Method and device for scheduling control of computing and power coordination reserved resources, terminal and storage medium

Abstract

The application discloses a kind of algorithm electricity collaborative reservation resource scheduling control method, device, terminal and storage medium, it is related to algorithm electricity collaborative scheduling technical field, the scheme is based on task image, vGPU resource state and future electricity carbon window information constructs time slot vGPU reservation matrix to determine the reservation of vGPU resource in future different time slice, then according to the SLA latest start time of target task, the window start time of target electricity carbon window and the bindable state time of reserved vGPU resource with the earliest time mode to determine the recovery deadline reference time, again in combination vGPU recovery cost time to determine the latest recovery time, simultaneously, the expected recovery time consumption is calculated for resource borrowing request task, to ensure that high priority task can be started in time under the premise, allow low priority task to safely borrow reserved resource, avoid resource idling waste, while ensuring the timeliness of resource recovery to maintain service level agreement commitment and low carbon window execution effect.

Description

Technical Field

[0001] This application relates to the field of computing-powered collaborative scheduling technology, and in particular to a computing-powered collaborative reserved resource scheduling control method, device, terminal and storage medium. Background Technology

[0002] As tasks such as large-scale model training, inference, video inspection, fault diagnosis, power flow calculation, and simulation analysis are centrally deployed in intelligent computing centers, acceleration resources such as GPUs are gradually becoming key resources affecting task response time, cluster throughput, energy costs, and carbon emission intensity. Currently, intelligent computing centers typically allocate GPUs or vGPUs through Kubernetes, Slurm, container platforms, or cloud resource management platforms, and combine priority, queue, quota, preemption, resource reservation, or elastic scaling mechanisms to meet the operational needs of different tasks.

[0003] In computing-powered collaborative application scenarios where electricity prices, green electricity ratios, and carbon emission intensity change over time, high-priority tasks or tasks suitable for execution within the target low-carbon window typically require pre-reserving GPU or vGPU resources in future time slots to ensure timely startup while meeting Service Level Agreements (SLAs). However, reserved GPU or vGPU resources are prone to being idle before the target task arrives or the target low-carbon window begins. Completely prohibiting other tasks from using these resources would result in resource waste; simply allowing low-priority tasks to borrow them might not allow timely recovery before high-priority tasks arrive or the low-carbon window begins. Furthermore, since existing reserved pools or idle resource borrowing mechanisms typically only determine whether resources are currently idle, in practical application scenarios, the recovery of borrowed resources is often delayed, severely impacting the timeliness of target task execution. Summary of the Invention

[0004] This application provides a computing-powered collaborative reserved resource scheduling and control method, device, terminal, and storage medium to achieve accurate prediction of resource recovery time windows, realize efficient reuse of reserved resources under the constraints of high-priority task service level agreements, effectively avoid resource idleness and waste, and ensure the timeliness of resource recovery.

[0005] To achieve the aforementioned objectives, the first aspect of this application provides a computer-computer collaborative reserved resource scheduling and control method, comprising:

[0006] Obtain task profiles for each task to be scheduled, as well as the vGPU resource status and future carbon dioxide window information of the intelligent computing center;

[0007] A time-slotted vGPU reservation matrix is ​​generated based on the task profile, the vGPU resource status, and the future carbon window information. The time-slotted vGPU reservation matrix is ​​used to record the reserved vGPU resources and their reservation status in several future time slices, as well as the bindable state time of the reserved vGPU resources. The bindable state time is the time when the reserved vGPU resources should enter the bindable state.

[0008] For any future target time slice, extract the target task information, target carbon window information, and target bindable state time corresponding to the target time slice. Use the earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time as the recycling cutoff benchmark time. Then, after deducting the vGPU recycling cost time based on the recycling cutoff benchmark time, obtain the latest recycling time corresponding to the target time slice.

[0009] Based on the task information of the resource borrowing request task, the estimated recovery time of the resource borrowing request task is calculated using a preset estimated recovery time calculation formula. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow the reserved vGPU resources within the target time slice.

[0010] When the estimated reclamation time is less than or equal to the safety margin time, a borrowing permission instruction is generated to execute the resource borrowing request task using the reserved vGPU resources before the resource reclamation condition is triggered, wherein the safety margin time is the difference between the latest reclamation time and the current time.

[0011] Preferably, the task profile includes: task priority, task service level agreement (SLA), predicted runtime, vGPU resource requirements, video memory requirements, deferral permission flag, scaling down permission flag, checkpoint support flag, migration permission flag, and rated release time.

[0012] The vGPU resource status includes: vGPU profile, video memory status, computing power share, physical GPU topology, and slice status;

[0013] The future carbon window information includes: electricity price, green electricity ratio, carbon emission intensity, and start time of the future time slice.

[0014] Preferably, the vGPU reclamation cost time includes one or more of the following: vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and security buffer time.

[0015] Preferably, the formula for calculating the estimated recovery time is:

[0016] ;

[0017] In the formula, T_reclaim_est represents the estimated reclaim time, T_scale_down represents the task's scaling down or de-rating time, T_checkpoint represents the checkpoint time of the current resource borrowing request task, T_migration_release represents the task's migration or release time, T_rebind represents the vGPU rebinding time, and T_verify represents the status verification time.

[0018] Preferably, the step of generating the borrowing permission instruction to execute the resource borrowing request task using the reserved vGPU resources before the resource reclamation condition is triggered further includes:

[0019] When the resource reclamation condition is triggered, a reclamation action sequence is executed according to the latest reclamation time to reclaim the vGPU resources borrowed by the resource borrowing request task;

[0020] After the recycling action sequence is completed, a status check is performed on the recycled reserved vGPU resources, and the time-slotted vGPU reservation matrix is ​​updated according to the status check result.

[0021] Preferably, the resource recycling conditions include: the target task has arrived, the target electrocarbon window is about to begin, the current time has reached the latest recycling time, the safety margin time is lower than the preset margin time threshold, the vGPU profile status is abnormal, the video memory status is abnormal, the physical GPU topology status is abnormal, or the slice status is abnormal.

[0022] Preferably, it further includes:

[0023] When the reserved vGPU resource reclamation fails or the status verification result fails, the affected time slice, reserved vGPU resources and / or target task are determined, and the affected time slice, reserved vGPU resources and / or target task are locally recalculated, and the time-slotted vGPU reservation matrix is ​​updated according to the recalculation result.

[0024] The second aspect of this application provides a computer-computer collaborative reserved resource scheduling and control device, comprising:

[0025] The basic information acquisition unit is used to acquire the task profiles of each task to be scheduled, as well as the vGPU resource status and future carbon window information of the intelligent computing center;

[0026] A time-slotted vGPU reservation determination unit is used to generate a time-slotted vGPU reservation matrix based on the task profile, the vGPU resource status, and the future carbon window information. The time-slotted vGPU reservation matrix is ​​used to record the reserved vGPU resources and their reservation status in several future time slices, as well as the bindable state time of the reserved vGPU resources. The bindable state time is the time when the reserved vGPU resources should enter the bindable state.

[0027] The latest recycling time determination unit is used to extract the target task information, target carbon window information, and target bindable state time corresponding to any future target time slice for any future target time slice. The earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time is used as the recycling cutoff benchmark time. Then, the latest recycling time corresponding to the target time slice is obtained by deducting the vGPU recycling cost time based on the recycling cutoff benchmark time.

[0028] The resource borrowing and recycling time calculation unit is used to calculate the estimated recycling time of the resource borrowing request task based on the task information of the resource borrowing request task and through a preset estimated recycling time calculation formula. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow the reserved vGPU resources in the target time slice.

[0029] The resource borrowing determination unit is used to generate a borrowing permission instruction when the estimated recycling time is less than or equal to the safety margin time, so as to execute the resource borrowing request task using the reserved vGPU resources before the resource recycling condition is triggered, wherein the safety margin time is the difference between the latest recycling time and the current time.

[0030] A third aspect of this application provides a computing-powered collaborative reserved resource scheduling and control terminal, comprising: a memory and a processor;

[0031] The memory is used to store program code, which corresponds to the computing and power-coordinated reserved resource scheduling and control method provided in the first aspect of this application.

[0032] The processor is used to read and execute the program code to implement the computing and power collaborative reserved resource scheduling and control method.

[0033] The fourth aspect of this application provides a computer-readable storage medium storing program code, which is read and executed by a processor to implement the computer-computer collaborative reserved resource scheduling and control method provided in the first aspect of this application.

[0034] As can be seen from the above technical solutions, this application has the following advantages:

[0035] The solution provided in this application constructs a time-slotted vGPU reservation matrix based on the task profile of the task to be scheduled, the vGPU resource status of the intelligent computing center, and future carbon emission window information to determine the reservation status of in-slice vGPU resources at different future time periods. Then, it determines the recycling deadline by taking the earliest time of the latest start time of the target task's SLA, the window start time of the target carbon emission window, and the bindable state time of the reserved vGPU resources. Combined with the vGPU recycling cost time, it determines the latest recycling time. At the same time, it calculates the estimated recycling time for resource borrowing request tasks. Thus, while ensuring that high-priority tasks can start in a timely manner, it allows low-priority tasks to safely borrow reserved resources. Under the constraints of the service level agreement for high-priority tasks, it realizes the dynamic borrowing and timely recycling of reserved resources, thereby accurately predicting the resource recycling time window and achieving efficient reuse of reserved resources under the constraints of the service level agreement for high-priority tasks. This effectively avoids resource idleness and waste, while ensuring the timeliness of resource recycling to maintain the service level agreement commitment and the low-carbon window execution effect. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating an embodiment of a computing-powered collaborative reserved resource scheduling and control method provided in this application.

[0038] Figure 2 This is a timing diagram illustrating how the reserved vGPU resources in the method provided in this application are borrowed by low-priority tasks and reclaimed before the target carbon window begins.

[0039] Figure 3 This is an overall flowchart of an embodiment of a computing and power-coordinated reserved resource scheduling and control method provided in this application.

[0040] Figure 4 This is a schematic diagram of the architecture of an embodiment of a computing and power collaborative reserved resource scheduling and control device provided in this application. Detailed Implementation

[0041] In computing-powered collaborative application scenarios, intelligent computing centers reserve vGPU resources in time slots through resource management platforms to meet the service level agreement requirements of high-priority tasks or low-carbon window tasks. However, the existing mechanism only judges the feasibility of borrowing based on the current idle status of resources and cannot dynamically quantify the technical constraints in the resource recycling process. As a result, the reserved resources cannot be released in time before the target task arrives because the recycling time of the borrowed task exceeds the safety window. This leads to extended task response time, decreased cluster throughput, increased energy costs, and hindered optimization of carbon emission intensity.

[0042] For example, when processing large model training tasks in an intelligent computing center, specific vGPU resources need to be reserved in advance for periods of low electricity prices and peak green electricity ratios. When low-priority power flow computing tasks borrow these reserved resources, the actual time spent on vGPU profile reconfiguration, task migration, and status verification is not included in the recycling time window calculation. This prevents the resource recycling process from being completed before the target low-carbon window begins, causing large model training tasks to start late, violating the latest start time requirement stipulated in the service level agreement (SLA). Simultaneously, it results in the carbon emission intensity failing to reach the expected optimization level during this period. If this technical problem is not resolved, the resource borrowing mechanism will continue to face the risk of uncontrolled recycling timing. High-priority tasks will frequently trigger SLA violations due to the inability to release resources on time, significantly reducing system reliability. Furthermore, the idle state of reserved resources and recycling failures will lead to a decrease in accelerated resource utilization, and energy consumption will not match the characteristics of the electricity-carbon window, making it difficult to achieve the overall carbon emission intensity optimization target of the intelligent computing center and affecting the sustainable operation of the computing infrastructure.

[0043] In view of this, this application provides a computing-powered collaborative reserved resource scheduling and control method, device, terminal and storage medium, which is used to accurately predict the resource recovery time window, realize the efficient reuse of reserved resources under the constraints of the service level agreement for high-priority tasks, effectively avoid resource idleness and waste, and at the same time ensure the timeliness of resource recovery.

[0044] To make the inventive objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] First, a detailed description of an embodiment of the computing and power-cooperative reserved resource scheduling and control method provided in this application is as follows:

[0046] Please see Figure 1The present application provides a computing-powered collaborative reserved resource scheduling and control method, the steps of which include:

[0047] Step 101: Obtain the task profile of each task to be scheduled, as well as the vGPU resource status and future carbon dioxide window information of the intelligent computing center;

[0048] Step 102: Generate a time-slotted vGPU reservation matrix based on the task profile, the vGPU resource status, and the future carbon window information;

[0049] Step 103: For any future target time slice, extract the target task information, target carbon window information, and target bindable state time corresponding to the target time slice. Use the earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time as the recycling cutoff benchmark time. Then, after deducting the vGPU recycling cost time based on the recycling cutoff benchmark time, obtain the latest recycling time corresponding to the target time slice.

[0050] Step 104: Based on the task information of the resource borrowing request task, calculate the estimated recovery time of the resource borrowing request task using a preset estimated recovery time calculation formula;

[0051] Step 105: When the estimated reclamation time is less than or equal to the safety margin time, a borrowing permission instruction is generated to execute the resource borrowing request task using the reserved vGPU resources before the resource reclamation conditions are triggered.

[0052] For ease of understanding, the following explains some key terms in this embodiment:

[0053] The computing-power collaborative reserved resource scheduling and control method aims to optimize the resource utilization and task execution efficiency of intelligent computing centers, and is especially suitable for scenarios that need to balance computing needs, power costs and carbon emissions.

[0054] Tasks awaiting scheduling refer to various tasks in the intelligent computing center that are waiting to be allocated vGPU resources to perform computing tasks.

[0055] A task profile is a dataset that describes the various attributes of a task to be scheduled, such as task priority, service level agreement requirements, estimated runtime, and vGPU resource requirements.

[0056] The intelligent computing center is a facility that possesses large-scale computing capabilities and integrates technologies such as artificial intelligence and big data to provide intelligent computing services.

[0057] vGPU resource status describes the current configuration, availability, performance metrics, and other information of virtualized graphics processing unit (vGPU) resources within the intelligent computing center.

[0058] Future electricity carbon window information includes data related to energy consumption and environmental impact, such as electricity prices, the proportion of green electricity, and carbon emission intensity, for a future period.

[0059] A time-slotted vGPU reservation matrix is ​​a data structure used to record, in the time dimension, the reservation status of vGPU resources in different future time slices, and when the resource can enter the binding state.

[0060] Reserve vGPU resources, which are pre-allocated and locked for specific high-priority or time-sensitive tasks.

[0061] The reserved status describes whether the reserved vGPU resources are currently occupied, idle, or in other specific states.

[0062] Bindable state time refers to the point in time when the reserved vGPU resources should be ready and available for the target task to begin binding.

[0063] A target time slice refers to a specific period of time in the future, usually the time period during which high-priority tasks are scheduled to be executed or the low-carbon window is open.

[0064] Target task information: For a specific target time slice, detailed attribute data of high-priority tasks associated with that time slice.

[0065] Targeted electricity carbon window information includes information such as electricity price, green electricity ratio, and carbon emission intensity associated with a specific target time slice.

[0066] The latest start time of an SLA is the latest time that a task must begin execution as specified in the Service-Level Agreement (SLA).

[0067] Window start time: The start time of a specific electricity carbon window (e.g., a period of low electricity prices or high green electricity ratio).

[0068] The recycling cutoff baseline time is a reference point used to determine the latest recycling time of resources, which is based on the earliest of several time factors.

[0069] vGPU reclamation cost time is the time required to reclaim a borrowed vGPU resource, including operations such as state saving, migration, and reconfiguration.

[0070] The latest recycling time is the latest point in time when reserved vGPU resources must be recycled and made ready for use by the target task.

[0071] Resource borrowing request tasks refer to low-priority tasks that plan to temporarily use vGPU resources that have been reserved but are currently idle.

[0072] The estimated recovery time calculation formula is a preset formula used to estimate the time required for a resource borrowing request task to be recovered.

[0073] Estimated recovery time: The actual recovery time required for the resource borrowing request task, calculated based on the formula.

[0074] Low-priority tasks awaiting scheduling are those with a lower execution priority than target tasks with reserved resources, and can be interrupted or delayed.

[0075] Safety margin time, the time difference between the latest recycling time and the current time, represents the remaining time that the system can use for recycling operations without affecting the target task.

[0076] The borrowing permission instruction is a system instruction that allows low-priority tasks to temporarily borrow reserved vGPU resources.

[0077] Resource reclamation conditions are specific events or states that trigger the system to perform vGPU resource reclamation operations.

[0078] More specifically, this embodiment provides a computing and power-coordinated reserved resource scheduling and control method, the specific steps of which are as follows:

[0079] This method first obtains task profiles for each scheduled task, as well as the vGPU resource status and future carbon emission window information from the intelligent computing center. Specifically, task profiles can be manually entered or obtained through simple database queries. The vGPU resource status can be manually checked and recorded periodically by the system administrator or obtained through periodic polling of the vGPU management interface. The future carbon emission window information can be obtained from external energy market data sources, such as obtaining current and future electricity prices, green electricity ratios, and carbon emission intensity data through simple API calls.

[0080] Furthermore, based on the acquired task profile, vGPU resource status, and future carbon window information, a time-slotted vGPU reservation matrix is ​​generated. This time-slotted vGPU reservation matrix can be a simple list or table, where each row records the reservation status of a reserved vGPU resource within a specific time slice. The reservation status can be a simple boolean value indicating whether the resource has been reserved. The bindable status time can be set to a fixed time offset, such as N minutes before the start of the reservation time slice.

[0081] More specifically, in one alternative implementation, the slotted vGPU reservation matrix is ​​represented as:

[0082] R={time_slot, node_id, gpu_id, vgpu_profile, compute_share, memory_gb, topology_group, reserved_task, reservation_state, bind_ready_time};

[0083] Wherein, time_slot represents the discretized scheduling slot; node_id represents the physical computing node identifier; gpu_id represents the physical GPU identifier within the node; vgpu_profile represents the vGPU resource specification; compute_share represents the share of computing resources allocated or reserved for the vGPU resource; memory_gb represents the video memory capacity allocated or reserved for the vGPU resource; topology_group represents the hardware topology group to which the vGPU resource belongs; reserved_task represents the reserved task identifier corresponding to the vGPU resource; reservation_state represents the current reservation and borrowing status of the vGPU resource; bind_ready_time represents the time boundary at which the vGPU resource completes reclamation, reconfiguration, and status verification and enters the bindable state, wherein reservation_state includes at least the following states: not reserved, reserved but not occupied, borrowed, reclamation in progress, reclamation completed, bindable, and verification failed.

[0084] Based on this, you can refer to Figure 2 The system considers the timing relationship between reserved vGPU resources being borrowed by low-priority tasks and reclaimed before the target carbon window begins. For any future target time slice, it extracts the target task information, target carbon window information, and target bindable state time corresponding to that time slice. For example, target task information can be directly read from the reservation matrix. Target carbon window information can be filtered from previously obtained future carbon window information. Target bindable state time can also be directly obtained from the reservation matrix. Subsequently, the earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time is used as the reclamation cutoff benchmark time. For example, the system can simply use the latest SLA start time as the sole reclamation cutoff benchmark time. Then, after deducting the vGPU reclamation cost time from this reclamation cutoff benchmark time, the latest reclamation time corresponding to that target time slice is obtained. The vGPU reclamation cost time can be a preset fixed value, such as uniformly set to 5 minutes, regardless of the specific vGPU configuration or task type.

[0085] Furthermore, based on the task information of the resource borrowing request task, the estimated recovery time of the resource borrowing request task is calculated using a preset formula. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow reserved vGPU resources within the target time slice. For example, the estimated recovery time can be a general, conservative estimate for all low-priority tasks, such as a uniformly set value of 10 minutes, or it can be calculated more precisely based on the specific attributes of the task.

[0086] When the estimated reclamation time is less than or equal to the safety margin time, a borrowing permission instruction is generated to execute the resource borrowing request task using the reserved vGPU resources before the resource reclamation condition is triggered. The safety margin time is the difference between the latest reclamation time and the current time. For example, the system can simply determine whether the reserved vGPU resources are currently idle. If they are idle, and the estimated reclamation time of the task is less than or equal to the difference between the latest reclamation time and the current time, then the low-priority task is allowed to temporarily borrow the vGPU resources; otherwise, borrowing is not allowed.

[0087] The following example will provide a more detailed explanation of the above technical solution:

[0088] Suppose that in a smart computing center, there exists a high-priority AI model training task (Task A). This task is required to start and run within the low-carbon electricity price window from 2 PM to 4 PM tomorrow to meet its Service Level Agreement (SLA) and reduce carbon emissions. For this purpose, the system has reserved a specific vGPU resource for Task A in advance. Before 2 PM tomorrow, this reserved vGPU resource will be idle. At this time, a low-priority log analysis task (Task B) requests to borrow this vGPU resource.

[0089] First, the system obtains the task profile of task A (e.g., high priority, latest SLA start time is 2 pm tomorrow, large vGPU demand), the task profile of task B (e.g., low priority, interruptible, small vGPU demand), the current status of the vGPU resource (e.g., idle, configuration information), and the electricity carbon window information for tomorrow afternoon (e.g., the low carbon electricity price window is from 2 pm to 4 pm tomorrow).

[0090] Secondly, based on this information, the system generates or updates a time-slotted vGPU reservation matrix. This matrix records that the vGPU resource is reserved for task A from 2 PM to 4 PM tomorrow, and its bindable state time is set to 2 PM tomorrow.

[0091] Next, for the target time slice of 2 PM tomorrow, the system extracts the latest SLA start time of task A (2 PM tomorrow), the start time of the low-carbon window (2 PM tomorrow), and the bindable state time of the vGPU resource (2 PM tomorrow). The earliest of these three times is 2 PM tomorrow, which is used as the base time for recycling. Subsequently, the system deducts the vGPU recycling cost time (for example, assuming it is 10 minutes, used for saving the context of task A, reconfiguring the vGPU, etc.), thus obtaining the latest recycling time for the vGPU resource as 1:50 PM tomorrow.

[0092] Then, based on the task information of task B, the system calculates the estimated recovery time of task B using a preset formula. For example, based on the type and scale of task B, the system estimates its estimated recovery time to be 5 minutes.

[0093] Finally, assuming the current time is 10:00 AM tomorrow. The system calculates the safety margin time, which is the difference between the latest recycling time (1:50 PM tomorrow) and the current time (10:00 AM tomorrow), to be 3 hours and 50 minutes. Since the estimated recycling time for task B (5 minutes) is less than or equal to the safety margin time (3 hours and 50 minutes), the system generates a borrowing permission instruction. This means that task B is allowed to temporarily borrow the reserved vGPU resources and complete recycling before 1:50 PM tomorrow (specifically, before 1:45 PM, minus 5 minutes), to ensure that task A can start on time at 2:00 PM tomorrow.

[0094] The solution provided in this embodiment achieves intelligent borrowing of reserved resources by finely calculating the "latest recycling time" and "estimated recycling time" and dynamically determining this in conjunction with the "safety margin time." Specifically, by comprehensively considering the latest SLA start time of the target task, the start time of the carbon window, and the bindable state time of the vGPU resources, and deducting the vGPU recycling cost time, the system can accurately determine the final deadline for resources to be recycled. This is more accurate and reliable than simply relying on a single point in time or a fixed recycling time.

[0095] Furthermore, for resource borrowing request tasks, the system no longer uses a rough fixed recovery time, but instead estimates its "estimated recovery time" through a preset calculation formula. This dynamic evaluation based on task attributes allows the system to more accurately determine whether a borrowing task can be recovered in a timely manner without affecting high-priority tasks.

[0096] Ultimately, by comparing the "estimated recovery time" with the "safety margin time," the system can make a decision on whether to allow borrowing. This mechanism ensures that even if low-priority tasks are allowed to borrow reserved resources, there is always sufficient time margin to complete the resource recovery operation, thus avoiding the risk of delays for high-priority tasks due to untimely recovery. This solution improves the utilization efficiency of reserved resources in the intelligent computing center while ensuring the service quality of high-priority tasks, achieving intelligent and refined resource scheduling and control.

[0097] Based on the aforementioned basic implementation method, this application further proposes detailed definitions and supplements for the task profile, vGPU resource status, and future carbon emission window information. The task profile includes: task priority, task service level agreement (SLA), predicted runtime, vGPU resource requirements, memory requirements, delay permission flag, scaling-down permission flag, checkpoint support flag, migration permission flag, and rated release time. The vGPU resource status includes: vGPU profile, memory status, computing power share, physical GPU topology, and slice status. The future carbon emission window information includes: electricity price for the future time slice, green electricity ratio, carbon emission intensity, and the start time of the carbon emission window.

[0098] Specifically, a task profile is a collection of attributes describing a task to be scheduled. Its purpose is to provide the scheduling system with a comprehensive view of the task, enabling accurate resource matching and scheduling decisions. Task priority distinguishes the importance of different tasks, with higher-priority tasks typically enjoying greater resource guarantees. The Task Service Level Agreement (SLA) specifies when a task must start or complete, serving as a key indicator of task timeliness. Predicted runtime is the estimated execution time of the task, helping to estimate resource usage cycles. vGPU resource requirements and memory requirements clarify the specific quantitative requirements of the task for computing and storage resources. Delay permission flags, scaling permission flags, checkpoint support flags, and migration permission flags indicate the flexibility of the task during scheduling, such as whether start delays are allowed, whether runtime resource scaling is allowed, whether breakpoint resumption is supported, and whether migration between different physical resources is allowed. These flags are crucial for formulating resource reclamation and borrowing strategies. Rated release time refers to the time required for the task to release its occupied resources upon normal completion. This information can be obtained from user-specified submissions, predictions based on historical data analysis, or automatic generation by the system based on task type.

[0099] vGPU resource status provides detailed information describing the current and future availability of vGPU resources within the intelligent computing center, ensuring the accuracy and effectiveness of resource allocation. Specifically, the vGPU profile defines the type and configuration of the vGPU, such as its computing power and memory capacity; memory status indicates the current memory usage and available capacity of the vGPU; computing power share represents the proportion of computing power that the vGPU can provide; physical GPU topology describes the physical GPUs to which the vGPU is attached and its connections to other physical GPUs, which is crucial for considering data transfer efficiency and resource isolation; and slice status indicates whether the vGPU has been further subdivided and its current subdivision status. This status information can be queried in real-time through the intelligent computing center's resource management system, obtained through the virtualization layer interface, or collected through hardware monitoring tools.

[0100] Future electricity carbon window information provides forecast data on electricity supply and environmental costs over a future period. Its purpose is to guide the dispatch system to optimize energy efficiency and carbon emissions while meeting performance requirements. Specifically, the electricity price for the future time slice reflects changes in electricity costs across different time periods; the green electricity ratio indicates the proportion of renewable energy in electricity generation, with a high green electricity ratio generally meaning lower carbon emissions; carbon emission intensity directly quantifies the carbon emissions generated per unit of electricity; and the start time of the electricity carbon window specifies the exact time when this information on electricity prices, green electricity ratio, and carbon emission intensity takes effect. This information can be obtained from electricity market data providers, grid operators, environmental monitoring agencies, or estimated using predictive models based on historical data and weather forecasts.

[0101] This application's solution, by acquiring the aforementioned detailed and comprehensive task profiles, vGPU resource status, and future electricity carbon window information, enables the scheduling system to more accurately understand the characteristics of each task to be scheduled, the actual availability of vGPU resources in the intelligent computing center, and the cost-effectiveness of the future power environment. For example, the latest SLA start time in the task profile can be directly used to determine the recycling deadline, ensuring the timely start of high-priority tasks; details such as the vGPU profile and memory status in the vGPU resource status allow the time-slotted vGPU reservation matrix to more accurately reflect the actual capabilities and limitations of the resources; and the future electricity carbon window information provides the scheduling system with a basis for optimizing energy costs and carbon emissions while meeting performance requirements. The introduction of this detailed information allows the system to fully consider the flexibility of tasks, the fine-grained characteristics of resources, and external environmental factors when generating the time-slotted vGPU reservation matrix. This enables a more accurate assessment of the expected recycling time and safety margin time in resource borrowing decisions, effectively balancing resource utilization and task SLA guarantees, and avoiding scheduling blind spots and risks caused by insufficient information.

[0102] Through the above technical solutions, the scheduling system can obtain more comprehensive and detailed information on tasks, resources, and the environment, thereby significantly improving the decision-making accuracy and robustness of the computing-electricity collaborative reserved resource scheduling and control method. This enables the system to more accurately assess the feasibility and risks of resource borrowing, effectively avoid scheduling errors caused by insufficient information, ensure the reliable guarantee of SLAs for high-priority tasks, maximize the utilization of idle resources, and achieve more economical and environmentally friendly resource scheduling by combining electricity carbon information.

[0103] In some of the above embodiments, the computing and power collaborative reserved resource scheduling and control method proposed in this application proposes that when calculating the latest recycling time, the vGPU recycling cost time needs to be deducted from the recycling deadline reference time. Based on this, this application further proposes that the vGPU recycling cost time includes one or more of the following: vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and security buffer time.

[0104] It's important to note that vGPU reclamation time refers to the total time required to restore borrowed vGPU resources to a state usable by reserved tasks when reclaiming them. This time can consist of multiple sub-times, including one or more of the following: vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and security buffer time. vGPU profile reconfiguration time refers to the time spent switching the vGPU configuration from the profile currently used by the borrowing task back to the profile required by the reserved task. This can include adjusting parameters by calling the GPU driver through a software interface, or changing the vGPU's virtualization parameters and resource allocation mode through hardware-level register settings. Checkpoint time refers to the time required to save the state of a low-priority task currently borrowing vGPU resources. This can include saving the task's running state, memory data, register information, etc., to persistent storage media (such as a hard disk), or creating a snapshot of the task state in memory for later recovery. Migration time refers to the time required to migrate a low-priority task currently borrowing vGPU resources from the vGPU or its physical GPU to other computing resources. This can include real-time migration of the task's memory and file system states over the network, or offline migration after the task stops. Release time refers to the time required to unbind vGPU resources from the currently borrowed task and return them to the available resource pool. This can include resource de-allocation operations performed by the operating system or hypervisor, or releasing vGPU memory and compute shares through specific API calls. Profile verification time refers to the time required to verify the new profile configuration after vGPU profile reconfiguration is complete. This can include running a series of diagnostics to check if vGPU parameters meet expectations, or simulating load tests to ensure vGPU functions correctly under the new configuration. Topology verification time refers to the time required to check the physical GPU topology and its mapping to vGPUs during vGPU resource reclamation. This can include confirming the physical GPU's connectivity and availability by querying system hardware information, or verifying the correct slicing of vGPUs on the physical GPUs. Safety buffer time refers to the additional time reserved to handle unforeseen delays or system uncertainties that may occur during the reclamation process. This can be a fixed time period, such as a few seconds or tens of seconds, or it can be a flexible time period that is dynamically calculated based on historical system data, current load, or network conditions.

[0105] Based on this, the expression for the latest collection time mentioned above can be further refined as follows:

[0106]

[0107] Wherein, T_reconfig represents the vGPU profile or slice reconfiguration time, T_checkpoint represents the current borrowed task checkpoint time, T_migration represents the migration time, T_release represents the release time, T_profile_check and T_topology_check represent the profile verification and topology verification times, respectively, and T_buffer represents the safety buffer time. In actual implementation, the appropriate cost item can be selected based on whether the task supports checkpointing, whether migration is allowed, whether re-slicing is required, whether cross-node migration is required, and the topology requirements of the target task. For example, if the task profile shows that migration is not allowed, the migration time T_migration can be set to 0 when calculating the latest recycling time, and so on.

[0108] in, This refers to the recycling cutoff time, where the start time of the target carbon window is T_carbon_window_start, the latest start time allowed for high-priority tasks under SLA conditions is T_sla_latest_start, and the time when the vGPU resource should enter the bindable state in the reservation matrix is ​​T_reservation_start. The system selects the earliest of these three as the recycling cutoff time to ensure that the resource is recycled, reconfigured, and its status is verified before the earliest constraint arrives.

[0109] When the target carbon window arrives before the latest start time of the SLA, the system uses the start time of the target carbon window as the recycling cutoff benchmark; when the task SLA is tighter, the system uses the latest start time of the SLA as the recycling cutoff benchmark; when the reservation matrix requires resources to be bindable in advance, the system uses the bindable state time in the matrix as the recycling cutoff benchmark.

[0110] After obtaining the recycling deadline, the system deducts the recycling cost based on the actual state of the current borrowed task and the target vGPU resources. If the borrowed task supports checkpointing, the checkpointing time is deducted; if it supports migration, the migration time is deducted; if it can only be released, the release time is deducted; if the target task requires a different vGPU profile or a different slice size, the profile or slice reconfiguration time is deducted.

[0111] This application's solution refines the vGPU reclamation cost time into multiple components, including vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and safety buffer time. This allows for a more comprehensive consideration of the various time-consuming aspects of the resource reclamation process when calculating the latest reclamation time. After determining the reclamation deadline, the system no longer simply deducts a vague reclamation time but precisely accumulates the time required for each specific reclamation operation. For example, when a low-priority task borrows vGPU resources, if the task supports checkpointing and migration, the calculation of the reclamation cost will include checkpointing time, migration time, and subsequent vGPU profile reconfiguration, profile verification, and topology verification times. If the task does not support migration, the release time will be considered. Furthermore, a safety buffer time can be introduced to effectively address uncertainties such as system load fluctuations and network latency, providing additional fault tolerance for resource reclamation. This refined time decomposition and accumulation mechanism ensures that the calculation result of the latest recovery time is closer to reality, thus providing a more reliable basis for resource borrowing decisions and avoiding resource conflicts or scheduling failures caused by insufficient recovery time estimation.

[0112] The above technical solution refines the vGPU reclamation cost time into specific components such as vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and security buffer time, making the calculation of the latest reclamation time more accurate and comprehensive. This effectively avoids resource reclamation delays or failures caused by inaccurate reclamation time estimation, thereby improving the reliability and efficiency of the computing and power-coordinated reserved resource scheduling control method. A precise latest reclamation time ensures that high-priority tasks can obtain the necessary vGPU resources in a timely manner before the latest SLA start time or at the start of the optimal power window, thus optimizing the resource utilization and energy efficiency of the intelligent computing center while guaranteeing the service quality of tasks.

[0113] In some of the above embodiments, the method of this application proposes the need to accurately assess the estimated recovery time of the resource borrowing request task to ensure that the borrowed vGPU resources can be recovered in a timely and smooth manner when the resource recovery condition is triggered, avoiding impact on the normal execution of high-priority tasks or the effective utilization of the carbon window. Based on this, this application further proposes the following formula for calculating the estimated recovery time:

[0114] ;

[0115] In the formula, T_reclaim_est represents the estimated reclaim time, T_scale_down represents the task's scaling down or de-rating time, T_checkpoint represents the current checkpoint time of the resource borrowing request task, T_migration_release represents the task's migration or release time, T_rebind represents the vGPU rebinding time, and T_verify represents the status verification time.

[0116] More specifically, the estimated reclaim time T_reclaim_est is a key metric for assessing whether a resource borrowing request task can safely borrow reserved vGPU resources. It represents the total time required from the decision to reclaim resources to their full availability. This time can be obtained by summing the times of each reclaiming phase or by prediction based on historical data and machine learning models. The task scaling down or de-rating time T_scale_down refers to the time required to perform resource reduction or performance degradation operations on a running resource borrowing request task before reclaiming resources. This can include gradually reducing the number of vGPUs allocated to the task or lowering the task's computational priority. The current resource borrowing request task checkpoint time T_checkpoint refers to the time required to save the current running state of the resource borrowing request task to persistent storage so that the task can be recovered from that state later. This may involve writing the task's memory, register state, and other critical data to disk or a distributed storage system. The task migration or release time T_migration_or_release refers to the time required to migrate the resource borrowing request task from its current vGPU resources to other available resources, or to directly terminate the task and release the vGPU resources it occupies. Migration can employ either hot or cold migration techniques, while release typically involves process termination and resource cleanup. The vGPU rebinding time, T_rebind, refers to the time required to reconfigure and bind a vGPU resource to its intended high-priority task or bring it into a usable state after it has been reclaimed. This may include operations such as vGPU driver reinitialization and virtualization-level configuration updates. The state verification time, T_verify, refers to the time required to verify the functionality and state of a vGPU resource after it has been reclaimed and rebinded to ensure it is in a normal working state and available for subsequent tasks. This may include running a series of diagnostic tests or checking whether various vGPU parameters meet expectations.

[0117] This application's solution constructs a comprehensive and accurate model for calculating the estimated recovery time by meticulously decomposing and summing multiple key time-consuming steps involved in the resource recovery process. This model considers the characteristics of the task itself (such as whether it supports scaling down, checkpointing, and migration) and the inherent time consumption of vGPU resource operations (such as rebinding and status verification), thereby quantifying the complexity of resource recovery. This refined time evaluation mechanism is closely integrated with the steps in the basic method that determine the latest recovery time and safety margin time through the recovery deadline benchmark time, enabling the system to make more reliable time assessments when judging the safety of borrowing requests for low-priority tasks. This not only effectively avoids the risk of untimely resource recovery due to inaccurate recovery time estimation and ensures the effective utilization of the Service Level Agreement (SLA) and carbon window for high-priority tasks, but also maximizes the utilization of idle reserved resources while ensuring safety, thus improving the overall resource utilization efficiency of the intelligent computing center.

[0118] In some of the above embodiments, this application proposes a computing-power-coordinated reserved resource scheduling and control method. This method proposes generating a time-slotted vGPU reservation matrix based on task profiles, vGPU resource status, and power consumption window information, and allows low-priority tasks to borrow reserved vGPU resources under specific conditions, thereby improving resource utilization. Based on this, such as... Figure 3 As shown in the overall process logic diagram, this application further proposes that after generating the borrowing permission instruction to execute the resource borrowing request task using the reserved vGPU resources before the resource reclamation condition is triggered, the following steps are also included:

[0119] When the resource reclamation condition is triggered, a reclamation action sequence is executed according to the latest reclamation time to reclaim the vGPU resources borrowed by the resource borrowing request task; after the reclamation action sequence is completed, a status check is performed on the reclaimed reserved vGPU resources, and the time-slotted vGPU reservation matrix is ​​updated according to the status check result.

[0120] It's important to note that "when resource reclamation conditions are triggered" refers to the occurrence of a specific event or state that requires reserved vGPU resources to be reclaimed. This condition can be a preset time point, such as the target task's latest SLA start time approaching, or the target electrocarbon window about to begin. It can also be a change in resource status detected by the system, such as abnormal vGPU profile status, abnormal memory status, abnormal physical GPU topology status, or abnormal slice status. Furthermore, resource reclamation can also be triggered when the current time reaches the latest reclamation time, or when the safety margin time falls below a preset margin time threshold. These conditions serve as signals to initiate the resource reclamation process, ensuring that the system can respond promptly to changes in resource demand.

[0121] "Executing a sequence of reclamation actions based on the latest reclamation time" refers to the system initiating a series of operations aimed at releasing vGPU resources from the borrowed task and restoring them to a usable state based on a predetermined latest reclamation time. This sequence of actions may include interrupting the borrowed task, saving the task state (e.g., through a checkpoint mechanism), unbinding the vGPU resources from the borrowed task, and cleaning up residual resource data. The system can plan and execute these actions based on the various time-consuming components included in the vGPU reclamation cost time (such as vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and security buffer time) to ensure that resources are effectively reclaimed within the specified time. For example, a forced reclamation mechanism can be used to directly terminate the borrowed task and release resources to handle emergencies; or a negotiated reclamation mechanism can be used to send a reclamation notification to the borrowed task, allowing it to complete its cleanup work within a certain time.

[0122] "Performing status verification on reclaimed reserved vGPU resources" refers to the system checking the vGPU resources after the reclamation action sequence is completed to confirm whether they are in the expected state, thereby ensuring their availability and correctness. This verification may include checking whether the vGPU profile is correct, whether the video memory has been cleared, whether the computing power share has been restored, whether the physical GPU topology is complete, and whether the slicing status is normal. The system can use preset verification scripts or tools for automated checks, or it can interface with the vGPU resource management system to obtain and compare resource status information. In addition, simple functional tests can be performed, such as attempting to assign a small test task to verify resource availability.

[0123] "Updating the slotted vGPU reservation matrix based on the status verification result" means that the system adjusts the reservation status and bindable status time of the corresponding resources in the slotted vGPU reservation matrix according to the status verification result. If the verification passes, the reservation status of the corresponding resource in the matrix is ​​updated to "available" or "ready," and the bindable status time is updated as needed. If the verification fails, the reservation status of the corresponding resource in the matrix is ​​marked as "abnormal" or "unavailable," and an alarm or further fault handling process may be triggered. This measure aims to ensure that the resource scheduling system has an accurate understanding of the availability status of vGPU resources and avoid scheduling errors.

[0124] This application's solution establishes a complete resource reclamation and verification mechanism, ensuring that while allowing low-priority tasks to borrow reserved vGPU resources, the intelligent computing center can reliably reclaim these resources and verify their status when needed. When resource reclamation conditions are triggered, the system initiates a series of carefully designed reclamation actions based on the pre-calculated latest reclamation time, releasing the vGPU resources from the borrowing tasks. This process fully considers the time required for reclamation to ensure resources are freed up in a timely manner. After reclamation, the system immediately performs status verification on the vGPU resources, comprehensively checking their profile, memory, computing power share, topology, and slice status to ensure the resources are clean, available, and meet the requirements of subsequent high-priority tasks. Finally, based on the verification results, the system updates the time-slotted vGPU reservation matrix in real time, providing accurate and reliable resource status information for subsequent resource scheduling. This forms a closed-loop resource management process, enabling the intelligent computing center to confidently borrow resources, improving resource utilization while effectively guaranteeing the service level agreements (SLAs) for high-priority tasks.

[0125] In some of the above embodiments, this application proposes generating a time-slotted vGPU reservation matrix based on task profiles, vGPU resource status, and future carbon window information, and allowing low-priority resource borrowing request tasks to borrow reserved vGPU resources while meeting the safety margin time. Furthermore, this application proposes that when reserved vGPU resource reclamation fails or the status verification result fails, the affected time slice, reserved vGPU resources, and / or target task are determined, and a partial recalculation is performed on the affected time slice, reserved vGPU resources, and / or target task, and the time-slotted vGPU reservation matrix is ​​updated based on the recalculation result.

[0126] It should be noted that "reserved vGPU resource reclamation failure" refers to the failure to successfully release or migrate the borrowed vGPU resources from the resource borrowing request task and restore them to a bindable state during the execution of the reclamation action sequence. This can be caused by various reasons, such as abnormal termination of the resource borrowing request task leading to the inability to release resources normally, network failures or insufficient target vGPU resource pools during migration, or timeouts in the reclamation action sequence itself. Another situation is "state verification result failed," which means that even if the reclamation action sequence appears to be completed, the state of the reclaimed reserved vGPU resources is found to be inconsistent with expectations during state verification, such as vGPU profile mismatch, abnormal memory state, or abnormal physical GPU topology or slice state. These abnormal states may prevent subsequent high-priority tasks from using the vGPU resources normally.

[0127] When any of the above-mentioned anomalies are detected, the system needs to "determine the affected time slice, reserved vGPU resources, and / or target task." The purpose of this step is to accurately identify the scope of the fault and avoid unnecessary interference with the entire scheduling system. Specifically, the system can locate which reserved vGPU resource within which time slice has the problem, and which target task originally scheduled to use that resource may be affected, based on the log information of the failed recycling operation, the anomaly report of the status verification, and the reservation information recorded in the time-slotted vGPU reservation matrix. For example, it can query the vGPU resource ID associated with the failed recycling operation and then reverse-engineer the future time slice and target task corresponding to that vGPU resource in the time-slotted vGPU reservation matrix.

[0128] After determining the affected scope, the system will perform a partial recalculation of the affected time slices, reserved vGPU resources, and / or target tasks. "Partial recalculation" refers to re-evaluating resource allocation and task scheduling schemes within the affected scope to correct scheduling plan defects caused by failed reclamation or state verification. For example, if the reclamation of a reserved vGPU resource fails, the system can attempt to find an alternative vGPU resource for the target task originally scheduled to use that resource; or, if no alternative resource can be found, it may need to adjust the latest SLA start time of the target task and reschedule it to another available time slice. Partial recalculation can employ various strategies, such as rerunning a lightweight scheduling algorithm based on the latest vGPU resource status and carbon window information, optimizing only the affected resources and tasks; or, according to a preset fault handling strategy, prioritizing attempts to repair the vGPU resource status (e.g., reconfiguring the vGPU profile), and migrating or rescheduling tasks if repair fails.

[0129] For example, after the system reclaims borrowed vGPU resources, it discovers that memory fragmentation on the target physical GPU prevents the successful creation of the target vGPU profile. In this case, the system does not perform a global rescheduling of all tasks and resources. Instead, it selects only a local region containing the physical GPU, the corresponding time slice, the target task, and related borrowed tasks for recalculation. Local recalculation prioritizes finding vGPU resources within the same topology group that meet the memory and profile requirements. If the same topology group cannot meet these requirements, resources from adjacent topology groups are selected based on the target task's bandwidth and latency requirements. If these requirements still cannot be met, the next level of reclamation is triggered early, or low-priority tasks are relocated to the target.

[0130] Local recalculation can reduce the overhead of global rescheduling and avoid affecting the existing scheduling results of other time slots and tasks.

[0131] Finally, the system updates the time-slotted vGPU reservation matrix based on the recalculation results. This step ensures the real-time performance and accuracy of the scheduling plan. After recalculation, all new scheduling decisions, including the latest reservation status of vGPU resources, bindable state times, and adjustment information for affected tasks, are promptly written into the time-slotted vGPU reservation matrix. For example, if a vGPU resource is marked as unavailable, or a target task is rescheduled to a new time slice, these changes will be reflected in the matrix, providing an accurate basis for subsequent scheduling decisions.

[0132] This embodiment effectively addresses anomalies such as failed vGPU resource reclamation or failed state verification by introducing mechanisms for fault detection, impact scope determination, local recalculation, and matrix updates. While allowing low-priority tasks to borrow reserved resources improves resource utilization in the aforementioned computational-powered collaborative reserved resource scheduling control method, it also introduces the risk of resource reclamation failure. This solution addresses this risk by providing a robust error handling mechanism. When problems arise during resource reclamation or state verification, the system no longer simply freezes or reports an error; instead, it intelligently identifies the scope of the problem and performs localized, targeted scheduling adjustments. This approach avoids the significant computational overhead and time delay associated with a global recalculation of the entire scheduling plan, thus ensuring the response speed and efficiency of the scheduling system. By locally recalculating the affected time slices, reserved vGPU resources, and / or target tasks, and promptly updating the time-slotted vGPU reservation matrix, the system can quickly recover to a consistent and executable scheduling state, minimizing the impact of localized failures on the overall scheduling plan and the execution of high-priority tasks.

[0133] The above is a detailed description of an embodiment of a computing and power-coordinated reserved resource scheduling and control method provided in this application. The following is a detailed description of an embodiment of a computing and power-coordinated reserved resource scheduling and control device provided in this application.

[0134] Please see Figure 4 This application provides a computer-aided collaborative reserved resource scheduling and control device, which includes:

[0135] The basic information acquisition unit 201 is used to acquire the task profile of each task to be scheduled, as well as the vGPU resource status and future carbon window information of the intelligent computing center;

[0136] The time-slotted vGPU reservation determination unit 202 is used to generate a time-slotted vGPU reservation matrix based on the task profile, the vGPU resource status, and the future carbon window information. The time-slotted vGPU reservation matrix is ​​used to record the reserved vGPU resources and their reservation status in several future time slices, as well as the bindable state time of the reserved vGPU resources. The bindable state time is the time when the reserved vGPU resources should enter the bindable state.

[0137] The latest recycling time determination unit 203 is used to extract the target task information, target carbon window information and target bindable state time corresponding to any future target time slice for any future target time slice. The earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information and the target bindable state time is used as the recycling cutoff benchmark time. Then, the latest recycling time corresponding to the target time slice is obtained by deducting the vGPU recycling cost time based on the recycling cutoff benchmark time.

[0138] The resource borrowing and recycling time calculation unit 204 is used to calculate the estimated recycling time of the resource borrowing request task based on the task information of the resource borrowing request task and through a preset estimated recycling time calculation formula. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow the reserved vGPU resources in the target time slice.

[0139] The resource borrowing determination unit 205 is used to generate a borrowing permission instruction when the estimated recycling time is less than or equal to the safety margin time, so as to execute the resource borrowing request task using the reserved vGPU resources before the resource recycling condition is triggered, wherein the safety margin time is the difference between the latest recycling time and the current time.

[0140] This embodiment determines the latest recycling time by combining the earliest of the target task's latest SLA start time, the target carbon window start time, and the bindable state time of reserved vGPU resources, and deducting the vGPU recycling cost time. Simultaneously, it calculates the estimated recycling time for resource borrowing request tasks. This allows low-priority tasks to safely borrow reserved resources while ensuring high-priority tasks can start in a timely manner. Specifically, this scheme dynamically determines the latest recycling time by finely calculating the difference between the recycling deadline and the vGPU recycling cost time, avoiding the coarse judgment of traditional mechanisms that rely solely on the current idle state of resources. Furthermore, the estimated recycling time calculation formula based on task attributes enables the system to accurately assess the recycling feasibility of low-priority tasks, ensuring resource recycling operations are completed within the safety margin. Through this technical solution, the problem of resource waste caused by reserved resources remaining idle before the target task arrives, and the contradiction of low-priority tasks not recycling in a timely manner after borrowing resources affecting the execution of high-priority tasks, are effectively solved. This improves the resource utilization rate of the intelligent computing center while ensuring the timeliness of task execution and compliance with service level agreements. For example, when a target task requires startup within a low-carbon electricity price window, the device can accurately calculate the time window for resource recovery and only allow low-priority tasks with expected recovery times within a safety margin to borrow resources, thereby balancing energy cost optimization and task scheduling reliability.

[0141] In other embodiments, this application proposes a computing and power collaborative reserved resource scheduling and control terminal, comprising: a memory and a processor; the memory is used to store program code corresponding to the computing and power collaborative reserved resource scheduling and control method described above; the processor is used to read and execute the program code to implement the computing and power collaborative reserved resource scheduling and control method.

[0142] The solution in this application, through the aforementioned technical means, ensures that even when low-priority tasks are allowed to borrow reserved vGPU resources, there is always sufficient time margin to complete the resource reclamation operation. Compared to traditional resource scheduling mechanisms, this terminal no longer simply determines whether the resource is currently idle, but comprehensively considers the latest SLA start time of the target task, the start time of the carbon window, and the bindable state time of the vGPU resource, and deducts the vGPU reclamation cost time, thereby accurately determining the final deadline for the resource to be reclaimed. In addition, for resource borrowing request tasks, the terminal estimates its expected reclamation time through a preset calculation formula. This dynamic evaluation based on task attributes enables the system to more accurately determine whether the borrowing task can be reclaimed in a timely manner without affecting high-priority tasks.

[0143] Through the above technical solution, this terminal significantly improves the utilization efficiency of reserved resources in the intelligent computing center while ensuring the service quality of high-priority tasks, achieving intelligent and refined resource scheduling and control. As a specific implementation method, this terminal can be deployed on the management server of the intelligent computing center. By executing stored program code, it can monitor the vGPU resource status in real time and dynamically adjust resource allocation strategies. This maximizes the utilization of the low-cost, low-carbon emission window of electricity while meeting service level agreements, thereby reducing overall operating costs and environmental impact.

[0144] In other embodiments, this application proposes a computer-readable storage medium storing program code that is read and executed by a processor to implement the above-described computer-computer collaborative reserved resource scheduling and control method.

[0145] Specifically, the program code enables the processor to generate a time-slotted vGPU reservation matrix based on the task profile, vGPU resource status, and future carbon window information. This time-slotted vGPU reservation matrix records the reserved vGPU resources and their reservation status within several future time slices, as well as the bindable state time of the reserved vGPU resources. For any future target time slice, the processor extracts the target task information, target carbon window information, and target bindable state time corresponding to the target time slice. The earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time is used as the recycling cutoff benchmark time. Then, the latest recycling time corresponding to the target time slice is obtained by subtracting the vGPU recycling cost time from the recycling cutoff benchmark time.

[0146] Based on this, the processor calculates the estimated recovery time of the resource borrowing request task according to the task information of the resource borrowing request task using a preset formula for estimated recovery time calculation. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow reserved vGPU resources within the target time slice. When the estimated recovery time is less than or equal to the safety margin time, the processor generates a borrowing permission instruction to execute the resource borrowing request task using the reserved vGPU resources before the resource recovery condition is triggered. The safety margin time is the difference between the latest recovery time and the current time. The solution in this application effectively solves the problem of reserved vGPU resources being idle when the target task has not arrived, while avoiding the risk of high-priority task delays due to untimely recovery.

[0147] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the terminals, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed terminals, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0149] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0150] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0151] 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0152] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0153] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or 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.

[0154] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for scheduling and controlling reserved resources in a computer-computer collaborative manner, characterized in that, include: Obtain task profiles for each task to be scheduled, as well as the vGPU resource status and future carbon dioxide window information of the intelligent computing center; A time-slotted vGPU reservation matrix is ​​generated based on the task profile, the vGPU resource status, and the future carbon window information. The time-slotted vGPU reservation matrix is ​​used to record the reserved vGPU resources and their reservation status in several future time slices, as well as the bindable state time of the reserved vGPU resources. The bindable state time is the time when the reserved vGPU resources should enter the bindable state. For any future target time slice, extract the target task information, target carbon window information, and target bindable state time corresponding to the target time slice. Use the earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time as the recycling cutoff benchmark time. Then, after deducting the vGPU recycling cost time based on the recycling cutoff benchmark time, obtain the latest recycling time corresponding to the target time slice. Based on the task information of the resource borrowing request task, the estimated recovery time of the resource borrowing request task is calculated using a preset estimated recovery time calculation formula. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow the reserved vGPU resources within the target time slice. When the estimated reclamation time is less than or equal to the safety margin time, a borrowing permission instruction is generated to execute the resource borrowing request task using the reserved vGPU resources before the resource reclamation condition is triggered, wherein the safety margin time is the difference between the latest reclamation time and the current time.

2. The computing-coordinated reserved resource scheduling and control method according to claim 1, characterized in that, The task profile includes: task priority, task service level agreement (SLA), predicted runtime, vGPU resource requirements, video memory requirements, deferral permission flag, scaling down permission flag, checkpoint support flag, migration permission flag, and rated release time. The vGPU resource status includes: vGPU profile, video memory status, computing power share, physical GPU topology, and slice status; The future carbon window information includes: electricity price, green electricity ratio, carbon emission intensity, and start time of the future time slice.

3. The computing-coordinated reserved resource scheduling and control method according to claim 1, characterized in that, The vGPU reclamation cost time includes one or more of the following: vGPU profile reconfiguration time, checkpoint time, migration time, release time, profile verification time, topology verification time, and security buffer time.

4. The computing-coordinated reserved resource scheduling and control method according to claim 1, characterized in that, The formula for calculating the estimated recovery time is: ； In the formula, T_reclaim_est represents the estimated reclaim time, T_scale_down represents the task's scaling down or de-rating time, T_checkpoint represents the checkpoint time of the current resource borrowing request task, T_migration_release represents the task's migration or release time, T_rebind represents the vGPU rebinding time, and T_verify represents the status verification time.

5. The computing-powered collaborative reserved resource scheduling and control method according to claim 2, characterized in that, The generation of the borrowing permission instruction, which executes the resource borrowing request task using the reserved vGPU resources before the resource reclamation condition is triggered, further includes: When the resource reclamation condition is triggered, a reclamation action sequence is executed according to the latest reclamation time to reclaim the vGPU resources borrowed by the resource borrowing request task; After the recycling action sequence is completed, a status check is performed on the recycled reserved vGPU resources, and the time-slotted vGPU reservation matrix is ​​updated according to the status check result.

6. The computing-powered collaborative reserved resource scheduling and control method according to claim 5, characterized in that, The resource recycling conditions include: the target task has arrived, the target electrocarbon window is about to begin, the current time has reached the latest recycling time, the safety margin time is lower than the preset margin time threshold, the vGPU profile status is abnormal, the video memory status is abnormal, the physical GPU topology status is abnormal, or the slice status is abnormal.

7. The computing-powered collaborative reserved resource scheduling and control method according to claim 5, characterized in that, Also includes: When the reserved vGPU resource reclamation fails or the status verification result fails, the affected time slice, reserved vGPU resources and / or target task are determined, and the affected time slice, reserved vGPU resources and / or target task are locally recalculated, and the time-slotted vGPU reservation matrix is ​​updated according to the recalculation result.

8. A computer-aided collaborative reserved resource scheduling and control device, characterized in that, include: The basic information acquisition unit is used to acquire the task profiles of each task to be scheduled, as well as the vGPU resource status and future carbon window information of the intelligent computing center; A time-slotted vGPU reservation determination unit is used to generate a time-slotted vGPU reservation matrix based on the task profile, the vGPU resource status, and the future carbon window information. The time-slotted vGPU reservation matrix is ​​used to record the reserved vGPU resources and their reservation status in several future time slices, as well as the bindable state time of the reserved vGPU resources. The bindable state time is the time when the reserved vGPU resources should enter the bindable state. The latest recycling time determination unit is used to extract the target task information, target carbon window information, and target bindable state time corresponding to any future target time slice for any future target time slice. The earliest time among the latest SLA start time in the target task information, the window start time in the target carbon window information, and the target bindable state time is used as the recycling cutoff benchmark time. Then, the latest recycling time corresponding to the target time slice is obtained by deducting the vGPU recycling cost time based on the recycling cutoff benchmark time. The resource borrowing and recycling time calculation unit is used to calculate the estimated recycling time of the resource borrowing request task based on the task information of the resource borrowing request task and through a preset estimated recycling time calculation formula. The resource borrowing request task refers to a low-priority scheduled task that plans to borrow the reserved vGPU resources in the target time slice. The resource borrowing determination unit is used to generate a borrowing permission instruction when the estimated recycling time is less than or equal to the safety margin time, so as to execute the resource borrowing request task using the reserved vGPU resources before the resource recycling condition is triggered, wherein the safety margin time is the difference between the latest recycling time and the current time.

9. A computer-aided collaborative reserved resource scheduling and control terminal, characterized in that, include: Memory and processor; The memory is used to store program code, which corresponds to the computing and power-coordinated reserved resource scheduling and control method as described in any one of claims 1 to 7; The processor is used to read and execute the program code to implement the computing and power collaborative reserved resource scheduling and control method.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that is read and executed by a processor to implement the computer-computer collaborative reserved resource scheduling and control method as described in any one of claims 1 to 7.

Images