Method and device for coordinated scheduling of computing power and power of data center

Abstract

The application provides a computing power and power collaborative scheduling method and device of a data center, relates to the technical field of data processing, and comprises the following steps: periodically collecting running state data of a plurality of computing nodes in the data center through a first data channel, and acquiring power supply parameters from intelligent power distribution units corresponding to the computing nodes in real time through a second data channel; a fusion decision model is constructed, the running state data is mapped to a computing power evaluation space, and the power supply parameters are mapped to an energy evaluation space; based on a fusion evaluation vector, a joint optimization problem containing both computing power constraints and power constraints is solved, a collaborative scheduling strategy is generated, the collaborative scheduling strategy comprises two interrelated decision sets, namely, a task allocation decision and a power supply adjustment decision; the task allocation decision is converted into a task scheduling instruction and sent to the computing nodes, and the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution units.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for the coordinated scheduling of computing power and power in a data center. Background Technology

[0002] In recent years, with the deepening of digital transformation and the rapid development of technologies such as cloud computing, big data, and artificial intelligence, data centers have become a key information infrastructure supporting the operation of modern society. However, data centers are also well-known energy consumers, with their energy consumption accounting for a growing proportion of global electricity consumption year by year. Therefore, improving power efficiency and achieving green, low-carbon, and efficient operation has become a major challenge for the industry.

[0003] Therefore, how to more effectively coordinate the scheduling of computing power and power in data centers has become an urgent problem to be solved in the industry. Summary of the Invention

[0004] This invention provides a method and apparatus for the coordinated scheduling of computing power and power in data centers, in order to solve the problem of how to more effectively coordinate the scheduling of computing power and power in data centers in the prior art.

[0005] This invention provides a method for coordinated scheduling of computing power and power in a data center, comprising: The system periodically collects operational status data from multiple computing nodes within the data center via the first data channel, and obtains power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing nodes via the second data channel, thereby achieving dual-channel data acquisition. A fusion decision model is constructed, which maps the operating status data to the computing power assessment space and the power supply parameters to the energy assessment space. A fusion assessment vector representing the comprehensive status of each computing node is generated in a unified multidimensional decision space. Based on the fusion evaluation vector, a joint optimization problem that simultaneously includes computing power constraints and power constraints is solved to generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The task allocation decision is converted into a task scheduling instruction and sent to the computing node, and the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution unit, thereby realizing the synchronous adjustment of the computing layer and the power layer.

[0006] According to a data center computing power and power coordinated scheduling method provided by the present invention, the solution of the joint optimization problem includes: Construct a multi-objective optimization function that comprehensively considers minimizing total data center energy consumption, maximizing load balancing among nodes, minimizing task migration costs, and minimizing carbon emissions; Set multi-level constraints, including upper limits on the processing capacity of computing nodes, upper limits on the power output of intelligent power distribution units, quality of service constraints for critical services, and network bandwidth constraints. The solution is obtained by using a reinforcement learning algorithm. The fused evaluation vector is used as the environmental state input, and task allocation and power supply adjustment are used as executable actions. The optimal strategy is obtained through continuous interaction and learning between the agent and the environment.

[0007] According to the present invention, a method for coordinated scheduling of computing power and power in a data center further includes: Based on historically collected operational status data and power supply parameters, a time-series prediction model is trained. The time-series prediction model adopts a deep learning network structure, which includes a recurrent neural network layer for extracting time-series features and an attention mechanism layer for identifying key factors. Using the trained time-series prediction model, the load change trend and power consumption evolution curve of each computing node are predicted within a future time window; When the prediction results show that the power consumption of a certain computing node will approach the safety threshold in the future, the preventive task migration should be given priority when generating the cooperative scheduling strategy, and some tasks should be allocated to other nodes in advance.

[0008] According to the computing power and power coordinated scheduling method for a data center provided by the present invention, the power supply control command is specifically used for: Independent control is implemented at the port level, and the output ports of each intelligent power distribution unit can independently adjust power supply parameters, including voltage amplitude adjustment, current limit setting, and power supply on / off control. The power supply mode is matched according to the task type. Stable high power supply is provided for compute-intensive tasks, dynamic power adjustment is adopted for storage-intensive tasks, and low power maintenance or complete power cut-off is performed for standby devices.

[0009] According to the present invention, a data center computing power and power coordinated scheduling method is provided, wherein the power supply control command includes at least one of the following control operations: turning on or off a designated output port of the intelligent power distribution unit; adjusting the power supply voltage value of the designated output port; setting the output power limit threshold of the designated output port; and switching the power supply mode of the intelligent power distribution unit.

[0010] According to the present invention, a method for coordinated scheduling of computing power and power in a data center further includes: The intelligent power distribution unit continuously monitors its own power parameters, analyzes parameter change characteristics in real time through an abnormal pattern recognition algorithm, and actively generates and reports alarm signals when an abnormal pattern is detected. A tiered response is implemented based on the severity of the anomaly. For mild anomalies, parameters are adaptively adjusted. For moderate anomalies, preventative migration of related tasks is triggered. For severe anomalies, fault isolation and emergency switching are implemented immediately. Record the exception handling process and results, and add exception patterns and their handling strategies to the knowledge base to enhance the system's ability to identify and handle similar exceptions.

[0011] According to the present invention, a method for coordinated scheduling of computing power and power in a data center is provided, the method specifically includes: The data center is divided into multiple autonomous management domains. Each autonomous management domain contains several computing nodes and their corresponding intelligent power distribution units, and is configured with an independent domain controller. Each autonomous management domain independently performs dual-channel data acquisition, builds a fusion decision model, solves joint optimization problems, and performs bidirectional linkage control to achieve local optimization scheduling within the domain. When the resources of a single autonomous domain are insufficient, resource support is requested from neighboring domains through an inter-domain negotiation mechanism to achieve cross-domain task migration and load balancing. A distributed ledger is used to record all cross-domain scheduling decisions and execution results.

[0012] The present invention also provides a computing power and power coordinated scheduling device for a data center, comprising the following modules: The acquisition module is used to periodically acquire the operating status data of multiple computing nodes in the data center through the first data channel, and to obtain the power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing node through the second data channel, so as to realize dual-channel data acquisition. The construction module is used to build a fusion decision model, map the operating status data to the computing power assessment space, map the power supply parameters to the energy assessment space, and generate a fusion assessment vector representing the comprehensive status of each computing node in a unified multi-dimensional decision space. The scheduling module is used to solve a joint optimization problem that includes both computing power constraints and power constraints based on the fusion evaluation vector, and generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The sending module is used to convert the task allocation decision into a task scheduling instruction and send it to the computing node, and to convert the power supply adjustment decision into a power supply control instruction and send it to the intelligent power distribution unit, so as to realize the synchronous adjustment of the computing layer and the power layer.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the computing power and power coordinated scheduling method of any of the above-described data centers.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the computing power and power coordinated scheduling method for a data center as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a data center computing power and power coordinated scheduling method as described above.

[0016] The computing power and power coordinated scheduling method and apparatus for data centers provided by this invention constructs a unified data foundation by synchronously collecting the operating status of computing nodes and the power parameters of intelligent power distribution units. Based on this foundation, through fusion decision models and joint optimization, the originally independent scheduling problem is transformed into a holistic optimization problem encompassing both computing power and power constraints. The generated coordinated scheduling strategy can be synchronously converted into task allocation and power supply adjustment instructions and issued for execution, thereby achieving real-time and precise matching of computing power consumption and power supply at the physical level. This closed-loop coordinated control not only significantly reduces the power usage efficiency of data centers by eliminating power wastage, but also mitigates physical risks such as overcurrent and overheating through proactive power management, ultimately improving energy efficiency while significantly enhancing the operational reliability of data centers. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the data center computing power and power coordinated scheduling method provided by the present invention; Figure 2 This is a schematic diagram of the control logic flow provided by the present invention; Figure 3 A schematic diagram of the computing power and power coordinated scheduling device for a data center provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Figure 1 This is a flowchart illustrating the data center computing power and power coordinated scheduling method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 110: Periodically collect the operating status data of multiple computing nodes in the data center through the first data channel, and obtain the power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing node through the second data channel to achieve dual-channel data acquisition; In this application, the first data channel is used to periodically collect the operating status data of multiple computing nodes within the data center.

[0021] The computing node in this application can be a standalone physical server, a virtual machine, a container, or, at a more macroscopic level, a rack or cabinet consisting of one or more servers.

[0022] Operational status data are parameters that characterize the workload and performance of computing nodes, including but not limited to: CPU load rate, memory utilization, disk I / O rate, network throughput, and node temperature.

[0023] Data acquisition in the first data channel of this application can be achieved through standard management protocols, such as obtaining data from the server's baseboard management controller via IPMI. The acquisition cycle can be fixed, such as once per second; or it can be dynamically adjusted, such as increasing the acquisition frequency when the system load is high and decreasing the frequency when the load is low, to balance the real-time performance of monitoring and the system overhead.

[0024] In this application, the second data channel is used to obtain power supply parameters in real time from the Smart Power Distribution Unit (PDU) that physically corresponds to the computing node.

[0025] A smart power distribution unit (PDU) specifically refers to a device with network communication and remote control capabilities that directly provides power to computing nodes.

[0026] Physical mapping means that a unique mapping relationship can be established between the output port of each PDU and the computing node it powers.

[0027] Power supply parameters are data that characterize the physical state of a power system, including but not limited to: real-time current, output voltage, active power, apparent power, power factor, and the operating temperature of the PDU itself at each output port. This data is typically acquired through power communication protocols such as Modbus TCP or Simple Network Management Protocol (SMMP).

[0028] In this application, by using dual-channel acquisition, a complete data profile of a data center on both the computing power consumption side and the energy supply side can be obtained, providing a solid data foundation for subsequent collaborative decision-making.

[0029] Step 120: Construct a fusion decision model, map the operating status data to the computing power assessment space, map the power supply parameters to the energy assessment space, and generate a fusion assessment vector representing the comprehensive status of each computing node in a unified multidimensional decision space. In this application, data from different dimensions and with different physical meanings are uniformly quantitatively evaluated.

[0030] First, a fusion decision model is constructed to process and fuse the two types of heterogeneous data collected.

[0031] Specifically, the fusion decision model may contain two subspace mapping processes: mapping the operating status data to the computing power assessment space and mapping the power supply parameters to the energy assessment space.

[0032] For example, a computing load score can be calculated by weighted summation of CPU load rate and memory utilization rate; at the same time, an energy efficiency score can be calculated by parameters such as the ratio of real-time power consumption to rated power consumption and power factor.

[0033] Then, a fusion evaluation vector representing the overall state of each computing node is generated in a unified multidimensional decision space. This vector is a combination of the results from the two evaluation spaces mentioned above.

[0034] For example, a simple fusion evaluation vector could be V = [Hydropower Load Score, Energy Efficiency Score, Node Temperature]. In more complex implementations, this vector can include more dimensions and undergo normalization to eliminate the influence of dimensions. This fusion evaluation vector comprehensively reflects the health and cost-effectiveness of each computing node at the current moment.

[0035] Step 130: Based on the fusion evaluation vector, solve the joint optimization problem that simultaneously includes computing power constraints and power constraints, and generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. In this application, unlike traditional scheduling which only considers computing power, the joint optimization problem includes both computing power constraints and power constraints.

[0036] The goal of optimization problems can be singular, such as minimizing total energy consumption, or it can be complex, such as minimizing energy consumption while ensuring performance.

[0037] After solving this problem, a cooperative scheduling strategy is generated. This strategy is a composite strategy, which contains two interrelated decision sets: task allocation decisions and power supply adjustment decisions.

[0038] Task allocation decisions specify which computing node a new computing task should be assigned to, or whether an existing task needs to be migrated.

[0039] Power supply adjustment decisions specify what adjustments need to be made to the power supply of the relevant computing nodes in order to coordinate with the task allocation.

[0040] For example, if a high-load task is assigned to a node, the power supply adjustment decision might be to appropriately increase the power supply voltage of the corresponding PDU port of that node to ensure stable operation.

[0041] Step 140: The task allocation decision is converted into a task scheduling instruction and sent to the computing node; the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution unit, thereby realizing the synchronous adjustment of the computing layer and the power layer.

[0042] In this application, the task allocation decision is converted into a task scheduling instruction and sent to the compute node. This operation is similar to that of a traditional data center scheduler, such as scheduling the creation, migration, or destruction of virtual machines or containers through platforms like Kubernetes and OpenStack.

[0043] Simultaneously, the power supply adjustment decision is converted into a power supply control command and sent to the intelligent power distribution unit.

[0044] Power supply control commands can be in structured data formats (such as JSON or XML) and sent to the target smart PDU via a secure network channel (such as a TCP / IP connection based on TLS encryption). After parsing the commands, the smart PDU executes the corresponding operations through its internal relays or power electronic devices.

[0045] In this application, a unified data foundation is constructed by synchronously collecting the operating status of computing nodes and the power parameters of intelligent power distribution units. Based on this foundation, a fusion decision model and joint optimization transform the originally independent scheduling problem into a holistic optimization problem with dual constraints of computing power and power. The generated collaborative scheduling strategy can be synchronously translated into task allocation and power supply adjustment instructions and issued for execution, thereby achieving real-time and precise matching of computing power consumption and power supply at the physical level. This closed-loop collaborative control not only significantly reduces the power usage efficiency of data centers by eliminating power wastage, but also mitigates physical risks such as overcurrent and overheating through proactive power management, ultimately improving energy efficiency while significantly enhancing the operational reliability of data centers.

[0046] Optionally, solving the joint optimization problem includes: Construct a multi-objective optimization function that comprehensively considers minimizing total data center energy consumption, maximizing load balancing among nodes, minimizing task migration costs, and minimizing carbon emissions; Set multi-level constraints, including upper limits on the processing capacity of computing nodes, upper limits on the power output of intelligent power distribution units, quality of service constraints for critical services, and network bandwidth constraints. The solution is obtained by using a reinforcement learning algorithm. The fused evaluation vector is used as the environmental state input, and task allocation and power supply adjustment are used as executable actions. The optimal strategy is obtained through continuous interaction and learning between the agent and the environment.

[0047] In this application, a multi-objective optimization function can be constructed to comprehensively measure the merits of scheduling strategies. This function aims to achieve multiple, and sometimes even conflicting, operational objectives simultaneously.

[0048] Minimizing total energy consumption in a data center is one of the core optimization goals, directly impacting operating costs and green computing metrics. Total energy consumption can be obtained by summing the power consumption data reported by all intelligent power distribution units in real time.

[0049] Maximizing load balancing across nodes involves distributing tasks evenly across all computing nodes. This avoids situations where some nodes are overloaded while others are idle, thereby improving resource utilization and extending the average lifespan of the hardware. Load balancing can be measured by calculating the variance or standard deviation of the fused evaluation vector across all nodes; the smaller the variance, the higher the load balancing.

[0050] Minimize task migration costs. Task migration consumes network bandwidth and computing resources, and may cause temporary business interruptions or performance fluctuations. Therefore, migration should be minimized whenever possible. Migration costs can be quantified by factors such as the size of the data being migrated and the migration time.

[0051] In some advanced application scenarios, if a data center is connected to multiple energy sources and can obtain real-time energy carbon footprint information, the optimization function can also include the goal of minimizing carbon emissions, prioritizing the use of clean energy.

[0052] Secondly, in order to ensure the feasibility and security of scheduling decisions, it is necessary to set multi-level constraints, which are boundary conditions that must be satisfied when solving optimization problems.

[0053] The processing capacity of computing nodes is constrained. Each computing node has limited resources such as CPU and memory, and the total resource requirements of tasks allocated to that node cannot exceed its physical limit.

[0054] The power output limit of the intelligent power distribution unit is constrained. Each intelligent PDU and each of its output ports has a rated maximum power or maximum current. The actual power at any time must not exceed this safety threshold to prevent overload tripping or equipment damage.

[0055] Service Level Agreement (SLA / QoS) constraints for critical services: For high-priority critical services, it is essential to ensure that the computing resources they receive meet the requirements of the service level agreement, such as response latency and processing throughput.

[0056] Network bandwidth constraints are a factor; the execution and migration of tasks consume network bandwidth, and scheduling decisions must take into account the bandwidth limitations of the data center's internal network to avoid network congestion.

[0057] Finally, for the complex optimization problem constructed above, reinforcement learning (RL) algorithms can be used to solve it. Reinforcement learning is particularly suitable for solving dynamic optimization problems that require sequential decision-making. In specific implementation: The entire data center can be viewed as an environment. The fused evaluation vector set of all computing nodes at each sampling time, together with the current task queue state, constitutes the environment state input of the reinforcement learning model.

[0058] The set of executable actions consists of two parts: assigning a task to a compute node and adjusting the power supply parameters of a PDU port.

[0059] The design of the reward function is closely related to the multi-objective optimization function. When an action is performed, if it reduces the total energy consumption of the data center and makes the load more balanced, a positive reward is given; conversely, if it leads to SLA default or power overrun, a negative reward (penalty) is given.

[0060] Through continuous interaction and learning between the agent and its environment, reinforcement learning algorithms can gradually learn an optimal policy. This policy is essentially a mapping from the environmental state to the best action, and can automatically generate optimal cooperative scheduling decisions under various complex situations.

[0061] In this application, the scheduling problem can be transformed into a formalized, solvable mathematical model, and advanced algorithms such as reinforcement learning can be used to make intelligent decisions, thereby finding the optimal balance between multiple operational objectives and strict physical constraints, and realizing highly intelligent, precise and adaptive computing power and power coordinated scheduling.

[0062] Optionally, the method further includes: Based on historically collected operational status data and power supply parameters, a time-series prediction model is trained. The time-series prediction model adopts a deep learning network structure, which includes a recurrent neural network layer for extracting time-series features and an attention mechanism layer for identifying key factors. Using the trained time-series prediction model, the load change trend and power consumption evolution curve of each computing node are predicted within a future time window; When the prediction results show that the power consumption of a certain computing node will approach the safety threshold in the future, the preventive task migration should be given priority when generating the cooperative scheduling strategy, and some tasks should be allocated to other nodes in advance.

[0063] In this application, historical data refers to a time-stamped data sequence accumulated over a long period through the first and second data channels during system operation. This data contains patterns and regularities in the load and power consumption changes of computing nodes. For example, some services exhibit periodic load peaks during the day, or the power consumption of some servers slowly increases with rising ambient temperature.

[0064] The time series prediction model employs a deep learning network architecture that can effectively capture time series dependencies. For example, the model might include: One or more layers of a recurrent neural network (RNN), such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU). These network structures have memory capabilities and are well-suited for processing time series data, enabling them to learn the dependencies and trends of data over time.

[0065] An attention mechanism layer can be added above the RNN layer. This mechanism allows the model to automatically assign different weights to different time steps in the input sequence when making predictions, thereby paying more attention to key historical moments that have the greatest impact on future predictions, such as the starting point of the previous load spike, thus improving prediction accuracy.

[0066] Secondly, the trained time-series prediction model is used for forward-looking predictions. During system operation, the most recent time window, such as the real data from the past hour, is input into the trained time-series prediction model.

[0067] The model outputs predictions of load change trends and power consumption evolution curves for each computing node within a future time window, such as the next 15 minutes. This means the system not only knows the current situation but can also anticipate potential changes within a short period of time.

[0068] Finally, preventative scheduling is performed based on the prediction results. The system continuously monitors the prediction results. For example, when the prediction results show that the power consumption of a computing node will approach its safe threshold (such as 95% of the rated power of the PDU port) in the next 5 minutes due to the upcoming start of a large computing task, the system identifies a potential overload risk.

[0069] In this scenario, the system prioritizes preventative task migration when generating collaborative scheduling strategies. Instead of waiting for an actual overload to occur, it proactively allocates or migrates one or more minor tasks on a given node, or new tasks that are about to be assigned to that node, to other nodes with lower current and predicted future loads and sufficient power redundancy before the risk materializes.

[0070] In this application, by introducing the aforementioned prediction and prevention mechanisms, potential power risks and performance bottlenecks can be identified and avoided in advance, eliminating many accidents that may lead to service interruptions or hardware damage in their infancy. This greatly enhances the continuity, security, and robustness of data center operations, achieving a qualitative leap from passive response to proactive management.

[0071] Optionally, the power supply control command is specifically used for: Independent control is implemented at the port level, and the output ports of each intelligent power distribution unit can independently adjust power supply parameters, including voltage amplitude adjustment, current limit setting, and power supply on / off control. The power supply mode is matched according to the task type. Stable high power supply is provided for compute-intensive tasks, dynamic power adjustment is adopted for storage-intensive tasks, and low power maintenance or complete power cut-off is performed for standby devices.

[0072] In this application, power supply control commands are used for independent control at the port level. Therefore, an intelligent power distribution unit with multiple output ports is no longer controlled coarsely as a whole, but each of its output ports can have its power supply parameters adjusted independently and precisely.

[0073] More specifically, voltage amplitude adjustment commands can instruct the PDU to fine-tune the output voltage of a specific port within a safe range (e.g., 220V ± 5%). Appropriate boosting can provide more stable power support for high-performance computing tasks, while appropriate bucking can save energy under lower loads.

[0074] The current limit setting command allows you to dynamically set a software-defined current limit for each port, serving as a second layer of protection in addition to physical fuses.

[0075] Power supply on / off control commands are the most basic controls, which can precisely turn the power supply to any port on or off, thereby completely eliminating the power consumption of idle servers.

[0076] In this application, based on this port-level fine-grained control capability, the system can implement a strategy of matching the power supply mode according to the task type.

[0077] Providing stable, high-power supply for computationally intensive tasks, such as a server performing large-scale scientific computing or training an artificial intelligence model, where the CPU and GPU are under high load for extended periods, is crucial. In this case, the cooperative scheduling strategy generates instructions requiring the corresponding PDU port to provide a relatively high and very stable voltage (e.g., 230V) and sets a high power limit to ensure that the computing process is not interrupted or malfunctions due to voltage fluctuations or power limitations.

[0078] For storage or I / O intensive tasks, dynamic power regulation can be employed. For example, a server primarily providing database or file services typically experiences bursty loads, with potentially long idle periods between I / O operations. For such tasks, the power supply mode can be configured for dynamic adjustment. During peak I / O periods, the PDU provides normal or slightly higher power; during idle periods, it can work in conjunction with the server's power management unit to dynamically reduce the supply voltage or frequency, entering an energy-saving mode.

[0079] For standby devices, the scheduling policy executes low-power maintenance or complete power-off. For servers that have not been assigned any tasks for a period of time, the scheduling policy generates instructions to switch their power supply mode to deep sleep or complete power-off. For scenarios requiring rapid wake-up, a low-power maintenance mode can be used, for example, maintaining power only to the BMC; for scenarios that can tolerate longer startup times, the main power supply can be directly cut off to maximize energy savings.

[0080] In this application, the granularity of power control is reduced from the rack level to the server-level PDU port, and intelligent linkage between power supply strategy and upper-layer service types is achieved. This refined and scenario-based power supply mode matching enables precise on-demand power allocation, avoiding significant energy waste and further optimizing the data center's PUE without affecting service performance.

[0081] Optionally, the power supply control command includes at least one of the following control operations: turning on or off a designated output port of the intelligent power distribution unit; adjusting the power supply voltage value of the designated output port; setting the output power limit threshold of the designated output port; and switching the power supply mode of the intelligent power distribution unit.

[0082] In this application, opening or closing a designated output port of the intelligent power distribution unit is the most direct power control operation. The instruction will include the identifier of the target intelligent PDU (such as an IP address) and the specific output port number. When the instruction is "open," the relay inside the intelligent PDU will close, supplying power to the port; when the instruction is "close," the relay will open, cutting off the power to the port.

[0083] Adjusting the supply voltage of the specified output port requires the intelligent PDU to have voltage regulation capabilities (usually achieved through a built-in AC-DC adjustable module or similar power electronic device). The instruction will include the target port number and a specific voltage value (e.g., 225V). Upon receiving the instruction, the PDU will adjust its internal circuitry to stabilize the output voltage of that port at the target value.

[0084] Setting the output power limit threshold for the specified output port is a software-level setting of a dynamic fuse threshold for each port. The instruction includes the target port number and a power value (e.g., 250W). The intelligent PDU continuously monitors the real-time output power of the port, and once the set threshold is exceeded, the PDU can perform preset actions, such as issuing an alarm or limiting the power.

[0085] Switching the power supply mode of the intelligent power distribution unit allows the dispatching system to dynamically switch the operating mode of the PDU according to the overall strategy. Specifically, some advanced intelligent PDUs may have multiple preset operating modes, such as high-efficiency mode, performance mode, energy-saving mode, etc., and this operation allows them to switch between these modes.

[0086] This invention provides a powerful and flexible physical execution interface for complex upper-layer collaborative scheduling algorithms. These refined control methods enable the coordination of computing power and electricity to move beyond the theoretical level and be executed accurately and reliably, ensuring the feasibility and practicality of the entire technical solution and precisely transmitting the intent of the scheduling strategy to the power physical layer.

[0087] Optionally, the method further includes: The intelligent power distribution unit continuously monitors its own power parameters, analyzes parameter change characteristics in real time through an abnormal pattern recognition algorithm, and actively generates and reports alarm signals when an abnormal pattern is detected. A tiered response is implemented based on the severity of the anomaly. For mild anomalies, parameters are adaptively adjusted. For moderate anomalies, preventative migration of related tasks is triggered. For severe anomalies, fault isolation and emergency switching are implemented immediately. Record the exception handling process and results, and add exception patterns and their handling strategies to the knowledge base to enhance the system's ability to identify and handle similar exceptions.

[0088] In this application, an anomaly pattern recognition algorithm is deployed in the PDU's firmware. This algorithm can range from simple to complex: Simple implementation: Based on static threshold comparison. For example, if the current exceeds 120% of the rated value for 3 consecutive seconds, or the temperature is higher than 85°C, it is judged as abnormal.

[0089] Complex implementation: Based on a dynamic baseline and machine learning model using historical data. The algorithm learns the normal parameter fluctuation range for each port under different loads. When real-time data deviates from this dynamic baseline to a certain extent, such as the appearance of atypical spikes or continuous abnormal jitter, it will be identified as an abnormal pattern even if it does not reach a hard threshold.

[0090] Once an abnormal pattern is detected, the intelligent PDU will proactively generate an alarm signal and report it to the central dispatch system, instead of waiting for the dispatch system to poll it. Reporting can be achieved through hardware interrupts or high-priority network messages, ensuring low latency for alarms.

[0091] Secondly, the system executes tiered responses based on the severity of the anomaly. This differentiated approach can efficiently and appropriately address risks at different levels.

[0092] For minor anomalies (such as transient small voltage fluctuations or slightly elevated temperatures), the intelligent PDU can first attempt adaptive parameter adjustments. For example, if the temperature is slightly high, the PDU can work with the server's BMC to slightly increase the fan speed. This localized, rapid self-healing can resolve most minor issues without needing to report to the central scheduler for complex task migration.

[0093] For moderate anomalies (e.g., power consistently approaching but not exceeding a threshold, or temperature showing a slow but clear upward trend), after a PDU reports an alarm, the central scheduling system will mark it as sub-healthy and trigger a preventative migration of related tasks. The system will smoothly migrate non-critical tasks or tasks sensitive to performance fluctuations on that node to healthy nodes to relieve the load on that node and restore it to normal.

[0094] For severe anomalies (e.g., a short circuit causing a surge in current, or a rapid temperature rise triggering over-temperature protection), the PDU may execute local emergency actions, such as immediately cutting off the power to the port, while generating the highest-level alarm. Upon receiving the alarm, the central dispatch system will immediately perform fault isolation and emergency switchover, migrating all tasks on the node (especially high-availability services) to the backup node as quickly as possible, marking it as faulty, prohibiting the assignment of new tasks, and notifying operations and maintenance personnel to intervene.

[0095] Finally, the system records the exception handling process and results, and adds them to the knowledge base for continuous learning and optimization.

[0096] Every occurrence, identification, response, and handling result of an anomaly is recorded in detail, forming an anomaly event log.

[0097] This log data is used to improve the system, particularly its ability to identify and handle similar anomalies. For example, by analyzing repeated occurrences of the same anomaly, the system can optimize the parameters of the anomaly pattern recognition algorithm, or add penalty weights for such anomaly precursors to the reinforcement learning model, enabling future scheduling strategies to proactively avoid scheduling combinations that may lead to the anomaly.

[0098] This application's embodiments construct a complete self-healing closed loop, from monitoring, identification, tiered response to learning and optimization. This enables data centers to no longer be vulnerable and passive in the face of power outages, but rather to possess strong resilience and intelligence, capable of quickly and automatically handling various anomalies and maximizing business continuity and data security.

[0099] Optionally, the method specifically includes: The data center is divided into multiple autonomous management domains. Each autonomous management domain contains several computing nodes and their corresponding intelligent power distribution units, and is configured with an independent domain controller. Each autonomous management domain independently performs dual-channel data acquisition, builds a fusion decision model, solves joint optimization problems, and performs bidirectional linkage control to achieve local optimization scheduling within the domain. When the resources of a single autonomous domain are insufficient, resource support is requested from neighboring domains through an inter-domain negotiation mechanism to achieve cross-domain task migration and load balancing. A distributed ledger is used to record all cross-domain scheduling decisions and execution results.

[0100] In this application, an autonomous management domain is a logically or physically independent management unit. For example, a row of server racks in a data center, a specific business cluster, or a containerized data center can all be defined as an autonomous management domain.

[0101] More specifically, each autonomous management domain contains several computing nodes and their corresponding physical intelligent power distribution units, forming a complete computing-power subsystem. Each domain is configured with one or a group of independent domain controllers.

[0102] Secondly, each autonomous domain independently executes local optimization scheduling within its domain.

[0103] In this application, each domain controller independently performs dual-channel data acquisition, constructs a fusion decision model, solves joint optimization problems, and issues bidirectional linkage control commands within its jurisdiction.

[0104] Most scheduling decisions and control loops are completed within the autonomous management domain, which greatly reduces the communication pressure and computational load on the central node, and improves the system's response speed and robustness. Even if the central node or cross-domain network fails, each autonomous management domain can still operate independently, achieving local optimized scheduling within the domain and ensuring basic service capabilities.

[0105] When cross-domain resource requirements arise, global collaboration is achieved through inter-domain negotiation mechanisms. When the resources of a single autonomous domain are insufficient, for example, if domain A receives a large computing task that exceeds the total capacity of all its nodes, or if a large-scale failure occurs within domain A, causing a sharp reduction in available resources, its domain controller will not directly reject the task, but will initiate an inter-domain negotiation mechanism.

[0106] It requests resource support from one or more physically or logically adjacent domain controllers. The request message will contain the required resource specifications (such as the number of CPU cores, memory size, estimated power consumption, etc.).

[0107] Upon receiving the request, the domain controller, such as domain B, will assess its own resource sufficiency. If it can meet the requirements, it will respond to the request and reserve the corresponding resources. Subsequently, the controllers of the two domains will collaborate to complete cross-domain task migration and load balancing. For example, tasks originally assigned to domain A may be directly scheduled to run on nodes in domain B.

[0108] Finally, to ensure the traceability and fairness of cross-domain scheduling, a distributed ledger can be used to record all cross-domain scheduling decisions and execution results.

[0109] Distributed ledger technologies, such as blockchain, can be used to create an immutable, shared record of transactions. Whenever a cross-domain resource call occurs (e.g., domain A uses resources from domain B), the relevant decision information is recorded as a transaction on the ledger. This ledger can then be used for subsequent billing, auditing, and resource accounting, ensuring fairness, transparency, and trustworthiness in resource usage within a multi-domain, multi-tenant environment.

[0110] In this application, by introducing this layered, distributed autonomous system architecture, a large and complex centralized management problem is decomposed into several relatively simple local optimization problems and a clear inter-domain coordination problem. This can effectively support the smooth expansion from a single rack to a hyperscale data center with hundreds of thousands of servers, while ensuring high availability and management efficiency of the system.

[0111] In one alternative embodiment, Figure 2 This is a schematic diagram of the control logic flow provided by the present invention, such as... Figure 2 As shown, it includes: First, the scheduler performs real-time monitoring. The scheduler continuously collects the operating status data of the computing nodes and simultaneously obtains the power supply parameters of the intelligent power distribution units that supply them.

[0112] The scheduler then performs a step to calculate the composite weight. The composite weight here is a quantitative indicator characterizing the overall state of a computing node; it is derived by fusing the node's computing layer operational data and physical layer power data. A lower composite weight typically indicates that the node is currently under lighter load and has higher energy efficiency, making it an ideal choice for undertaking new tasks.

[0113] Based on the composite weights of all nodes, the scheduler generates a task allocation strategy and initially selects one or more target nodes to carry out new computing tasks.

[0114] Before officially issuing a task, the scheduler first performs a crucial preparatory step: sending a power supply command to the smart distribution unit. This command is generated based on the expected load of the task to be assigned. For example, if the new task is a high-load type, the scheduler sends a boost command to the smart distribution unit port corresponding to the target node.

[0115] Upon receiving instructions, the intelligent power distribution unit performs power supply adjustments. This physical-level adjustment creates conditions for the central processing unit of the computing node to adjust its operating voltage and frequency, enabling it to enter a performance state that matches the expected task.

[0116] After the power supply environment is adjusted, the compute nodes report their latest performance status to the scheduler. Only after confirming that the physical performance of the nodes is ready does the scheduler finally execute the step of adjusting the compute node's tasks, that is, instantiating the compute tasks onto the node. This order of adjusting power first and then loading tasks ensures a precise match between tasks and resources.

[0117] Once the task begins, the system enters an iterative optimization loop. The scheduler continuously receives real-time status feedback from the computing nodes and intelligent power distribution units, and determines whether the current state has achieved the preset optimization task. This optimization task can be a specific objective, such as stabilizing the node's power consumption below a certain value.

[0118] If the judgment result is negative, it means that the current state has not yet reached the optimal state. The process will jump back to the step of calculating the composite weight, and use the latest feedback data to perform a new round of fine-tuning, forming a rapid local optimization loop. If the judgment result is positive, it means that the current state has met the optimization objective. In this case, the scheduling event is successful, the process ends, and the next cycle begins.

[0119] Through the iterative closed-loop control method described in this embodiment, the present invention achieves a more prudent collaborative scheduling mechanism. It continuously fine-tunes computing resources and power supply throughout the entire lifecycle of task execution via a trial-and-error, feedback-and-ready-adjustment process. This enables more precise matching of resources and load, avoiding resource waste or performance bottlenecks caused by inaccurate forecasts, thereby significantly improving the overall operational efficiency and stability of the data center.

[0120] The computing power and power coordinated scheduling device for data centers provided by the present invention will be described below. The computing power and power coordinated scheduling device for data centers described below can be referred to in correspondence with the computing power and power coordinated scheduling method for data centers described above.

[0121] Figure 3 This is a schematic diagram of the computing power and power coordinated scheduling device for a data center provided by the present invention, as shown below. Figure 3 As shown, it includes: The acquisition module 310 is used to periodically acquire the operating status data of multiple computing nodes in the data center through the first data channel, and to obtain the power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing node through the second data channel, so as to realize dual-channel data acquisition. The construction module 320 is used to construct a fusion decision model, map the operating status data to the computing power assessment space, map the power supply parameters to the energy assessment space, and generate a fusion assessment vector representing the comprehensive status of each computing node in a unified multi-dimensional decision space. The scheduling module 330 is used to solve a joint optimization problem that includes both computing power constraints and power constraints based on the fusion evaluation vector, and generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The sending module 340 is used to convert the task allocation decision into a task scheduling instruction and send it to the computing node, and to convert the power supply adjustment decision into a power supply control instruction and send it to the intelligent power distribution unit, so as to realize the synchronous adjustment of the computing layer and the power layer.

[0122] In this application, a unified data foundation is constructed by synchronously collecting the operating status of computing nodes and the power parameters of intelligent power distribution units. Based on this foundation, a fusion decision model and joint optimization transform the originally independent scheduling problem into a holistic optimization problem with dual constraints of computing power and power. The generated collaborative scheduling strategy can be synchronously translated into task allocation and power supply adjustment instructions and issued for execution, thereby achieving real-time and precise matching of computing power consumption and power supply at the physical level. This closed-loop collaborative control not only significantly reduces the power usage efficiency of data centers by eliminating power wastage, but also mitigates physical risks such as overcurrent and overheating through proactive power management, ultimately improving energy efficiency while significantly enhancing the operational reliability of data centers.

[0123] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a data center computing power and power coordinated scheduling method. This method includes: periodically collecting operating status data of multiple computing nodes in the data center through a first data channel, and obtaining power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing nodes through a second data channel to achieve dual-channel data acquisition; A fusion decision model is constructed, which maps the operating status data to the computing power assessment space and the power supply parameters to the energy assessment space. A fusion assessment vector representing the comprehensive status of each computing node is generated in a unified multidimensional decision space. Based on the fusion evaluation vector, a joint optimization problem that simultaneously includes computing power constraints and power constraints is solved to generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The task allocation decision is converted into a task scheduling instruction and sent to the computing node, and the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution unit, thereby realizing the synchronous adjustment of the computing layer and the power layer.

[0124] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute the data center computing power and power coordinated scheduling method provided by the above methods, the method including: periodically collecting the operating status data of multiple computing nodes in the data center through a first data channel, and obtaining power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing node through a second data channel, so as to realize dual-channel data acquisition; A fusion decision model is constructed, which maps the operating status data to the computing power assessment space and the power supply parameters to the energy assessment space. A fusion assessment vector representing the comprehensive status of each computing node is generated in a unified multidimensional decision space. Based on the fusion evaluation vector, a joint optimization problem that simultaneously includes computing power constraints and power constraints is solved to generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The task allocation decision is converted into a task scheduling instruction and sent to the computing node, and the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution unit, thereby realizing the synchronous adjustment of the computing layer and the power layer.

[0126] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a data center computing power and power coordinated scheduling method provided by the above methods. The method includes: periodically collecting operating status data of multiple computing nodes in the data center through a first data channel, and obtaining power supply parameters in real time from an intelligent power distribution unit physically corresponding to the computing nodes through a second data channel to achieve dual-channel data acquisition. A fusion decision model is constructed, which maps the operating status data to the computing power assessment space and the power supply parameters to the energy assessment space. A fusion assessment vector representing the comprehensive status of each computing node is generated in a unified multidimensional decision space. Based on the fusion evaluation vector, a joint optimization problem that simultaneously includes computing power constraints and power constraints is solved to generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The task allocation decision is converted into a task scheduling instruction and sent to the computing node, and the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution unit, thereby realizing the synchronous adjustment of the computing layer and the power layer.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for coordinated scheduling of computing power and power in a data center, characterized in that, include: The system periodically collects operational status data from multiple computing nodes within the data center via the first data channel, and obtains power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing nodes via the second data channel, thereby achieving dual-channel data acquisition. A fusion decision model is constructed, which maps the operating status data to the computing power assessment space and the power supply parameters to the energy assessment space. A fusion assessment vector representing the comprehensive status of each computing node is generated in a unified multidimensional decision space. Based on the fusion evaluation vector, a joint optimization problem that simultaneously includes computing power constraints and power constraints is solved to generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The task allocation decision is converted into a task scheduling instruction and sent to the computing node, and the power supply adjustment decision is converted into a power supply control instruction and sent to the intelligent power distribution unit, thereby realizing the synchronous adjustment of the computing layer and the power layer.

2. The data center computing power and power coordinated scheduling method according to claim 1, characterized in that, Solving the joint optimization problem includes: Construct a multi-objective optimization function that comprehensively considers minimizing total data center energy consumption, maximizing load balancing among nodes, minimizing task migration costs, and minimizing carbon emissions; Set multi-level constraints, including upper limits on the processing capacity of computing nodes, upper limits on the power output of intelligent power distribution units, quality of service constraints for critical services, and network bandwidth constraints. The solution is obtained by using a reinforcement learning algorithm. The fused evaluation vector is used as the environmental state input, and task allocation and power supply adjustment are used as executable actions. The optimal strategy is obtained through continuous interaction and learning between the agent and the environment.

3. The data center computing power and power coordinated scheduling method according to claim 1, characterized in that, The method further includes: Based on historically collected operational status data and power supply parameters, a time-series prediction model is trained. The time-series prediction model adopts a deep learning network structure, which includes a recurrent neural network layer for extracting time-series features and an attention mechanism layer for identifying key factors. Using the trained time-series prediction model, the load change trend and power consumption evolution curve of each computing node are predicted within a future time window; When the prediction results show that the power consumption of a certain computing node will approach the safety threshold in the future, the preventive task migration should be given priority when generating the cooperative scheduling strategy, and some tasks should be allocated to other nodes in advance.

4. The data center computing power and power coordinated scheduling method according to claim 1, characterized in that, The power supply control command is specifically used for: Independent control is implemented at the port level, and the output ports of each intelligent power distribution unit can independently adjust power supply parameters, including voltage amplitude adjustment, current limit setting, and power supply on / off control. The power supply mode is matched according to the task type. Stable high power supply is provided for compute-intensive tasks, dynamic power adjustment is adopted for storage-intensive tasks, and low power maintenance or complete power cut-off is performed for standby devices.

5. The data center computing power and power coordinated scheduling method according to claim 1, characterized in that, The power supply control command includes at least one of the following control operations: turning on or off a designated output port of the intelligent power distribution unit; adjusting the power supply voltage value of the designated output port; Set the output power limit threshold of the specified output port; switch the power supply mode of the intelligent power distribution unit.

6. The data center computing power and power coordinated scheduling method according to claim 1, characterized in that, The method further includes: The intelligent power distribution unit continuously monitors its own power parameters, analyzes parameter change characteristics in real time through an abnormal pattern recognition algorithm, and actively generates and reports alarm signals when an abnormal pattern is detected. A tiered response is implemented based on the severity of the anomaly. For mild anomalies, parameters are adaptively adjusted. For moderate anomalies, preventative migration of related tasks is triggered. For severe anomalies, fault isolation and emergency switching are implemented immediately. Record the exception handling process and results, and add exception patterns and their handling strategies to the knowledge base to enhance the system's ability to identify and handle similar exceptions.

7. The data center computing power and power coordinated scheduling method according to claim 1, characterized in that, The method specifically includes: The data center is divided into multiple autonomous management domains. Each autonomous management domain contains several computing nodes and their corresponding intelligent power distribution units, and is configured with an independent domain controller. Each autonomous management domain independently performs dual-channel data acquisition, builds a fusion decision model, solves joint optimization problems, and performs bidirectional linkage control to achieve local optimization scheduling within the domain. When the resources of a single autonomous domain are insufficient, resource support is requested from neighboring domains through an inter-domain negotiation mechanism to achieve cross-domain task migration and load balancing. A distributed ledger is used to record all cross-domain scheduling decisions and execution results.

8. A data center computing power and power coordinated scheduling device, characterized in that, include: The acquisition module is used to periodically acquire the operating status data of multiple computing nodes in the data center through the first data channel, and to obtain the power supply parameters in real time from the intelligent power distribution unit physically corresponding to the computing node through the second data channel, so as to realize dual-channel data acquisition. The construction module is used to build a fusion decision model, map the operating status data to the computing power assessment space, map the power supply parameters to the energy assessment space, and generate a fusion assessment vector representing the comprehensive status of each computing node in a unified multi-dimensional decision space. The scheduling module is used to solve a joint optimization problem that includes both computing power constraints and power constraints based on the fusion evaluation vector, and generate a collaborative scheduling strategy. The collaborative scheduling strategy includes two interrelated decision sets: task allocation decision and power supply adjustment decision. The sending module is used to convert the task allocation decision into a task scheduling instruction and send it to the computing node, and to convert the power supply adjustment decision into a power supply control instruction and send it to the intelligent power distribution unit, so as to realize the synchronous adjustment of the computing layer and the power layer.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the data center computing power and power coordinated scheduling method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data center computing power and power coordinated scheduling method as described in any one of claims 1 to 7.

Images