Task scheduling method for computing power cluster, electronic equipment and storage medium

Abstract

The invention relates to a task scheduling method of a computing power cluster, electronic equipment and a storage medium. The method comprises the steps that at the starting moment of a planning time domain corresponding to each scheduling period, scheduling domain state data of a computing power cluster at the starting moment and out-of-domain state data in the planning time domain are obtained, the out-of-domain state data comprise observation data at the starting moment and prediction data at future moments in the planning time domain, and the planning time domain covers a plurality of time slices; according to the scheduling domain state data and the out-of-domain state data, a scheduling decision corresponding to each time slice in the planning time domain is generated with the purpose of optimizing the total operation cost of the computing power cluster in the planning time domain, the scheduling decision comprises a task object, execution hardware equipment and power consumption constraints, and a scheduling decision storage area is updated; and in the current time slice, controlling the computing power cluster to operate according to the corresponding scheduling decision in the scheduling decision storage area. Through prospective planning and power consumption constraint setting, the total operation cost of the computing power cluster is globally and remarkably reduced.

Description

Technical Field

[0001] This disclosure relates to the field of computing resource management technology, and more specifically, to a task scheduling method for a computing power cluster, an electronic device, and a non-transitory computer-readable storage medium. Background Technology

[0002] A computing cluster is a collection of resources used to execute computing tasks, and it is widely deployed in scenarios such as artificial intelligence, big data, high-performance computing, and cloud computing. A computing cluster contains multiple computing nodes, and during operation, task scheduling is required to match computing tasks with computing nodes. Existing task scheduling schemes mostly focus on real-time decisions based on the current hardware resource status and task status, making it difficult to control the global operating cost of the computing cluster within the expected level and lacking the ability to achieve sustainable global optimization. Summary of the Invention

[0003] One objective of this disclosure is to provide a new technical solution for computing cluster scheduling, enabling cross-time-slice collaborative optimization of computing cluster operating costs.

[0004] According to a first aspect of this disclosure, a task scheduling method for a computing power cluster is provided, comprising: At the start of the planning time domain corresponding to this scheduling, the scheduling domain status data of the computing power cluster at the start of the scheduling time domain and the external status data related to the operating cost of the computing power cluster within the planning time domain are obtained; wherein, the scheduling domain status data includes task status data and hardware resource status data, and the external status data includes observation data at the start of the scheduling time domain and prediction data for future times within the planning time domain, and the planning time domain covers multiple time slices; Based on the scheduling domain state data and external state data corresponding to the planning time domain, with the goal of optimizing the total operating cost of the computing cluster within the planning time domain, a scheduling decision is generated for each time slice within the planning time domain, and the scheduling decision storage area is updated based on the generated scheduling decision; wherein, the scheduling decision includes multiple scheduling units for the corresponding time slice, and each scheduling unit includes a task object, a hardware device for executing the task object, and a power consumption constraint set for the hardware device; In the current time slice, the control of the computing cluster is to run according to the scheduling decision corresponding to the time slice in the scheduling decision storage area.

[0005] Optionally, the time length of the planning time domain covers multiple scheduling cycles, wherein the scheduling cycle is the time interval between two adjacent scheduling operations.

[0006] Optionally, the time length of the planning time domain allows the external state data to cover multiple levels of unit operating costs.

[0007] Optionally, the scheduling planning operation of generating the scheduling decision corresponding to each time slice in the planning time domain and updating the scheduling decision storage area is decoupled from the scheduling control operation of controlling the computing cluster to run according to the scheduling decision corresponding to the time slice in the scheduling decision storage area.

[0008] Optionally, the method further includes: In response to the event that the current scheduling has ended, the actual operating cost of the computing cluster during the current scheduling is obtained; The reward value for this scheduling is determined based on the actual operating cost; wherein the reward value is negatively correlated with the actual operating cost. The decision parameters of the agent used to generate the scheduling decision are updated based on the reward value.

[0009] Optionally, generating the scheduling decision corresponding to each time slice within the planning time domain includes: For time slices where the unit operating cost indicated by the out-of-domain state data is at a preset minimum level, the upper limit of the power consumption constraint for at least some scheduling units within that time slice is greater than the rated power consumption of the corresponding hardware device; and / or, For scheduling units whose remaining time for a task object is less than a set duration, the upper limit of their power consumption constraint is greater than the rated power consumption of the corresponding hardware device; wherein, the remaining time is the time difference between the deadline of the corresponding task object and the start time of the corresponding time slice, and the set duration is less than twice the length of the time slice.

[0010] Optionally, the computing power cluster includes hardware devices of various specifications. The step of generating scheduling decisions for each time slice within the planned time domain, based on the scheduling domain state data and external state data corresponding to the planned time domain, with the goal of optimizing the total operating cost of the computing power cluster within the planned time domain, includes: Based on the scheduling domain state data and external state data corresponding to the planning time domain, as well as the performance data of each type of hardware device, a scheduling decision is generated for each time slice in the planning time domain with the goal of optimizing the total operating cost of the computing cluster within the planning time domain; wherein, the performance data reflects the mapping relationship between the computing performance and power consumption of the corresponding hardware device.

[0011] Optionally, the total operating cost includes a time penalty cost; the time penalty cost is determined by matching the priority and / or timeout duration of the planned task in a preset penalty cost mapping table. The penalty cost mapping table includes timeout penalty costs under different priorities, different timeout durations, or combinations of different priorities and different timeout durations. At least some items in the penalty cost mapping table have their timeout penalty costs set to a preset unacceptable penalty upper limit.

[0012] According to a second aspect of this disclosure, an electronic device is also provided, the electronic device comprising: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method according to the first aspect of this disclosure.

[0013] According to a third aspect of this disclosure, a non-transitory computer-readable storage medium is also provided, the non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method according to the first aspect of this disclosure.

[0014] In this embodiment, each scheduling cycle corresponds to a preset planning time domain, which covers multiple time slices. By acquiring the external state data related to the computing cluster's operating cost within the corresponding preset planning time domain when each scheduling cycle arrives, and combining it with the current scheduling domain state data, scheduling decisions can proactively respond to changes in external state over a future period, achieving cross-time slice collaborative planning and significantly improving the global planning capability of computing cluster scheduling. Furthermore, in this embodiment, by actively incorporating the operating power consumption of hardware devices into the decision variables for optimization, the expected operating cost of the scheduling decision can be well matched with the actual execution result, thereby ensuring the reliable achievement of the optimization objective.

[0015] The features and advantages of the embodiments of this specification will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments of this specification and, together with their description, serve to explain the principles of these embodiments.

[0017] Figure 1 This is a schematic diagram of the composition structure of a computing power cluster provided in this disclosure; Figure 2 This is a flowchart illustrating a computing cluster scheduling method according to some embodiments; Figure 3 This is a flowchart illustrating a computing cluster scheduling method according to other embodiments; Figure 4 This is a service architecture of a computing cluster scheduling system according to some embodiments; Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to some embodiments. Detailed Implementation

[0018] Various exemplary embodiments of this specification will now be described in detail with reference to the accompanying drawings.

[0019] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the embodiments of this specification or their application or use.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0021] It should be noted that all data acquisition actions in this disclosure were carried out in compliance with the relevant data protection laws and policies of the country where the data is located, and with the authorization of the relevant equipment owner.

[0022] This disclosure relates to a technical solution for task scheduling of hardware resources in a computing power cluster. Figure 1 A schematic diagram of the composition structure of a computing power cluster is shown. For example... Figure 1 As shown, the computing cluster comprises multiple nodes, denoted as node n1, node n2, ..., node nX, where X is a positive integer. These nodes are computing nodes with computing capabilities, and each node contains at least one type of processor, such as at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Neural-network Processing Unit (NPU), or Tensor Processing Unit (TPU). Furthermore, each node is typically configured with a communication interface for inter-node data communication, such as an Ethernet interface, fiber optic communication interface, or other high-speed interconnect interface. Some nodes may further include local memory for storing temporary or long-term data required for node operation. For example, ... Figure 1 As shown, node N2 includes at least one processor 201, at least one memory 202, and a communication interface 203. When node N2 contains multiple processors, these processors can be of the same type or different types to adapt to different computing task requirements.

[0023] Since a computing cluster comprises multiple computing nodes, efficient task execution requires matching tasks with computing nodes through task scheduling during operation. Traditional task scheduling systems (such as Hadoop YARN, Kubernetes, and Slurm) prioritize maximizing resource utilization and ensuring Quality of Service (QoS). They employ static or quasi-static strategies based on resource requests (CPU, memory, etc.), limiting decision-making to internal cluster resource status (such as load and memory utilization) and ignoring external environmental factors related to the operating costs of the computing cluster, such as electricity prices and grid carbon emission intensity. Under the backdrop of volatile energy prices and the "dual carbon" objective, such scheduling cannot dynamically adjust task execution strategies based on energy costs, easily leading to the execution of non-urgent tasks during peak electricity price periods, resulting in economic waste and unnecessary carbon emissions. Furthermore, it fails to differentiate between grid carbon intensity (i.e., carbon emission intensity) differences across different time periods and regions, resulting in an increased carbon footprint.

[0024] To optimize the operating costs of computing clusters, related technologies attempt to collect real-time task queues, hardware resource status, real-time electricity prices, and carbon intensity at each scheduling moment, and use deep reinforcement learning or heuristic rules to generate task scheduling instructions (task-node matching) for the current moment. However, the scheduling decisions of these technologies are instantaneous and cannot proactively utilize the changing trends of the external environment over multiple future time slices to achieve cross-time-slice resource allocation and timing coordination among multiple tasks, easily getting trapped in local optima. Moreover, such scheduling decisions are limited to task-resource matching; the actual operating power consumption of hardware devices is managed by default at the underlying level and is not included in the decision variables, resulting in a large deviation between the expected operating cost and the actual execution result, making it difficult to reliably achieve the optimization goal.

[0025] To this end, this disclosure proposes a prediction-driven task scheduling method, which can be implemented by a task scheduling system 100 of a computing power cluster. Based on the predicted external state data for a future period of time, the method aims to optimize the total operating cost of the computing power cluster in multiple time slices in the future. It performs forward-looking global planning across time slices and incorporates the power consumption of the hardware executing the assigned tasks into the decision variables. From the two dimensions of scheduling decision and hardware operation constraints, it achieves effective optimization of the total operating cost of the computing power cluster in that period of time.

[0026] like Figure 1As shown, in terms of hardware composition, the task scheduling system 100 includes at least one processor 101 and at least one memory 102. The processor 101 executes a computer program to perform the task scheduling method according to embodiments of this disclosure. The computer program can be written based on various instruction set architectures. The memory 102 stores the executable computer program and may be of a type including non-volatile storage media such as read-only memory (ROM), random access memory (RAM), and hard disk. Furthermore, the task scheduling system 100 may also include a communication interface 103 for data communication between itself and nodes within the computing facility, as well as with external resources, supporting the acquisition and transmission of various types of data. The deployment of the task scheduling system 100 is highly flexible. For example, it can be deployed on one or more specific nodes in a computing cluster; it can be deployed independently of the computing cluster, such as on a dedicated monitoring node or edge device; or it can adopt a hybrid deployment mode, where some functions are deployed on nodes of the computing cluster, such as data acquisition functions, while other functions are deployed in the cloud, etc. This disclosure does not limit this approach.

[0027] The following is combined with Figure 1 The schematic diagram of the computing cluster and task scheduling system 100 illustrates various embodiments of this disclosure.

[0028] <First Embodiment> This embodiment provides a task scheduling method for a computing power cluster. Figure 2 A flowchart illustrating the task scheduling method according to this embodiment is shown. This method, based on prediction-driven state awareness, decision-making, and execution, achieves cross-time-slice global collaborative optimization of computing cluster operating costs while ensuring computing performance. The method of this embodiment can be derived from... Figure 1 The task scheduling system 100 shown is implemented.

[0029] like Figure 2 As shown, the task scheduling method of this embodiment may include the following steps S210 to S230: Step S210: At the start time of the planning time domain corresponding to this scheduling, obtain the scheduling domain status data of the computing power cluster at the start time, as well as the external domain status data related to the operating cost of the computing power cluster within the planning time domain.

[0030] The task scheduling system 100 can perform scheduling according to the set scheduling period, which is the time interval between two adjacent scheduling sessions.

[0031] In this embodiment, each scheduling corresponds to a planning time domain, and the start time of the planning time domain corresponding to the current scheduling (or the current scheduling) is the start time of this scheduling. The planning time domain represents the time range within which the task scheduling system 100 performs cross-time-slice global collaborative planning, covering multiple time slices, with a duration such as 12 hours or 24 hours. The time slice is the basic time unit for the task scheduling system 100 to make scheduling decisions, with a duration such as 15 minutes, 30 minutes, or 60 minutes.

[0032] In each scheduling process, the task scheduling system 100 will execute a cross-time-slice scheduling plan for the corresponding planning time domain and generate a scheduling decision for each time slice within that planning time domain.

[0033] In this embodiment, the scheduling domain status data required for this scheduling is the real-time observation data of the computing cluster at the start of this scheduling, including task status data and hardware resource status data.

[0034] Task status data includes, for example, some or all of the data for each task in the task list, such as task identifier, hardware resource requirements, remaining workload, deadline, priority, and historical execution status. Hardware resource requirements include the number of CPU cores and / or memory size; historical execution status reflects whether the corresponding task has been assigned to one or more hardware devices for execution, and the identifiers of the historically assigned hardware devices. For example, for an AI model training task Task_A, its task identifier is "Task_A", the remaining workload is 1G floating-point operations (FLOPs), the deadline is 12 hours after the start of the current scheduling cycle, the priority is set to high priority, and the hardware resource requirements are at least 64 CPU cores.

[0035] Hardware resource status data reflects the current usage status of hardware resources in the computing cluster. For example, hardware resource status data includes some or all of the data such as the availability status, current power consumption, current temperature, and current utilization rate of each hardware device in the computing cluster. For example, for a GPU device, its availability status is idle, its current power consumption is 150W, its temperature is 45℃, and its utilization rate is 0%.

[0036] In this embodiment, the external state data refers to external data related to the operating cost of the computing power cluster. The external state data within the planning time domain includes observation data at the start of the planning time domain and predicted data for future times within the planning time domain. The external state data acquired by the task scheduling system 100 in step S210 can be represented as a data sequence, where each data point is arranged in timestamp order, with the first data point being the observation data at the start time and the remaining data points being the predicted data.

[0037] For example, off-domain status data may include at least one of electricity price time series data and carbon intensity time series data.

[0038] Taking an external state data set including electricity price time-series data and carbon intensity time-series data, with a planned time domain of 24 hours, as an example, the observation data includes real-time electricity price and real-time carbon intensity at the initial time. For example, at the initial time t0, the real-time electricity price... Yuan / kWh, real-time carbon intensity The forecast data includes electricity price forecasts and carbon intensity forecasts for the next 24 hours. For example, the forecast data shows that during the period from 2:00 AM to 5:00 AM, the electricity price will drop to 0.3 yuan / kWh, the off-peak hour price, and the carbon intensity will drop to... However, during the evening hours of 6:00 PM to 10:00 PM, the electricity price rises to 1.2 yuan / kWh, the peak period, and the carbon emission intensity rises to [missing information]. The task scheduling system 100 can obtain this extra-domain status data from the APIs of the power grid operator and the environmental protection department through an external data adapter, so as to provide input information for the system to perform global optimization across time slices.

[0039] Step S220: Based on the scheduling domain state data and external state data corresponding to the planning time domain, and with the goal of optimizing the total operating cost of the computing power cluster within the planning time domain, generate scheduling decisions for each time slice within the planning time domain.

[0040] This embodiment uses the total operating cost of the computing cluster within the planned time domain as the optimization objective, and performs task scheduling planning from a global perspective across time slices to control this total operating cost. For example, the task scheduling system 100 can be configured to perform scheduling planning with the objective of minimizing the total operating cost of the computing cluster within the planned time domain. Alternatively, the task scheduling system 100 can also be configured to perform scheduling planning with the objective of controlling the total operating cost of the computing cluster within the planned time domain below a set cost.

[0041] In some examples, the total operating cost of the computing cluster within the planning time domain. It consists of three weighted terms: electricity price cost Carbon emission costs and time penalty cost Therefore, the objective function for optimization can be expressed as: ; in, For electricity cost The weighting coefficients, For carbon emission costs The weighting coefficients, Time penalty cost Weighting coefficients for electricity costs. This can be expressed as: the sum of the power consumption of all hardware devices within all time slices, the execution duration, and the electricity price during the execution period. Carbon emission cost. This can be expressed as: the sum of the power consumption of all hardware devices within all time slices multiplied by the execution duration, and then multiplied by the carbon intensity of the execution period. Time penalty cost. The penalty cost can be determined by matching against a pre-defined penalty cost mapping table based on at least one of the priority and timeout duration of the planned task. For example, the penalty cost mapping table may include timeout penalty costs under different priorities, different timeout durations, or combinations of different priorities and different timeout durations.

[0042] In this example, the objective function can be minimized to perform scheduling planning for the objective, thereby generating scheduling decisions for each time slice within the planning time domain.

[0043] In this example, the task scheduling system 100 can provide an interactive interface, allowing users to set the weight coefficients of each weighting item to match different application scenarios.

[0044] In another example, total operating cost It can also be composed of any two of the three weighting terms mentioned above, and may include more weighting terms; this disclosure does not limit this.

[0045] In the example where the total operating cost includes time penalty costs, the timeout penalty costs corresponding to at least some items in the penalty cost mapping table are set to a preset unacceptable penalty upper limit. This unacceptable penalty upper limit makes it impossible for the corresponding scheduling scheme to meet the optimization objective, and thus it is naturally excluded during the optimization process, effectively prohibiting scheduling schemes that meet these timeout items from being used as the generated scheduling decision. In this example, the time penalty cost... The corresponding weighting term dominates the objective function. This setting ensures that the task scheduling system prioritizes the completion of planned tasks before the deadline during the optimization process. For tasks that are close to the deadline or have already exceeded the deadline, the task scheduling system will temporarily disregard electricity costs and carbon emission costs and prioritize their scheduling. Furthermore, the upper limit of the power consumption constraint corresponding to the relevant task can be forcibly increased to above the rated power consumption of the hardware device, or even to the maximum power consumption value allowed by the hardware device, and the most powerful hardware device can be assigned to it.

[0046] In this example, for application scenarios where timeouts are not allowed, optimization conditions can be set to prohibit timeouts in the optimization algorithm, or the timeout penalty cost corresponding to all timeout items can be set to an unacceptable upper limit in the penalty cost mapping table.

[0047] In this embodiment, the scheduling decision generated by the task scheduling system 100 for each time slice within the planning time domain includes multiple scheduling units for the corresponding time slice. Each scheduling unit is a three-dimensional array, including a task object, the hardware resources for executing the task object, and power consumption constraints set for the hardware resources. Both the task object and the hardware resources in the scheduling unit can be represented by their unique identifiers. In this embodiment, the power consumption constraint can be a power consumption upper limit or a power consumption range. In this embodiment, a task object can be assigned to one hardware device or multiple hardware devices within a time slice. The combination of the task object and each hardware device forms different scheduling units.

[0048] For example, for time slice t1 within the planning time domain, there are scheduling units such as (Task_A, GPU_1, 450W) and (Task_B, GPU_2, 250W). The scheduling unit (Task_A, GPU_1, 450W) means that task Task_A is assigned to GPU_1 for execution and the power consumption limit of GPU_1 is set to 450W. The scheduling unit (Task_B, GPU_2, 250W) means that task Task_B is assigned to GPU_2 for execution and the power consumption limit of GPU_2 is set to 250W.

[0049] In some examples, the duration of the planning time domain corresponding to this scheduling is set to ensure that the out-of-domain state data covers multiple levels of unit operating cost. Here, unit operating cost refers to the operating cost incurred by the computing cluster consuming a unit of electricity (1 kWh), which may include, for example, electricity price costs and / or carbon emission costs. By ensuring that the out-of-domain state data acquired in each scheduling session covers multiple cost levels (such as preset high, medium, and low levels), the task scheduling system 100 can fully integrate the unit operating cost distribution within the planning time domain in a single scheduling plan, performing global optimization across time slices, thereby facilitating the achievement of lower total operating costs through scheduling planning.

[0050] In some examples, when generating scheduling decisions, the task scheduling system 100 can adopt different power consumption constraint strategies based on the level of unit operating cost indicated by the out-of-domain state data. For time slices where the unit operating cost indicated by the out-of-domain state data is at a preset minimum level, such as during off-peak hours in the late night or early morning when electricity prices are lowest and carbon intensity is also low, the upper limit of the power consumption constraint for at least some scheduling units within these time slices can be set to be greater than the rated power consumption of the corresponding hardware devices. For example, the power consumption constraint for at least some scheduling units within these time slices can be set to the maximum power consumption allowed by the corresponding hardware devices, so as to consume cheap / clean energy as quickly as possible, accumulate computing progress for subsequent time slices, and thus minimize electricity cost and carbon emission cost.

[0051] In some examples, when generating scheduling decisions, the task scheduling system 100 can adopt different power consumption constraint strategies based on the unit operating cost level indicated by the external state data. At least for time slices where the external state data indicates that the unit operating cost is at a preset minimum level, such as during off-peak hours in the late night or early morning when electricity prices and carbon intensity are low, the upper limit of the power consumption constraint for at least some scheduling units in this time slice can be set to be greater than the rated power consumption of the corresponding hardware device. For example, it can be set to the maximum power consumption value allowed by the corresponding hardware device, so as to make full use of cheap or clean energy to quickly complete the calculation, reserve the calculation margin for subsequent time periods, and thereby reduce the total operating cost of the computing power cluster in the corresponding planned time domain.

[0052] In other examples, for time slices where the unit operating cost indicated by the out-of-domain state data is at a preset medium level, such as during the daytime, the upper limit of the power consumption constraint for at least some scheduling units within that time slice can be set to be equal to the rated power consumption of the corresponding hardware device.

[0053] In other examples, for time slices where the unit operating cost indicated by the out-of-domain state data is at a preset high level, such as the evening peak period, the upper limit of the power consumption constraint for at least some scheduling units in that time slice can be set below the rated power consumption of the corresponding equipment to avoid high electricity prices and / or high carbon intensity.

[0054] In some other examples, when generating scheduling decisions, if the remaining time for a task object corresponding to a scheduling unit is less than a set duration, the upper limit of the power consumption constraint for that scheduling unit can be set to be greater than the rated power consumption of the corresponding hardware resources. Here, the remaining time is the time difference between the deadline of the corresponding task object and the start time of its time slice, and the set duration is less than twice the length of the time slice. This setting aims to accelerate task execution by increasing hardware power consumption above the rated value within the current time slice, thereby reducing or avoiding timeouts and lowering the total operating cost of the computing cluster within the corresponding planned time domain.

[0055] In some examples, the computing cluster employs a heterogeneous hardware architecture, including various specifications of hardware devices. These specifications are manifested in differences in type and / or model. For example, the computing cluster includes at least two of the following hardware devices: CPU, GPU, GPGPU, NPU, and TPU. Furthermore, it may also include different models of the same type of hardware device. Different specifications of hardware devices have different computing performances. The task scheduling system 100 can maintain a hardware configuration table for the computing cluster. This table includes performance data for each specification of hardware device, used to characterize the mapping relationship between the computing performance and power consumption of the corresponding hardware device. For example, for GPU_1, its performance data can be represented as Eff(P, R) = {150W: 50 samples / s, 200W: 80 samples / s, 300W: 130 samples / s, 400W: 160 samples / s}, indicating that at power consumptions of 150W, 200W, 300W, and 400W, its computing performance is 50, 80, 130, and 160 samples per second, respectively. These performance data can be obtained by running standardized benchmark tests on each type of hardware device in advance, providing a basis for the task scheduling system 100 to perform fine-grained scheduling planning. Compared with the general adaptation method based on standard computing performance, the scheduling planning based on the hardware configuration table enables the task scheduling system to accurately estimate the time required for task completion and the corresponding energy consumption under the power consumption when setting power constraints, thereby ensuring that the operating cost of the scheduling decision is highly consistent with the actual execution result.

[0056] In these examples, the hardware configuration table can be updated in real time based on requests to add or remove hardware from the computing cluster.

[0057] Step S230: Update the scheduling decision storage area based on the generated scheduling decision.

[0058] The scheduling decision storage area, for example, is implemented using a high-speed Redis storage system, used to store the complete scheduling decisions generated by the optimization. When the scheduling decision storage area is empty, the current scheduling decision generated in this scheduling can be written to the scheduling decision storage area to update it; when the scheduling decision storage area stores historical scheduling decisions, the current scheduling decision can overwrite the historical scheduling decisions to update the scheduling decision storage area.

[0059] The scheduling strategy for each time slice within the optimized planning time domain can be serialized and stored in the decision storage area. The storage format for scheduling decisions can be: the key is the time slice identifier (e.g., "2025-01-15_02:00"), and the value is the list of scheduling units corresponding to that time slice.

[0060] In some examples, the planning time domain can cover multiple scheduling cycles, meaning that the planning time domains corresponding to two adjacent scheduling operations overlap in time, thus enabling dynamic rolling optimization. For example, the scheduling cycle can be set to 30 minutes or 60 minutes, and the planning time domain can be set to 12 hours or 24 hours; the scheduling cycle can correspond to one time slice or contain two or more time slices. Dynamic rolling optimization can adjust the scheduling strategy corresponding to each time slice in a timely manner based on the dynamic changes in scheduling domain state data, external domain state data, etc., making the expected operating cost of scheduling decisions more consistent with the actual execution cost.

[0061] In another example, the length of the planning time domain can also be equal to the scheduling period. In this example, the planning time domain can have a relatively short length; for example, the planning time domain can be set to 6 hours or 12 hours, etc.

[0062] Step S240: In the current time slice, control the computing cluster to run according to the scheduling decision corresponding to the time slice in the scheduling decision storage area.

[0063] In this embodiment, the task scheduling system 100 reads the scheduling decision corresponding to the current time slice from the scheduling decision storage area and translates it into control instructions for the hardware devices. For each scheduling unit (Taskid, ResourceR, PowerP), the task scheduling system 100 can allocate the task object Taskid to the hardware device ResourceR for execution through the cluster scheduler interface, and PowerP represents the power consumption constraint set for the hardware device ResourceR. The cluster scheduler interface acts as a bridge between this system and existing cluster management frameworks (such as Kubernetes), and is responsible for translating the scheduling units into API calls of the corresponding scheduler. For example, for the scheduling unit (Task_A, GPU_1, 450W), the cluster scheduler interface can call the Kubernetes API to create a Pod and schedule the container of task_A to run on GPU_1 of the GPU node. Then, the task scheduling system 100 can dynamically set the power consumption constraints of the hardware devices in the computing node through the hardware control module deployed on the computing node. For example, for a scheduling unit (Task_A, GPU_1, 450W), the hardware control module executes a power consumption setting command on the GPU node to set the power consumption limit of GPU_1 to 450W. When GPU_1 executes Task_A, if the actual power consumption reaches the set power consumption limit of 450W, the clock frequency of GPU_1 will be automatically reduced to maintain the power consumption within the limit. In this way, the task scheduling system can precisely control the power consumption of hardware devices at the hardware level, ensuring that the actual operating state of the hardware devices matches the scheduling decisions.

[0064] In this embodiment, steps S220 to S230 are scheduling planning operations, and step S240 is a scheduling control operation. In some examples, such as Figure 3 As shown, scheduling planning operations and scheduling control operations can be decoupled, meaning the two operations are independent of each other. In these examples, the task scheduling system 100 can directly read the scheduling policy corresponding to the current time slice from the scheduling policy storage area when each time slice arrives. If the scheduling policy storage area does not contain a scheduling policy corresponding to the current time slice, it will loop through the storage area until the scheduling policy is found, without needing to check whether a scheduling planning operation has been started or is being started, which simplifies the scheduling logic. In addition, this example, combined with dynamic rolling optimization, can adjust the scheduling policy in a timely manner according to changes in the scheduling domain state and external state, while ensuring the real-time performance of the scheduling response.

[0065] In this embodiment, steps S210 to S230 correspond to scheduling planning operations, and step S240 corresponds to scheduling control operations. In some examples, scheduling planning operations and scheduling control operations can be decoupled. This decoupling design allows scheduling planning and scheduling control to be performed asynchronously. The task scheduling system 100 can directly read the scheduling policy corresponding to the current time slice from the scheduling policy storage area when each time slice arrives, without waiting for the ongoing scheduling planning operation to complete, thereby improving the system's response speed and stability.

[0066] In another example, the task scheduling system 100 can also be configured such that when each time slice arrives, if there is no ongoing scheduling planning operation, the scheduling policy corresponding to the current time slice is read directly from the scheduling policy storage area; if there is an ongoing scheduling planning operation, it can wait for it to complete before reading. For example, if a scheduling planning operation takes about 2 minutes, in a time slice that is aligned with the start time of this scheduling, the task scheduling system 100 can wait for it to complete before reading the scheduling policy corresponding to that time slice.

[0067] The task scheduling system 100 can obtain the execution status data of the executed tasks by monitoring the execution status of the executed tasks within the current time slice, and update the task status data of the corresponding tasks based on the execution status data, such as updating the remaining workload, for use in the next scheduling.

[0068] According to steps S210 to S240, this embodiment uses a prediction-driven mechanism to perform forward-looking global planning of the operating cost of the computing power cluster across time slices, which can effectively break through the limitations of local optima and obtain a scheduling strategy with sustainable low-cost operation capabilities. At the same time, this embodiment incorporates the power consumption of the hardware executing the assigned tasks into the decision variables, and from the two dimensions of scheduling decision and hardware operation constraints, it realizes effective control over the total operating cost of the computing power cluster.

[0069] <Second Embodiment> This embodiment provides a task scheduling method for computing power clusters based on reinforcement learning, which enables the agent used to generate scheduling decisions to update its policies through interaction with the environment, thereby allowing the generated scheduling decisions to obtain the expected cumulative rewards in actual operation.

[0070] In this embodiment, the agent employs reinforcement learning (RL) as the core optimization algorithm, which can model the scheduling problem as a Markov decision process (MDP). The MDP consists of quintuples... Define $\frac{ ... This is the discount factor.

[0071] The state space S includes scheduling domain state data and external domain state data, etc.

[0072] Action space A includes the mapping between tasks and hardware devices, as well as power consumption constraints. Action vector. The elements are (Taskid, ResourceR, PowerP), indicating that task Taskid is assigned to hardware resource ResourceR for execution, and the power consumption constraint of ResourceR is set to PowerP. The action space can be a complex space that mixes discrete and continuous components. The discrete part is the task-hardware device mapping, and the continuous part is the power consumption constraint setting, such as a power consumption range of 150W-500W.

[0073] reward function It can be defined as the negative of the objective function, that is, the reward value of this scheduling and the actual operating cost generated by the scheduling decision movement based on this scheduling. Negative correlation: ; in, , , These are the actual electricity price cost, actual carbon emission cost, and actual time penalty cost, respectively. Actual electricity price cost It equals the sum of the power consumption of all devices that have performed tasks, the actual execution time, and the actual electricity price. Actual carbon emission cost. This equals the sum of the power consumption of all executed tasks, multiplied by the actual execution time, and then multiplied by the actual carbon intensity. Actual time penalty cost. It equals the sum of the penalty values ​​for all timed-out tasks.

[0074] Therefore, minimizing electricity costs, carbon emission costs, and time penalty costs is equivalent to maximizing cumulative rewards. At the end of the scheduling cycle, the task scheduling system 100 obtains the actual operating costs of the computing cluster during the scheduling cycle. And based on actual operating costs The reward value corresponding to this scheduling is determined; the lower the actual operating cost, the higher the reward value.

[0075] In some examples, the agent's policy network This can be achieved using a deep neural network (DNN), with the input state vector. The output layer shows the probability distribution of each action. The policy network structure can be, for example, but is not limited to, a 3-layer fully connected neural network: input layer (corresponding to the state space dimension) → 512-dimensional hidden layer → 256-dimensional hidden layer → output layer (corresponding to the action space dimension). The output layer can use the Softmax activation function to output the probability distribution of each action. The value network V(S) can also be implemented using a deep neural network, with the input state vector... The value network estimates the value of the output state. The structure of a value network is, for example, but not limited to, a three-layer fully connected neural network: input layer → 512-dimensional hidden layer → 256-dimensional hidden layer → 1-dimensional output layer. Value networks are used to estimate the value of a state. Next, according to the current strategy The expected cumulative reward that can be obtained.

[0076] Based on the reward value, the system updates the decision parameters of the agent used to generate scheduling decisions, i.e., the parameters of the policy network. The update method can employ the Proximal Policy Optimization (PPO) algorithm. The core idea of ​​the PPO algorithm is to limit the step size of policy updates to avoid performance degradation caused by excessively large policy updates.

[0077] Through training and continuous optimization of the reinforcement learning model described above, the task scheduling system can learn an optimal scheduling strategy, which enables it to maximize long-term cumulative rewards and minimize total operating costs in complex and dynamic environments.

[0078] <Third Embodiment> This embodiment provides a service architecture for an alternative task scheduling system. For example... Figure 4 As shown, the task scheduling system can adopt a layered, loosely coupled microservice architecture to ensure high scalability, high availability, and compatibility with heterogeneous environments.

[0079] In this embodiment, the task scheduling system is divided into three logical layers: a data acquisition layer, an intelligent decision-making layer, and a resource control layer. Each layer can be developed, deployed, and expanded independently to adapt to rapid changes in business needs. For example, when support for new hardware types is required, corresponding adapters can be added to the data acquisition layer and the resource control layer without modifying the core intelligent decision-making layer. Furthermore, the task scheduling system supports individual algorithm upgrades and optimizations without affecting the data and control pathways.

[0080] The data acquisition layer is responsible for comprehensively and in real-time collecting the internal and external data required by the intelligent decision-making layer. The internal data required by the intelligent decision-making layer is scheduling domain state data, including task state data and hardware resource state data; the external data required by the intelligent decision-making layer is off-domain state data related to the operating cost of the computing power cluster, such as at least one of electricity price time-series data and carbon intensity time-series data.

[0081] The external data adapter in the data acquisition layer can be a pluggable module responsible for interacting with external system interfaces. For example, it can obtain electricity price time-series data for each scheduling operation from the external interface of the power grid operator, corresponding to the planned time domain. And / or, obtain time-series carbon intensity data of the power grid from external interfaces of environmental protection departments or third-party data providers. .

[0082] The Cluster Telemetry Service, deployed on each node of the computing cluster, is responsible for monitoring and collecting scheduling domain status data. This data provides the task scheduling system with an accurate real-time snapshot of the computing cluster's status. This data includes, for example, the actual power consumption, utilization, temperature, and memory usage of the hardware devices within the nodes, as well as a list of currently running tasks.

[0083] The Performance Benchmark Library in the data acquisition layer stores performance data for each hardware device in the computing cluster. This performance data reflects the mapping relationship between the computing performance and power consumption of the corresponding hardware device. The performance data of each hardware device can be obtained by running standardized benchmark tests (e.g., ResNet50 training) on ​​that hardware device beforehand. The performance data describes the computing performance (e.g., the number of samples processed per second) that the hardware device can provide under different power consumption limits.

[0084] The intelligent decision-making layer is responsible for optimizing the total operating cost of the computing cluster and generating scheduling decisions based on global planning.

[0085] The Task Manager in the intelligent decision-making layer is responsible for maintaining metadata for all pending and executing tasks. For each task, it tracks its submission time, user-defined deadline, priority, and the total workload estimated based on the task type and input data. The Task Manager is also responsible for updating the execution progress of tasks in real time and dynamically adjusting the remaining workload of each task based on actual performance data fed back by the cluster telemetry module, providing accurate input data for the agent.

[0086] The agent receives task status data from the task manager, external status data from the data acquisition layer, and hardware resource status data, and runs an optimization algorithm aimed at optimizing the total operating cost of the computing power cluster, outputting corresponding scheduling decisions.

[0087] The scheduling policy storage area stores the scheduling decisions generated by the agent in each scheduling cycle. Scheduling decisions in later cycles overwrite those in earlier cycles. Since the agent's optimization calculations require time, to ensure real-time scheduling, the complete scheduling policy output by the agent for each cycle is stored in the scheduling policy storage area. The resource control layer can directly read the scheduling policy for each time slice from the scheduling policy storage area without waiting for the agent's real-time calculations. This decoupling ensures that decision generation and execution can be performed asynchronously, improving the system's response speed and stability.

[0088] The resource control layer is responsible for translating the abstract scheduling strategies generated by the intelligent decision-making layer into specific control instructions for hardware resources. The resource control layer may include a cluster scheduler interface, a hardware control module, and a fault-tolerant processing module.

[0089] The cluster scheduling interface acts as a bridge between the task scheduling system and existing cluster management frameworks (such as Kubernetes, Slurm, and Yarn), translating scheduling policy instructions (such as "start task A to node 2") into corresponding scheduler API calls (such as Kubernetes Pod creation requests or Slurm's srun command). This allows the task scheduling system of this disclosure to be seamlessly integrated into existing IT infrastructure as a "plugin" or "extension" without replacing the original scheduling system.

[0090] The hardware control module is a lightweight daemon deployed on each node. It is responsible for receiving power control instructions from the scheduling policy storage area and calling the underlying hardware management tools to execute them.

[0091] The fault tolerance module is responsible for handling potential failures during scheduling execution. For example, it is triggered when the hardware control module attempts to set a power consumption limit exceeding the hardware's allowable range, or when the cluster scheduling interface rejects a task allocation request due to insufficient hardware resources. The fault tolerance module records the failed instruction and attempts to execute a pre-defined rollback strategy, such as setting the power consumption back to the previous stable value or re-queuing the task indicated by the instruction, to ensure system stability and the eventual completion of the task.

[0092] Through the collaborative work of the above three-layer architecture, the computing cluster task scheduling system of this disclosure realizes the closed-loop control of "perception-decision-execution" driven by predictive data. It can optimize operating costs and carbon emissions globally while ensuring computing performance, and has a high level of intelligence and automation.

[0093] <Fourth Embodiment> This disclosure also provides an electronic device, such as... Figure 5 As shown, the electronic device 500 includes at least one processor 510 and at least one memory 520, the memory 520 being used to store computer program instructions, which, when executed by the processor 510, cause the electronic device 500 to perform a task scheduling method according to any embodiment of the present disclosure.

[0094] The electronic device 500 deploys the aforementioned task scheduling system to execute the task scheduling method; it can be a server or other types of devices.

[0095] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0096] Embodiments of this specification may be devices, methods, and / or computer program products. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the embodiments of this specification.

[0097] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0098] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0099] Computer program instructions used to perform the operations of the embodiments described herein may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the embodiments described herein.

[0100] Various aspects of embodiments of this specification are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this specification. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0101] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0102] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this specification. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be well known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0104] Various embodiments of this specification have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A task scheduling method for a computing power cluster, characterized in that, include: At the start of the planning time domain corresponding to this scheduling, the scheduling domain status data of the computing power cluster at the start of the scheduling time domain and the external status data related to the operating cost of the computing power cluster within the planning time domain are obtained; wherein, the scheduling domain status data includes task status data and hardware resource status data, and the external status data includes observation data at the start of the scheduling time domain and prediction data for future times within the planning time domain, and the planning time domain covers multiple time slices; Based on the scheduling domain state data and external state data corresponding to the planning time domain, with the goal of optimizing the total operating cost of the computing cluster within the planning time domain, a scheduling decision is generated for each time slice within the planning time domain, and the scheduling decision storage area is updated based on the generated scheduling decision; wherein, the scheduling decision includes multiple scheduling units for the corresponding time slice, and each scheduling unit includes a task object, a hardware device for executing the task object, and a power consumption constraint set for the hardware device; In the current time slice, the control of the computing cluster is to run according to the scheduling decision corresponding to the time slice in the scheduling decision storage area.

2. The method according to claim 1, characterized in that, The time length of the planned time domain covers multiple scheduling cycles, wherein the scheduling cycle is the time interval between two adjacent scheduling operations.

3. The method according to claim 1, characterized in that, The time length of the planned time domain enables the external state data to cover multiple levels of unit operating costs.

4. The method according to any one of claims 1 to 3, characterized in that, The scheduling planning operation, which generates the scheduling decision corresponding to each time slice within the planning time domain and updates the scheduling decision storage area, is decoupled from the scheduling control operation, which controls the computing cluster to run according to the scheduling decision corresponding to the time slice in the scheduling decision storage area.

5. The method according to any one of claims 1 to 3, characterized in that, The method further includes: In response to the event that the current scheduling has ended, the actual operating cost of the computing cluster during the current scheduling is obtained; The reward value for this scheduling is determined based on the actual operating cost; wherein the reward value is negatively correlated with the actual operating cost. The decision parameters of the agent used to generate the scheduling decision are updated based on the reward value.

6. The method according to any one of claims 1 to 3, characterized in that, The generation of scheduling decisions for each time slice within the planning time domain includes: For time slices where the unit operating cost indicated by the out-of-domain state data is at a preset minimum level, the upper limit of the power consumption constraint for at least some scheduling units within that time slice is greater than the rated power consumption of the corresponding hardware device; and / or, For scheduling units whose remaining time for a task object is less than a set duration, the upper limit of their power consumption constraint is greater than the rated power consumption of the corresponding hardware device; wherein, the remaining time is the time difference between the deadline of the corresponding task object and the start time of the corresponding time slice, and the set duration is less than twice the length of the time slice.

7. The method according to any one of claims 1 to 3, characterized in that, The computing power cluster includes hardware devices of various specifications. Based on the scheduling domain state data and external state data corresponding to the planning time domain, and with the goal of optimizing the total operating cost of the computing power cluster within the planning time domain, scheduling decisions are generated for each time slice within the planning time domain, including: Based on the scheduling domain state data and external state data corresponding to the planning time domain, as well as the performance data of each type of hardware device, a scheduling decision is generated for each time slice in the planning time domain with the goal of optimizing the total operating cost of the computing cluster within the planning time domain; wherein, the performance data reflects the mapping relationship between the computing performance and power consumption of the corresponding hardware device.

8. The method according to any one of claims 1 to 3, characterized in that, The total operating cost includes time penalty cost; the time penalty cost is determined by matching and / or overtime duration in a preset penalty cost mapping table based on the priority of the planned task and / or the overtime duration. The penalty cost mapping table includes timeout penalty costs under different priorities, different timeout durations, or combinations of different priorities and different timeout durations. At least some items in the penalty cost mapping table have their timeout penalty costs set to a preset unacceptable penalty upper limit.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1-8.

Images