it into a PENDING state until the specified time before it starts execution.

[0079] (4) Pause / Resume (for Si.s) interrupt = 1 task): Pause: The decision-making module issues a pause command, which this module translates as scontrol hold.<job_id> .

[0080] Recovery: The decision module issues a recovery command, which this module translates as scontrol release.<job_id> .

[0081] 3. Status Feedback (Optional): After issuing a command, the module can subscribe to Kubernetes events or periodically poll Slurm's sacct or squeue commands to obtain the actual task start time T. actual_start and actual completion time T actual_finish This information is then fed back to the data acquisition module to form a closed-loop control and optimization model f.

[0082] Example 2 Using the system described in Example 1, a computing-electricity collaborative task scheduling method for hybrid heterogeneous computing power is executed, such as... Figure 2 As shown, it includes the following steps: Step 1: Real-time multi-dimensional state perception. Through its data acquisition module, it acquires multi-source heterogeneous data in real time and periodically from the task scheduling queue, heterogeneous computing hardware (GPU, NPU, etc.), and power monitoring system (SCADA / EMS) within the intelligent computing center via preset APIs (such as REST API), out-of-band management interfaces (such as IPMI), and industrial protocols (such as Modbus). After timestamp alignment and pre-processing, the data is stored in a time-series database.

[0083] Content retrieved: Task Information {J i}: Includes task type J type,i Task parameter J params,i and business needs J req,i ; Computing power status {H k (t)}: Including hardware model H type,k Real-time load L k (t) and real-time power consumption P k,actual (t); Electricity information E(t): includes real-time (or forecast) electricity price Price(t) and forecasted output of renewable energy sources. .

[0084] Step 2: Build and update the energy consumption characteristic model. Extract time series database data through the energy consumption characteristic modeling module, complete feature engineering, model training and model maintenance, and output a high-precision energy consumption prediction model. Step 3: Quantify the schedulable potential of tasks. The task schedulability analysis module performs rule matching and vector calculation on each task to be scheduled, generating a schedulability vector S. i =[s time ,s interrupt ,s qos ]; Step 4: Solve the computational-power-coordinated scheduling optimization problem, including algorithm routing, algorithm selection, and execution. Invoke the energy consumption prediction model to obtain the expected power consumption, calculate the electricity cost based on the electricity price, and solve the optimization problem that minimizes the total operating cost while satisfying constraints such as task completion, resource capacity, latency, and interruptibility. Generate the optimal scheduling strategy. This strategy defines which hardware each task should be assigned to. k And when to start executing T start,i Specifically, it includes: 4.1. Algorithm Routing: The system first analyzes the state of the current queue of tasks to be scheduled, especially the number of tasks N = |Queue(t)| and the task arrival rate R. arrival (t).

[0085] 4.2. Algorithm Selection: The judgment is made according to the preset rules of the hybrid decision routing function: If N <N threshold If Rarrival(t)≈0 (indicating a static batch scheduling scenario), then the Genetic Algorithm (GA) is selected as the solver for global optimization search.

[0086] Otherwise, if the scenario is classified as a large-scale or dynamically scheduled scenario, Deep Reinforcement Learning (DRL) is selected as the solver to achieve rapid real-time decision response.

[0087] 4.3. Perform the solution: The system call selects the algorithm (GA or DRL) from step 4.2, which minimizes the total operating cost C. total Let (π) be the objective function, and strictly adhere to the following constraints: (1) Task completion constraints (ensure that the task is completed).

[0088] (2) Resource capacity constraints (ensure that hardware resources are not overloaded).

[0089] (3) Delay constraints (using S) i .s time (Allows flexible tasks to be completed within a window).

[0090] (4) Interruptible constraints (using S) i .s interrupt (Uninterruptible tasks may be paused).

[0091] During the solution process, the algorithm obtains P by calling the energy consumption model f generated in step 2. est And combine the electricity price Price(t) obtained in step 1 to calculate C. elec (π).

[0092] Step 5: Issue and execute scheduling instructions. The scheduling instruction execution module translates the optimal scheduling strategy into underlying platform instructions and executes them. Step 6: Status feedback and closed loop. The scheduling instruction execution module monitors the actual execution status of the task and feeds it back to the data acquisition module, providing a basis for subsequent data perception and model updates.

[0093] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.

[0094] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A computing-powered collaborative task scheduling system for hybrid heterogeneous computing power, characterized in that: The management server deployed in the intelligent computing center includes a data acquisition module, an energy consumption characteristic modeling module, a task schedulability analysis module, a collaborative scheduling decision module, and a scheduling instruction execution module. The scheduling instruction execution module is communicatively connected to the intelligent computing center's resource management platform. Each module works in sequence to collaboratively implement computing and power scheduling. The data acquisition module, as the system perception layer, aggregates heterogeneous data from multiple sources, including tasks, computing power, and electricity, in real time and periodically through preset interfaces and protocols. After preprocessing, the data is stored in a time-series database. The energy consumption characteristic modeling module connects to the data acquisition module, extracts historical and real-time data from the time series database to construct a feature-label dataset, and uses the gradient boosting decision tree algorithm to train the energy consumption prediction model. The energy consumption prediction model is periodically updated or retraining is triggered when a new task / hardware type is added. The task schedulability analysis module connects to the data acquisition module and has a built-in task attribute rule library. It queries the rule library according to task type and business requirements and dynamically generates a schedulability vector that includes time elasticity, interruptibility, and service quality adjustability. The collaborative scheduling decision module, as the core of the system, receives power information, energy consumption prediction models, tasks to be scheduled, and schedulability vectors. It has a built-in hybrid decision routing function and dynamically selects deep reinforcement learning or genetic algorithms based on the number of tasks and the task arrival rate. Under the premise of meeting the constraints, it solves the constrained optimization problem with the goal of minimizing the total operating cost and generates the optimal scheduling strategy. The scheduling instruction execution module connects the collaborative scheduling decision module and the intelligent computing center resource management platform. It translates the optimal scheduling strategy into API calls or command-line instructions that the underlying platform can recognize, executes task deployment, startup, pause and resumption operations, and provides feedback on the actual execution status of the tasks.

2. The computing power collaborative task scheduling system for hybrid heterogeneous computing power according to claim 1, characterized in that: The preset interfaces and protocols of the data acquisition module include: For the task layer: connect to the task scheduling queue and resource management platform of the intelligent computing center via API; For the computing hardware layer: obtain the physical state and power consumption of the hardware through out-of-band management interfaces and in-band monitoring tools; For the power layer: connect to the power monitoring system or smart meters of the intelligent computing center via standard industrial protocols.

3. The computing-powered collaborative task scheduling system for hybrid heterogeneous computing power according to claim 2, characterized in that: The features extracted by the energy consumption characteristic modeling module include task type, task parameters, hardware model, hardware load and construction interaction features, and the extracted label is the real-time power consumption of the hardware. The model training process includes data partitioning, hyperparameter tuning, and model training. The training objective is to minimize the root mean square error or mean absolute percentage error of the validation set. The model is stored in a standard format or deployed as a microservice API. The formula for calculating the root mean square error is: ； in, To determine the number of samples in the validation set, To predict power consumption for the model, This represents the actual power consumption for data acquisition.

4. The computing power collaborative task scheduling system for hybrid heterogeneous computing power according to claim 3, characterized in that: The schedulability vector of the task schedulability analysis module is represented as follows: ； in: , indicating task The maximum time window during which execution can be postponed. This is the latest deadline for completing the task. For task submission time, The estimated shortest execution time for the task; This represents a Boolean value, obtained by querying the rule base. It takes the value 1 or 0, where 1 indicates interruptible and 0 indicates non-interruptible. Task type; , represents a floating-point value, obtained by querying the rule base, with a value range of 0 to 1, where 1 indicates fully adjustable and 0 indicates strict service quality.

5. A computing-powered collaborative task scheduling system for hybrid heterogeneous computing power according to claim 4, characterized in that: The hybrid decision-making routing logic of the collaborative scheduling decision module is implemented through a routing controller, specifically as follows: If the number of tasks to be scheduled And the task completion rate The scenario is determined to be a static batch task, and a genetic algorithm is selected. The genetic algorithm uses chromosome encoding to achieve global optimization search, and its core is used in static batch task scenarios. The implementation process is as follows: (1) Chromosome encoding: Each chromosome corresponds to a complete scheduling scheme. The length of the chromosome is the same as the number of tasks to be scheduled, N. Each gene is a tuple. , k For hardware indexing, For the task Startup time; decode chromosome π to obtain x for all tasks. i,k,t (π) and T finish,i (π); (2) Constraint check: Check whether scheme π violates resource capacity constraints and delay / interruption constraints. If the constraints are violated, then let C total (π)=∞; if no constraints are violated, then calculate C according to the objective function. total (π) = C elec (π) + C penalty (π); The selection of genetic operators during the calculation process is as follows: The selection operator uses tournament selection or roulette wheel selection to select high-fitness parent individuals; the crossover operator uses single-point crossover or uniform crossover, where parent individuals exchange some task hardware allocation and start time information to generate offspring; the mutation operator performs random mutation on the offspring. (3) Calculate fitness, fitness function ,in To prevent positive numbers from being divided by zero; (4) Termination condition: The algorithm terminates after the preset number of iterations G or the fitness function value converges, and the scheduling strategy corresponding to the chromosome with the highest fitness is output as the optimal strategy. ; If the number of tasks to be scheduled or task arrival rate The scenario was identified as a large-scale dynamic task flow scenario, and a deep reinforcement learning algorithm was selected. Deep reinforcement learning algorithms employ Markov decision process modeling and are primarily used for rapid decision-making in large-scale dynamic task flow scenarios, specifically including: (1) State modeling, state s t It consists of an electricity price vector for the next K time steps, a cluster hardware state matrix, and a task queue state matrix. The electricity price vector contains the predicted electricity price for each time step. Each row of the cluster hardware state matrix contains the real-time hardware load, hardware model, and idle / busy status. Each row of the task queue state matrix contains the task type and schedulability vector. Task submission time ; (2) Feature extraction: 1D CNN or LSTM is used to extract the temporal features of the electricity price vector and the task queue. MLP is used to process the cluster hardware state matrix. The features are concatenated and used as the input of the deep reinforcement learning agent. (3) Definition of action: Action For composite discrete actions, it means that the task will be performed at time t. Assigned to hardware The action space size is the product of the number of tasks to be scheduled N and the total number of hardware M, N×M; (4) Algorithm selection: The agent is trained using either the PPO or A2C algorithm; (5) Reward Design: The reward is represented as follows: ； in, It is the instantaneous power cost consumed by the cluster during time t. The timeout penalty cost for a task starting at time t; (6) Training process: The agent continuously interacts with the environment in the simulation environment or the digital twin environment of the real system, and learns the optimal scheduling strategy by maximizing the cumulative reward. ; in This is a threshold for the number of tasks. This represents the total number of tasks in the queue to be scheduled. This represents the number of new tasks added per unit of time.

6. The computing-electricity collaborative task scheduling method for hybrid heterogeneous computing power according to claim 5, characterized in that: The solution process for the collaborative scheduling decision module is as follows: The decision variable is defined as follows: π is a scheduling strategy. Let i be a binary decision variable, where i is the task index and k is the hardware index. If the task... Under policy π, at time t, the hardware is... If it runs on, then Otherwise, it is 0; The objective function of the total operating cost minimization optimization problem is: ； ； ； in, Let T be the electricity cost, M be the total duration of the scheduling cycle, M be the total number of hardware units in the heterogeneous computing resource pool, and N be the total number of tasks to be scheduled. For the task In hardware Expected power consumption at time t. Let be the electricity price at time t. For time step; As a penalty cost for delay, For the task i The penalty weight, For the task i Actual completion time under strategy π.

7. The computing-electricity collaborative task scheduling method for hybrid heterogeneous computing power according to claim 6, characterized in that: During the solution process of the collaborative scheduling decision module, the constraints include: Task completion constraints: Ensure that the task execution time is no less than the estimated minimum execution time; Resource capacity constraints: ; For the task For hardware Resource consumption For hardware Total resource capacity; Delay constraints: The stringency of the constraints is determined by the time elasticity in the schedulability vector. Decide, This is a flexible constraint; Interruptible constraints: ,but exist arrive The values ​​must be consecutively 1, and no breaks are allowed.

8. A computing-powered collaborative task scheduling system for hybrid heterogeneous computing power according to claim 7, characterized in that: The scheduling policy translation logic of the scheduling instruction execution module includes: For the Kubernetes platform: Translate tasks into Pod / Job YAML files, defining container images, startup commands, and parameters; translate hardware into node affinity or node selectors; translate time into immediate creation or delayed queues; translate pause for interruptible tasks into adjusting the number of Job replicas or deleting Pods; translate resume into restoring the number of replicas or recreating the Pod and resuming from the checkpoint. For the Slurm platform: tasks are translated into .sbatch batch scripts, including execution commands and parameters; hardware is translated into #SBATCH hardware constraint instructions or partition specification instructions; time is translated into immediate execution or future execution; and pause of interruptible tasks is translated into scontrol hold.<job_id> The correct translation is scontrol release.<job_id> .

9. A computing power collaborative task scheduling method for hybrid heterogeneous computing power, executed by a computing power collaborative task scheduling system for hybrid heterogeneous computing power as described in any one of claims 1-8, characterized in that, Includes the following steps: Step 1: Real-time perception of multi-dimensional status. Task information, computing power status and power information are acquired through the data acquisition module, and stored in the time series database after timestamp alignment and cleaning preprocessing. Step 2: Build and update the energy consumption characteristic model. Extract time series database data through the energy consumption characteristic modeling module, complete feature engineering, model training and model maintenance, and output a high-precision energy consumption prediction model. Step 3: Quantify the schedulable potential of tasks. Perform rule matching and vector calculation on each task to be scheduled through the task schedulability analysis module to generate a schedulability vector. Step 4: Solve the computational-electricity collaborative scheduling optimization problem, including algorithm routing, algorithm selection, and execution. During algorithm selection, a hybrid decision-making routing function is implemented through a routing controller, and either a deep reinforcement learning algorithm or a genetic algorithm is selected as the solution algorithm. During execution, the energy consumption prediction model is called to obtain the expected power consumption, and the electricity cost is calculated in combination with the electricity price. Under the constraints of task completion, resource capacity, latency, and interruptibility, the selected solution algorithm is used to solve the optimization problem that minimizes the total operating cost and generate the optimal scheduling strategy. Step 5: Issue and execute scheduling instructions. The scheduling instruction execution module translates the optimal scheduling strategy into underlying platform instructions and executes them. Step 6: Status feedback and closed loop. The scheduling instruction execution module monitors the actual execution status of the task and feeds it back to the data acquisition module, providing a basis for subsequent data perception and model updates.

Images