A cloud computing resource pool dynamic scheduling method

Abstract

The application discloses a kind of cloud computing resource pool dynamic scheduling method, it is related to cloud computing resource scheduling technical field, method is executed by the resource scheduler of deployment in scheduling node, when task is deployed to target physical node, official bottom quota and resource credit limit are configured for task, the sum of both is set as the upper limit of resource use of task.In the process of task running, periodically collect actual resource use data to generate sliding window average load value, according to the change of load, dynamically complete the mutual transformation of two types of quota under the premise of keeping the upper limit of resource use unchanged;While setting supporting mechanism of hierarchical downshift, oversell control, overload emergency recovery and multi-type resource differentiation scheduling mechanism.The application can effectively improve the resource utilization efficiency of cloud computing resource pool, enhance the load fluctuation response capability, balance business operation stability and resource use benefit, and can be widely applied to resource scheduling management of various cloud data centers.

Description

Technical Field

[0001] This invention relates to the field of cloud computing resource scheduling technology, specifically a dynamic scheduling method for cloud computing resource pools. Background Technology

[0002] With the large-scale deployment and application of cloud computing technology, cloud data centers abstract and integrate the computing, memory, network, and other hardware resources of massive physical servers into a unified resource pool, providing on-demand and elastically scalable computing power services to multiple tenants. Resource scheduling, as a core component of the cloud resource pool management system, is responsible for allocating, dynamically adjusting, and reclaiming resources according to task requirements. The rationality of its scheduling strategy directly determines the resource utilization efficiency and business operation stability of the data center. Improving resource utilization efficiency while ensuring service levels is the core objective of ongoing research in the field of cloud computing resource scheduling.

[0003] Currently, the mainstream resource allocation modes for cloud computing resource pools are mainly divided into two categories: static full allocation and dynamic elastic scaling. In the static full allocation mode, all resources are allocated and assigned rights at once according to the resource requests submitted by the user during the task startup phase. This mode has simple implementation logic and high operational reliability, and is the basic allocation method commonly used by current cloud platforms. However, in practical applications, users often overestimate the peak resource demand of tasks to ensure stable business operation, resulting in a large number of allocated resources remaining idle at low load for extended periods, leading to a low overall utilization efficiency of the resource pool. To alleviate resource waste, dynamic elastic scaling technology has been widely adopted. This technology monitors task load levels in real time and expands resources when the load increases and shrinks resources when the load decreases, thus adapting to dynamic load changes. However, existing dynamic elastic scaling solutions generally adopt a threshold-triggered post-event adjustment mechanism. The resource expansion operation has an inherent time delay, which cannot effectively cope with sudden load growth and is prone to causing business performance degradation or even service level agreement (SLA) breaches. At the same time, existing adjustment mechanisms mostly operate at the instance group or cluster level and do not build fine-grained resource allocation and adjustment logic for the complete lifecycle of individual tasks, making it difficult to achieve the dual goals of low resource idle rate and low burst response latency at the same time. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a dynamic scheduling method for cloud computing resource pools. This method addresses the problems of existing cloud computing resource scheduling technologies, such as difficulty in simultaneously achieving low resource idle rates and low burst load response latency, serious waste in static full allocation of resources, and lag in dynamic elastic scaling adjustments with insufficient scheduling fineness.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a dynamic scheduling method for a cloud computing resource pool, applied to a cloud computing resource pool containing multiple physical nodes, executed by a resource scheduler deployed on a scheduling node, including the steps of receiving resource request information submitted by a task, selecting a target physical node based on the resource request information, and allocating corresponding computing resources to the target physical node, characterized in that the step of allocating corresponding computing resources to the target physical node specifically includes:

[0006] When a task is deployed to a target physical node, the resource scheduler configures a formal guaranteed quota and a resource credit quota for the task. The formal guaranteed quota is set as the resource exclusively occupied by the task, the resource credit quota is set as the shared buffer resource that the task can temporarily use, and the upper limit of the task's resource usage is set as the sum of the formal guaranteed quota and the resource credit quota.

[0007] During task execution, the resource scheduler periodically collects actual resource usage data of the task and generates a sliding window average load value. When the sliding window average load value continuously exceeds the preset upgrade threshold of the current formal minimum quota, the corresponding share of resource credit quota is converted into the formal minimum quota according to the preset step size, and the resource credit quota value is reduced simultaneously to keep the upper limit of resource usage unchanged.

[0008] When the average load value of the sliding window continues to be lower than the preset downgrade threshold of the current official minimum quota, the resource credit limit value is reduced first, and then the official minimum quota value is reduced by a preset step size. At the same time, the resource credit limit value is replenished to keep the upper limit of resource usage stable until the official minimum quota falls back to the initial value.

[0009] Furthermore, in the process of configuring the formal minimum quota and resource credit limit for the task, the resource scheduler calculates the initial formal minimum quota value based on the total amount of resource requests submitted by the task and the benchmark allocation ratio corresponding to the task type. The initial resource credit limit value is the difference between the total amount of resource requests and the initial formal minimum quota.

[0010] The initial guaranteed minimum quota is calculated using the following formula: In the formula, This indicates the initial official minimum allocation for the task; This indicates the total amount of resources requested by the task. This indicates the baseline allocation ratio for the corresponding task type.

[0011] Furthermore, in the process of generating the sliding window average load value, the resource scheduler collects the real-time resource utilization rate of the task at a fixed period, and calculates the arithmetic average of the values ​​of multiple consecutive collection periods as the sliding window average load value.

[0012] The average load value of the sliding window is calculated using the following formula: In the formula, This represents the average load value of the sliding window; This indicates the total number of acquisition cycles contained within the sliding window; This represents the real-time resource utilization rate of the task within the i-th collection period. When the average load value of the sliding window is greater than the current guaranteed minimum quota multiplied by the preset upgrade coefficient within multiple consecutive judgment periods, an upgrade operation is triggered. The adjustment step size of each upgrade is a preset proportion of the total resource application amount. A cooldown period is set after the upgrade, and the upgrade operation is not triggered again during the cooldown period.

[0013] Furthermore, during the process of reducing the resource credit limit and lowering the official minimum quota, the resource scheduler first performs a credit limit reduction operation. When the average load value of the sliding window is continuously lower than the preset first tier reduction threshold and reaches the first duration, the resource credit limit is reduced to the current official minimum quota multiplied by a preset buffer ratio. When the load is continuously lower than the preset second tier reduction threshold and reaches the second duration, the official minimum quota is lowered. The step size of each reduction is a preset ratio of the total resource application amount. After the reduction, the corresponding resource credit limit is replenished simultaneously to keep the upper limit of resource usage stable.

[0014] Furthermore, the resource credit limit is managed globally using an over-selling mechanism. The resource scheduler calculates the total resource credit limit of all tasks on the target physical node, and calculates the over-selling coefficient in combination with the total amount of remaining available resources on the node, keeping the over-selling coefficient within a preset range.

[0015] The node resource overselling coefficient is calculated using the following formula: In the formula, This represents the resource overselling coefficient of the target physical node; This represents the total resource credit limit for all tasks on the target physical node; This represents the total remaining available resources of the target physical node. When the overall load of the node exceeds the preset load threshold, the resource scheduler stops granting resource credits to newly deployed tasks and simultaneously gradually reduces the resource credit values ​​of already deployed tasks to ensure stable node operation.

[0016] Furthermore, the computing resources include CPU resources and memory resources. Independent official minimum quotas and resource credit limits are set for CPU resources and memory resources respectively, and upgrade and downgrade operations are performed respectively. For the downgrade operation of memory resources, the resource scheduler first verifies the actual memory usage of the task. Only when the actual memory usage is lower than the preset ratio of the target downgraded official minimum quota, the official minimum quota downgrade operation is performed.

[0017] Furthermore, the computing resources also include heterogeneous computing power resources. For heterogeneous computing power resources, the formal guaranteed quota corresponds to a dedicated computing power shard, and the resource credit quota corresponds to a preemptible computing power shard. When a high-priority task has computing power requirements, the resource scheduler will prioritize reclaiming the preemptible computing power shards of low-priority tasks and allocating them to high-priority tasks for use. Low-priority tasks retain their dedicated computing power shards to maintain basic operation.

[0018] Furthermore, when the overall resource load of the target physical node reaches the overload warning threshold, the resource scheduler reclaims the resource credits of the corresponding tasks in order of task priority from low to high, retaining only the formal minimum quota of the task, until the overall resource load of the node falls back to the safe range; the tasks whose resource credits have been reclaimed are marked as pending migration and are asynchronously scheduled by the resource scheduler to migrate to physical nodes with sufficient resources.

[0019] Furthermore, after each upgrade and downgrade operation, the resource scheduler sets a corresponding cooldown period. During the cooldown period, quota adjustment operations in the same direction are not triggered. For sudden peak load scenarios, when the real-time resource utilization rate of a task continuously reaches the resource utilization limit and exceeds the preset time, the tiered adjustment process is skipped, and the official minimum quota is directly increased to a value that matches the current actual load.

[0020] Furthermore, the adjustment of the formal guaranteed quota and resource credit limit is executed through the resource control interface of the operating system kernel, and the resource scheduler obtains the real-time resource usage data of the task through the kernel-mode data acquisition component.

[0021] Compared with existing technologies, this dynamic scheduling method for cloud computing resource pools has the following advantages:

[0022] I. This invention adopts a binary resource allocation model that combines formal guaranteed quotas and resource credit limits. Based on the average load value of the sliding window, it realizes the dynamic conversion between the two types of quotas. Under the premise of keeping the upper limit of task resource usage constant, it can dynamically adjust the proportion of dedicated guarantee resources and shared buffer resources according to the actual load level of the task, thereby effectively alleviating the resource idle problem in the static full allocation mode. At the same time, it relies on pre-configured shared buffer resources to quickly take over the load growth, make up for the adjustment delay defects of traditional dynamic scaling, and take into account both the improvement of resource utilization efficiency and the guarantee of business operation performance.

[0023] Second, by setting tiered downgrade judgment rules and adjusting cooling cycles, combined with node-level resource credit quota over-selling control and overload priority recovery mechanism, this invention can filter out erroneous adjustments caused by short-term load fluctuations, avoid system operation oscillations caused by frequent quota changes, thereby improving the stability of the scheduling process. At the same time, it controls the overall over-selling level of nodes within a reasonable range, and orderly recovers shared resources according to task priority when nodes are overloaded, ensuring the stable operation of high-priority services and enhancing the reliability and fault tolerance of the overall resource pool scheduling.

[0024] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0026] Figure 1 This is a flowchart of the initial resource quota configuration process for this invention;

[0027] Figure 2 This is a flowchart illustrating the dynamic adjustment of quota levels during the operation of this invention.

[0028] Figure 3 This is a flowchart of the node resource overselling control and overload emergency handling process of the present invention. Detailed Implementation

[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0030] As attached Figure 1-3 As shown, this invention discloses a dynamic scheduling method for a cloud computing resource pool, applied to a cloud computing resource pool containing multiple physical nodes. It is executed by a resource scheduler deployed on a scheduling node. The resource scheduler is connected to each physical node through internal network communication and can call the resource control interface of the physical node's operating system kernel to perform resource quota configuration and adjustment. It can obtain real-time resource operation data of each task through a kernel-mode data acquisition component deployed on the physical node. The method includes three core steps: receiving resource request information submitted by the task, selecting the target physical node based on the resource request information, and allocating corresponding computing resources to the target physical node.

[0031] I. Initial Resource Quota Configuration

[0032] When a task is deployed to the target physical node, the resource scheduler configures two types of resource quotas for the task: a formal guaranteed quota and a resource credit quota.

[0033] Specifically, the formal guaranteed minimum quota is set as a resource exclusively for the task, with physical nodes reserving a dedicated share of resources for this task. This resource will not be preempted by other tasks during the task's runtime, ensuring the task's basic operational performance. The resource credit limit is set as a shared buffer resource that the task can temporarily use. It belongs to the node's globally shared resource pool. Tasks can temporarily occupy this resource when the load increases, but it does not have long-term exclusive rights and can be uniformly scheduled and reclaimed by the resource scheduler based on the overall node load status. The upper limit of resource usage for a task is set as the sum of the formal guaranteed minimum quota and the resource credit limit, remaining stable throughout the task's entire lifecycle. In some optional implementations, the initial formal guaranteed minimum quota value is calculated by the resource scheduler based on the total amount of resource requests submitted by the task and the baseline allocation ratio corresponding to the task type, while the initial resource credit limit value is the difference between the total amount of resource requests and the initial formal guaranteed minimum quota.

[0034] The initial guaranteed minimum quota is calculated using the following formula:

[0035]

[0036] In the formula, This indicates the initial official minimum allocation for the task, and the unit is consistent with the total amount of resources requested. This indicates the total amount of resource requests submitted for the task, which is reported by the task submitter during the resource request phase. This represents the baseline allocation ratio for the corresponding task type. It is a preset dimensionless coefficient. Different task types correspond to different baseline allocation ratios, which can be configured and adjusted based on historical business load data.

[0037] For example, task types can be divided into online service tasks, offline computing tasks, batch processing tasks, etc. Online service tasks have high requirements for operational stability and can be configured with a higher baseline allocation ratio; offline computing tasks have lower sensitivity to latency and can be configured with a lower baseline allocation ratio to improve the resource sharing ratio.

[0038] II. Load Acquisition and Sliding Window Calculation

[0039] During task execution, the resource scheduler periodically collects actual resource usage data for the task and generates a sliding window average load value, which serves as the core basis for determining quota adjustments.

[0040] Specifically, the resource scheduler periodically obtains the real-time resource usage of tasks from the kernel-mode data acquisition components of physical nodes. It then calculates the arithmetic average of the values ​​from multiple consecutive acquisition periods as the sliding window average load value. The sliding window moves forward with each acquisition period, always reflecting the average load level of the task within the most recent period, effectively filtering out interference from short-term load spikes. The sliding window average load value is calculated using the following formula:

[0041]

[0042] In the formula, This represents the average load value of the sliding window, with the same unit as the official minimum allowance. This represents the total number of acquisition cycles contained in the sliding window, and is a preset positive integer. This represents the actual resource usage of the task within the i-th collection period, obtained directly from the operating system kernel by the kernel-mode collection component. In some optional implementations, the upgrade operation is triggered by a continuous judgment mechanism. When the average load value of the sliding window is greater than the current guaranteed minimum quota multiplied by a preset upgrade coefficient for multiple consecutive judgment periods, it is determined that the task load is continuously increasing, triggering an upgrade operation. The adjustment step size for each upgrade is a preset proportion of the total resource request amount, and the step size can be configured according to the scheduling accuracy requirements. After the upgrade operation is completed, a cooling-off period of corresponding duration is initiated. During the cooling-off period, upgrade operations in the same direction are not triggered repeatedly to avoid frequent adjustments in a short period of time that could cause system resource fluctuations.

[0043] III. Quota Upgrade Adjustment Mechanism

[0044] When the upgrade trigger condition is met, the resource scheduler converts the corresponding share of resource credit into a formal guaranteed quota according to a preset step size, and simultaneously reduces the resource credit value to keep the upper limit of resource usage unchanged.

[0045] Specifically, in a single upgrade operation, the increase in the official guaranteed quota is equal to the preset adjustment step size, and the decrease in the resource credit limit is exactly equal to the increase in the official guaranteed quota. After the conversion, the exclusive resource share that the task can exclusively use increases, the share of shared buffer resources that can be used temporarily decreases accordingly, and the maximum total amount of resources that the task can use remains at the original level.

[0046] Understandably, this upgrade mechanism gradually confirms the shared resources actually continuously occupied by the task as dedicated resources without changing the upper limit of resources perceived by the business, thereby improving the resource guarantee level of the task and avoiding interference with business operations caused by changes in the upper limit of resources.

[0047] IV. Quota Downgrading Adjustment Mechanism

[0048] When the average load value of the sliding window continues to be lower than the preset downgrade threshold of the current official minimum quota, the resource scheduler performs a tiered downgrade operation. First, it reduces the resource credit limit value, then lowers the official minimum quota value by a preset step size, and simultaneously replenishes the resource credit limit value to keep the upper limit of resource usage stable until the official minimum quota falls back to the initial value.

[0049] Specifically, the downgrade operation consists of two phases. The first phase is the credit limit reduction phase. When the average load value of the sliding window remains below the preset first downgrade threshold for a sustained period, the resource credit limit is reduced to the current guaranteed minimum quota multiplied by a preset buffer ratio, retaining only a small amount of buffer resources to cope with minor load fluctuations. The second phase is the guaranteed minimum quota adjustment phase. When the load remains below the preset second downgrade threshold for a sustained period, the guaranteed minimum quota is adjusted downwards. Each adjustment step is a preset percentage of the total resource application volume. After the adjustment, an equivalent amount is simultaneously added to the resource credit limit to maintain a stable resource usage ceiling. This guaranteed minimum quota adjustment operation is repeated until the guaranteed minimum quota returns to its initial value.

[0050] In some optional implementations, the second downgrade threshold is lower than the first downgrade threshold, and the second duration is longer than the first duration. Through the design of graded judgment and gradual downgrade, the probability of false downgrade caused by short-term load fluctuations is reduced, and the service friendliness of the scheduling process is improved.

[0051] V. Control over over-selling of resource credit limits

[0052] Resource credit limits are managed globally using an over-sale mechanism. The resource scheduler calculates the total resource credit limits of all tasks on a single physical node, and combines this with the total remaining available resources of the node to calculate the over-sale coefficient, keeping the over-sale coefficient within a preset range.

[0053] The node resource overselling coefficient is calculated using the following formula:

[0054]

[0055] In the formula, This represents the resource oversupply coefficient of the target physical node, and is a dimensionless value. This represents the total resource credit limit of all deployed tasks on the target physical node. This represents the total remaining available resources of the target physical node, which is the remaining resource share after subtracting the total official minimum quota for all tasks from the node's total resource amount.

[0056] For example, the preset range of the overselling coefficient can be configured based on the node hardware reliability and business fault tolerance level. When the overselling coefficient is lower than the lower limit of the preset range, the initial resource credit limit ratio of newly deployed tasks can be appropriately increased to improve resource sharing. When the overselling coefficient is higher than the upper limit of the preset range, the resource credit limit configuration of newly deployed tasks can be reduced to control the overall overselling risk of the node.

[0057] When the overall load of a node exceeds a preset load threshold, the resource scheduler stops granting resource credits to newly deployed tasks and gradually reduces the resource credit values ​​of already deployed tasks to release shared buffer resources and ensure stable node operation.

[0058] VI. Differentiated scheduling of multiple resource types

[0059] In some alternative implementations, computing resources include CPU resources and memory resources. Separate official minimum quotas and resource credit limits are set for CPU resources and memory resources respectively, and upgrade and downgrade operations are performed separately. The scheduling processes of the two types of resources do not interfere with each other.

[0060] To address memory resource downgrading, the resource scheduler adds a pre-verification step. Before executing the formal minimum quota reduction, the actual memory usage of the task is verified. The formal minimum quota reduction is only executed if the actual memory usage is lower than the preset percentage of the target minimum quota after the reduction. If the actual memory usage does not meet the verification condition, the downgrading operation is paused, and load changes are monitored again. This design avoids memory shortage anomalies caused by passive memory resource reclamation, ensuring the operational stability of memory-intensive tasks.

[0061] In some optional implementations, computing resources also include heterogeneous computing power resources. For heterogeneous computing power resources, a formal guaranteed minimum quota corresponds to a dedicated computing power shard, and a resource credit quota corresponds to a preemptible computing power shard. A dedicated computing power shard is exclusively used by a task and is not scheduled or reclaimed during its runtime, ensuring the task's basic computing power needs. Preemptible computing power shards belong to the node's globally shared computing power pool and can be scheduled and reclaimed based on task priority.

[0062] When there is a computing power gap for high-priority tasks, the resource scheduler prioritizes reclaiming preemptible computing power fragments from low-priority tasks and allocating them to high-priority tasks. After the preemptible computing power fragments of low-priority tasks are reclaimed, they retain their own dedicated computing power fragments to maintain basic operation, and there will be no task interruption or exit.

[0063] VII. Emergency Handling of Node Overload

[0064] When the overall resource load of the target physical node reaches the overload warning threshold, the resource scheduler triggers an emergency resource reclamation process.

[0065] Specifically, the resource scheduler obtains the priority information of all running tasks on the node, sorts them in ascending order of priority, and reclaims all resource credits for the corresponding tasks in turn, retaining only the formal minimum quota for the tasks. After the resource credits of each task are reclaimed, the overall load status of the node is checked. If the node load falls back to a safe range, the subsequent reclamation operation is stopped.

[0066] Tasks whose resource credits have been revoked are marked as pending migration, and the resource scheduler asynchronously performs migration scheduling in the background. The resource scheduler selects physical nodes with sufficient resources from the resource pool as migration targets, schedules the tasks to be migrated to the target nodes, and restores the resource credit configuration of the task after the migration is completed.

[0067] 8. Adjusting Cooling and Handling Sudden Peak Performance

[0068] After each upgrade or downgrade operation, the resource scheduler sets a corresponding cooldown period. During the cooldown period, no quota adjustment operations in the same direction are triggered; only load data is continuously collected. This avoids frequent adjustments in a short period that could cause system oscillations and improves the operational stability of the scheduling system.

[0069] For sudden peak load scenarios, a rapid response channel is set up. When the real-time resource utilization rate of a task continuously reaches the resource utilization limit and continues for a preset duration, it is determined to be a sudden peak load scenario. The step-by-step adjustment process is skipped, and the formal minimum quota is directly increased to a value that matches the current actual load. At the same time, the corresponding share of resource credit is reduced to ensure the peak performance needs of the business are met.

[0070] IX. Low-level execution and data acquisition

[0071] The adjustment of the formal guaranteed quota and resource credit limit is executed through the resource control interface of the operating system kernel. The resource scheduler issues quota adjustment instructions to the target physical node, and the node's operating system kernel directly updates the resource limit parameters of the task, ensuring the real-time performance and accuracy of the quota adjustment.

[0072] Real-time resource usage data for tasks is acquired by a kernel-mode data acquisition component deployed on physical nodes. This component runs at the operating system kernel level, directly reading task resource statistics from the kernel and reporting them to the resource scheduler. Compared to user-mode acquisition, kernel-mode acquisition offers lower performance overhead and higher data accuracy, providing reliable data support for scheduling decisions.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A dynamic scheduling method for a cloud computing resource pool, applied to a cloud computing resource pool containing multiple physical nodes, executed by a resource scheduler deployed on a scheduling node, comprising the steps of receiving resource request information submitted by a task, selecting a target physical node based on the resource request information, and allocating corresponding computing resources to the target physical node, characterized in that, The step of allocating corresponding computing resources to the target physical node specifically includes: When a task is deployed to a target physical node, the resource scheduler configures a formal guaranteed quota and a resource credit quota for the task. The formal guaranteed quota is set as the resource exclusively occupied by the task, the resource credit quota is set as the shared buffer resource that the task can temporarily use, and the upper limit of the task's resource usage is set as the sum of the formal guaranteed quota and the resource credit quota. During task execution, the resource scheduler periodically collects actual resource usage data of the task and generates a sliding window average load value. When the sliding window average load value continuously exceeds the preset upgrade threshold of the current formal minimum quota, the corresponding share of resource credit quota is converted into the formal minimum quota according to the preset step size, and the resource credit quota value is reduced simultaneously to keep the upper limit of resource usage unchanged. When the average load value of the sliding window continues to be lower than the preset downgrade threshold of the current official minimum quota, the resource credit limit value is reduced first, and then the official minimum quota value is reduced by a preset step size. At the same time, the resource credit limit value is replenished to keep the upper limit of resource usage stable until the official minimum quota falls back to the initial value.

2. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, In the process of configuring the formal guaranteed quota and resource credit limit for the task, the resource scheduler calculates the initial formal guaranteed quota value based on the total amount of resource requests submitted by the task and the benchmark allocation ratio corresponding to the task type. The initial resource credit limit value is the difference between the total amount of resource requests and the initial formal guaranteed quota. The initial guaranteed minimum quota is calculated using the following formula: In the formula, This indicates the initial official minimum allocation for the task; This indicates the total amount of resources requested by the task. This indicates the baseline allocation ratio for the corresponding task type.

3. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, In the process of generating the sliding window average load value, the resource scheduler collects the real-time resource utilization rate of the task at a fixed period, and calculates the arithmetic average of the values ​​of multiple consecutive collection periods as the sliding window average load value. The average load value of the sliding window is calculated using the following formula: In the formula, This represents the average load value of the sliding window; This indicates the total number of acquisition cycles contained within the sliding window; This represents the real-time resource utilization rate of the task within the i-th collection cycle; When the average load value of the sliding window is greater than the current guaranteed minimum quota multiplied by the preset upgrade coefficient in multiple consecutive judgment periods, the upgrade operation is triggered. The adjustment step size of each upgrade is a preset proportion of the total resource application amount. After the upgrade, a cooldown period is set, and the upgrade operation will not be triggered again during the cooldown period.

4. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, During the process of reducing the resource credit limit and lowering the official minimum quota, the resource scheduler first performs a credit limit reduction operation. When the average load value of the sliding window is continuously lower than the preset first tier reduction threshold and reaches the first duration, the resource credit limit is reduced to the current official minimum quota multiplied by a preset buffer ratio. When the load is continuously lower than the preset second tier reduction threshold and reaches the second duration, the official minimum quota is lowered. The step size of each reduction is a preset ratio of the total resource application amount. After the reduction, the corresponding resource credit limit is replenished simultaneously to keep the upper limit of resource usage stable.

5. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, The resource credit limit is managed globally using an over-sale mechanism. The resource scheduler calculates the total resource credit limit of all tasks on the target physical node, and calculates the over-sale coefficient based on the total remaining available resources of the node, keeping the over-sale coefficient within a preset range. The node resource overselling coefficient is calculated using the following formula: In the formula, This represents the resource overselling coefficient of the target physical node; This represents the total resource credit limit for all tasks on the target physical node; This represents the total remaining available resources of the target physical node. When the overall load of the node exceeds the preset load threshold, the resource scheduler stops granting resource credits to newly deployed tasks and gradually reduces the resource credit values ​​of already deployed tasks to ensure stable node operation.

6. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, The computing resources include CPU resources and memory resources. Independent official minimum quotas and resource credit limits are set for CPU resources and memory resources respectively, and upgrade and downgrade operations are performed respectively. For memory resource downgrading operations, the resource scheduler first verifies the actual memory usage of the task. Only when the actual memory usage is lower than the preset ratio of the official minimum quota after the target downgrade, will the official minimum quota downgrade operation be executed.

7. A dynamic scheduling method for cloud computing resource pools according to claim 6, characterized in that, The computing resources also include heterogeneous computing power resources. For the formal guaranteed quota of heterogeneous computing power resources, there are dedicated computing power shards, and the resource credit quota corresponds to preemptible computing power shards. When a high-priority task has computing power requirements, the resource scheduler will prioritize reclaiming the preemptible computing power shards of low-priority tasks and allocating them to high-priority tasks. Low-priority tasks retain their dedicated computing power shards to maintain basic operation.

8. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, When the overall resource load of the target physical node reaches the overload warning threshold, the resource scheduler will reclaim the resource credit quota of the corresponding tasks in order of task priority from low to high, and retain only the formal minimum quota of the task, until the overall resource load of the node falls back to the safe range; the tasks whose resource credit quota has been reclaimed are marked as pending migration and will be asynchronously scheduled by the resource scheduler to migrate to physical nodes with sufficient resources.

9. A dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, After each upgrade and downgrade operation, the resource scheduler sets a corresponding cooldown period. During the cooldown period, quota adjustment operations in the same direction are not triggered. For sudden peak load scenarios, when the real-time resource utilization rate of a task continuously reaches the resource utilization limit and exceeds the preset time, the tiered adjustment process is skipped, and the official minimum quota is directly increased to a value that matches the current actual load.

10. A dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that, The adjustment of the formal guaranteed quota and resource credit limit is executed through the resource control interface of the operating system kernel. The resource scheduler obtains the real-time resource usage data of the task through the kernel-mode data acquisition component.

Images