A resource recycling method and system of a cloud computing power server
By collecting and analyzing the resource utilization of cloud computing servers in real time, and combining dynamic judgment and optimal time slot strategy, the problems of resource waste and false positives caused by static thresholds are solved, and efficient release of cloud computing power and seamless business transfer are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-21
- Publication Date
- 2026-06-16
Smart Images
Figure CN122220073A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cloud computing resource scheduling, and in particular to a method and system for resource recycling of a cloud computing server. Background Technology
[0002] With the large-scale deployment of cloud computing and massive applications, data centers need to dynamically and elastically schedule a large number of computing server nodes. This requires the scheduling system to monitor the underlying hardware operation indicators in real time, accurately identify idle nodes when tasks are completed or suspended, and trigger recycling operations so that the released computing power can be redistributed to waiting high-priority tasks, thereby improving the overall throughput of the cluster and reducing the operating costs of the data center.
[0003] However, existing recycling methods mostly rely on static thresholds and fixed waiting time windows for judgment. This mechanism lacks flexibility when facing complex and ever-changing cloud workloads, and is very likely to cause obviously idle nodes to waste computing power due to waiting window periods, or to mistakenly trigger recycling due to short-term load fluctuations, thereby killing normal critical business. Summary of the Invention
[0004] To overcome the above shortcomings, this invention provides a resource recycling method and system for cloud computing servers, aiming to improve the problem of computing power waste caused by existing static thresholds and fixed window mechanisms.
[0005] In a first aspect, the present invention provides the following technical solution: a resource recycling method for a cloud computing server, comprising: S1. Collect resource utilization data sequences of computing server nodes at a preset frequency, and calculate the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequences. S2. Compare the log-likelihood ratio with a preset decision upper boundary. If the log-likelihood ratio is greater than the decision upper boundary, then the corresponding node is determined to be a node to be recycled. S3. For the node to be recycled, the time within the preset time window after the judgment is discretized into multiple consecutive time slots, and the expected benefit value of performing the recycling operation in each time slot is calculated. The expected benefit value is the value obtained by subtracting the task migration cost value and the task interruption penalty value from the resource release benefit value. S4. For the multiple consecutive time slots, compare the expected benefit value of each time slot with the expected benefit value of waiting until the next time slot to perform the recycling operation, and determine the time slot with the largest expected benefit value as the target execution time slot. S5. During the target execution time slot, perform the recycling operation, including task migration, on the node to be recycled, and update the computing resource parameters released by the node to be recycled to the global available resource pool.
[0006] Preferably, in step S1, the step of collecting resource utilization data sequences of computing server nodes at a preset frequency includes: Deploy monitoring probes to synchronously collect underlying multi-dimensional hardware status indicators, including CPU utilization, GPU memory usage, and memory bandwidth. The collected multidimensional hardware status indicators are denoised and smoothed using filtering operators to remove transient abnormal peaks and obtain stable time-domain indicator data. The denoised time-domain index data are aligned by timestamps and spliced together by dimensions to construct a discrete time matrix to generate the resource utilization data sequence.
[0007] Preferably, in step S1, the step of calculating the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequence includes: Construct a probabilistic model and define the normal operation assumptions that represent nodes maintaining high load, and the idle state assumptions that represent nodes being suspended under low load. Using the probability density function, calculate the conditional probability values of the resource utilization data at the current sampling time under the assumptions of normal operation and idle state, respectively; By combining the sequential test algorithm, the logarithm of the ratio of the conditional probability value under the idle state assumption to the conditional probability value under the normal operation state assumption is calculated, and then superimposed with the historical cumulative value to generate the log-likelihood ratio.
[0008] Preferably, in step S2, the step of comparing the log-likelihood ratio with a preset upper decision boundary includes: Obtain the preset system misjudgment tolerance rate and the missed release tolerance rate, and map them to the decision upper boundary and decision lower boundary respectively through the logarithmic transformation formula; The log-likelihood ratio is extracted in each sampling period and compared with the decision upper boundary. If an upward overshoot occurs, the corresponding node is determined as the node to be recycled. If the log-likelihood ratio is between the upper and lower decision boundaries, then adaptive dynamic sampling observation is maintained until either boundary is exceeded.
[0009] Preferably, in step S3, the step of discretizing the time within the predetermined time window after determination into multiple consecutive time slots includes: Extract the scheduler clock frequency and the average task arrival rate to determine the minimum time granularity for capturing cluster state abrupt changes; Based on the minimum time granularity, the future time span from when a node is determined to be recycled to the preset mandatory timeout period is divided into the multiple consecutive time slots; The current network congestion index and the degree of computing power fragmentation are mapped to the environmental state matrix under each time slot, and a discretized inference space of the Markov decision process is constructed.
[0010] Preferably, in step S3, the step of calculating the expected revenue value of performing the recovery operation in each time slot includes: Evaluate the compatibility between the high-priority tasks to be processed and the computing power to be released from the nodes to be reclaimed, and convert it into the resource release benefit value. By combining cross-node bandwidth margin and the number of dirty pages in node memory, the task migration cost and the task interruption penalty for violating the service agreement are quantitatively calculated. Using a linear summation operator, the expected profit value is output by subtracting the task migration cost value and the task interruption penalty value from the resource release profit value.
[0011] Preferably, in step S4, the step of sequentially comparing the expected return value of each time slot with the expected return value of waiting until the next time slot to perform the recovery operation includes: A dynamic value function is constructed based on the Bellman equation, defining the total value of the current time slot system as the maximum value between immediate execution and continued waiting; The reverse derivation algorithm is used to perform reverse recursion from the end of the preset time window to solve for the expected mathematical benefit of continuing to wait until the next time slot; Perform a numerical comparison calculation. When the expected benefit value for immediate execution is greater than the expected mathematical benefit value for continued waiting, lock the current time slot as the target execution time slot.
[0012] Preferably, in step S5, the step of performing the recycling operation, including task migration, on the node to be recycled includes: Issue a scheduling isolation command to cut off the new task route of the node to be reclaimed, and simultaneously capture a snapshot of the memory context of the existing tasks within the node; Establish a data transmission channel for the target node, use a hot migration protocol to transfer the memory context snapshot and restore business operation, thereby completing the task migration; The process of the nodes to be recycled is terminated and the environment is reset to clean up the remaining routing and storage mount points in order to release the underlying computing resources.
[0013] Secondly, the present invention provides the following technical solution: a resource recycling system for a cloud computing server, comprising: The data acquisition and calculation module is used to acquire resource utilization data sequences of computing server nodes at a preset frequency, and calculate the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequences. The state determination module is used to compare the log-likelihood ratio with a preset decision upper boundary. If the log-likelihood ratio is greater than the decision upper boundary, the corresponding node is determined to be a node to be recycled. The benefit assessment module is used to discretize the time within the preset time window after the judgment of the node to be recycled into multiple consecutive time slots, and calculate the expected benefit value of performing the recycling operation in each time slot. The expected benefit value is the value obtained by subtracting the task migration cost value and the task interruption penalty value from the resource release benefit value. The timing determination module is used to compare the expected benefit value of each time slot with the expected benefit value of waiting until the next time slot to perform the recycling operation for the multiple consecutive time slots, and determine the time slot with the largest expected benefit value as the target execution time slot; The execution and update module is used to perform the recycling operation, including task migration, on the node to be recycled within the target execution time slot, and update the computing resource parameters released by the node to be recycled to the global available resource pool.
[0014] The present invention has the following beneficial effects: 1. In this invention, a sequential testing algorithm is introduced to break the traditional fixed time window limitation. By accumulating the log-likelihood ratio of hardware features in real time and comparing it with the dynamic boundary, the system can quickly identify obviously idle nodes and trigger their reclamation. At the same time, the observation period for fluctuating nodes is automatically extended, thereby eliminating the idle time of computing power while ensuring that the business is not mistakenly killed.
[0015] 2. In this invention, resource release is transformed into an optimal stopping time solution process based on Markov decision. By constructing a dynamic value function that considers release benefits, migration costs and default penalties and performing reverse derivation, the scheduler can capture the target execution slot that maximizes the net utility of the system, effectively avoiding cluster jitter caused by blind migration during peak network periods.
[0016] 3. In this invention, after locking the execution time slot, a software and hardware collaborative cleanup strategy based on memory snapshot and hot migration protocol is adopted. By seamlessly transferring the existing business to the target node while cutting off the route, the continuity of the user's critical business during the termination of the underlying process and the reset of the environment is ensured, and the lossless release and closed-loop flow of high-value cloud computing power are realized. Attached Figure Description
[0017] Figure 1 This is a flowchart of a resource recycling method for a cloud computing server proposed in this invention; Figure 2 This is an architecture diagram of a resource recycling system for a cloud computing server proposed in this invention. Detailed Implementation
[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Example 1 In a first embodiment of the present invention, the present invention provides a resource reclamation method for a cloud computing server, such as... Figure 1 As shown, it includes: S1. Collect resource utilization data sequences of computing server nodes at a preset frequency, and calculate the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequences. Further, step S1, the step of collecting resource utilization data sequences of computing server nodes at a preset frequency, includes: Deploy monitoring probes to synchronously collect underlying multi-dimensional hardware status indicators, including CPU utilization, GPU memory usage, and memory bandwidth. The collected multidimensional hardware status indicators are denoised and smoothed using filtering operators to remove transient abnormal peaks and obtain stable time-domain indicator data. The denoised time-domain index data are aligned by timestamps and spliced together by dimensions to construct a discrete time matrix to generate a resource utilization data sequence.
[0020] Further, in step S1, the step of calculating the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequence includes: Construct a probabilistic model and define the normal operation assumptions that represent nodes maintaining high load, and the idle state assumptions that represent nodes being suspended under low load. Using the probability density function, calculate the conditional probability values of the resource utilization data at the current sampling time under the assumptions of normal operation and idle state, respectively. By combining the sequential test algorithm, the logarithm of the ratio of the conditional probability value under the idle state assumption to the conditional probability value under the normal operation state assumption is calculated, and then superimposed with the historical cumulative value to generate the log-likelihood ratio.
[0021] Specifically, when executing the step of collecting resource utilization data sequences of computing server nodes at a preset frequency, hardware resource monitoring probes are first deployed at the underlying host operating system level of each computing server node. The monitoring probes synchronously read and extract underlying multi-dimensional hardware status indicators, including CPU utilization, GPU memory usage, and memory bandwidth, according to a set sampling period. Due to transient concurrent requests or network congestion in cloud services, directly collected underlying indicator data often contains spikes and sudden peaks. Therefore, the system uses a sliding median filter operator to denoise and smooth the collected multi-dimensional hardware status indicators. Specifically, a sliding window length is set for any one-dimensional hardware status indicator sequence, and the median of all sampled data within this continuous window is taken as the current filtered output value, thereby filtering out abnormal peaks caused by transient kernel interruptions and obtaining stable time-domain indicator data. Subsequently, the denoised time-domain index data is aligned in the time domain using a globally unified network time benchmark, and the multi-dimensional index features under the same timestamp are combined by column and dimensionally concatenated to construct a discrete time matrix that reflects the dynamic evolution characteristics of the underlying hardware. Finally, a resource utilization data sequence is generated as the input for subsequent hypothesis testing.
[0022] After acquiring the resource utilization data sequence, the system calculates the log-likelihood ratio of the corresponding node at the current sampling time based on this data sequence. First, a probabilistic model is constructed. Based on the large-sample statistical patterns of historical cluster load, assumptions are defined to represent the normal operation state of a node maintaining high-load computing tasks and the idle state of a node with low load and no actual task processing. Using a Gaussian probability density function, the conditional probability values of the resource utilization data at the current sampling time are calculated under the normal operation state assumption and the idle state assumption, respectively. Let the multidimensional resource utilization data vector at the current sampling time be... Normal operating condition is assumed to be The idle state is assumed to be Assuming normal operating conditions Below, the data vector follows the mean vector as follows: And the covariance matrix is The multidimensional Gaussian distribution is assumed in the idle state. Below, the data vector follows the mean vector as follows: And the covariance matrix is The data vector at the current sampling time is a multidimensional Gaussian distribution. Under the state assumption conditional probability value The following probability density function formula is used to calculate: ; in The value is or To correspond to the two hypothetical states mentioned above, The number of feature dimensions for resource utilization data. Let be the determinant of the covariance matrix. It is the inverse of the covariance matrix. This is the transpose operation of the vector difference. Through the above operation, the conditional probability values under the idle state assumption and the normal operation state assumption can be obtained respectively. Combining the sequential test algorithm, the natural logarithm of the ratio of the conditional probability value under the idle state assumption to the conditional probability value under the normal operation state assumption is calculated, and this ratio is accumulated with the historical cumulative log-likelihood ratio of the previous sampling time to generate the log-likelihood ratio of the current sampling time. The specific superposition formula is as follows: ; in This is the historical cumulative value saved from the previous sampling time. This is the natural logarithm operator, representing the historical cumulative value at the initial time. Set the value to zero.
[0023] Through the above implementation steps, the system achieves adaptive feature extraction and state probability mapping of underlying multi-dimensional hardware resources. This approach eliminates the interference of transient abnormal data and can accurately quantify the nonlinear state evolution trend of computing nodes from active to idle, providing high-fidelity data input for subsequent dynamic resource reclamation determination without fixed window limitations.
[0024] S2. Compare the log-likelihood ratio with the preset decision upper boundary. If the log-likelihood ratio is greater than the decision upper boundary, then the corresponding node is determined to be a node to be recycled. Further, step S2, the step of comparing the log-likelihood ratio with a preset upper decision boundary, includes: Obtain the preset system misjudgment tolerance rate and omission tolerance rate, and map them to the upper and lower decision boundaries respectively through a logarithmic transformation formula; The log-likelihood ratio is extracted in each sampling period and compared with the decision upper boundary. If an upward overshoot occurs, the corresponding node is determined to be a node to be recycled. If the log-likelihood ratio lies between the upper and lower decision boundaries, then adaptive dynamic sampling observations are maintained until either boundary is breached.
[0025] Specifically, after obtaining the log-likelihood ratio at the current sampling moment, the system needs to compare it with preset dynamic boundary conditions to make resource scheduling decisions. The system first obtains the system's misjudgment tolerance rate and missed release tolerance rate preset by the cluster global scheduler. In the cloud computing server resource reclamation scenario, the misjudgment tolerance rate specifically represents the maximum acceptable probability limit of the system incorrectly classifying a computing node that is actually in a high-load, normally operating state as an idle state, while the missed release tolerance rate specifically represents the maximum acceptable probability limit of the system incorrectly classifying a computing node that is actually in a truly idle state as a normally operating state and not reclamating it. The system maps the above two business-level tolerance rate parameters to the upper and lower decision boundaries in the statistical test space using a logarithmic transformation formula. Let the system misjudgment tolerance rate be... The tolerance rate for leaks is The upper boundary of decision-making is The lower boundary of the decision is The specific formula for the logarithmic transformation boundary mapping is as follows: ; ; in This is the natural logarithm operator. Through the above mathematical transformation, the system converts the abstract probabilistic constraints on scheduling security at the business level into threshold constants that can be directly used for judgment and comparison in the numerical space.
[0026] After completing the boundary constant mapping, the system extracts the accumulated log-likelihood ratio in real time within each hardware indicator sampling period and compares it with the aforementioned upper and lower decision boundaries. Let the log-likelihood ratio extracted at the current sampling time be... The system executes branch routing logic based on the numerical relationships. If an upward bounds out-of-bounds situation occurs, i.e., the current log-likelihood ratio... Strictness is greater than the upper limit of decision-making. This indicates that the accumulated data evidence showing the current node is in an idle state has reached a level that meets the stringent tolerance threshold for false positives. At this point, the system directly blocks subsequent sampling and verification processes, immediately classifies the corresponding node as a node to be reclaimed, and outputs a state change signal to the upper-layer orchestration system to trigger the subsequent resource eviction process. If a comparison reveals that the log-likelihood ratio at the current moment... Strictly located at the upper boundary of decision-making With decision lower boundary Between, that is, satisfying the algebraic inequality This indicates that the currently collected resource utilization data is insufficient to confirm either absolute idleness or absolute activity. At this point, the system maintains an adaptive dynamic sampling and observation state, the scheduler does not intervene in any node state changes, and instructs the underlying probes to continue collecting indicators and accumulating likelihood ratios for the next cycle, until the updated log-likelihood ratio value exceeds either the upper or lower decision boundary.
[0027] Through this implementation process, the resource scheduling system completely eliminates the rigid, fixed waiting time window settings of traditional cloud computing by utilizing a sequential statistical comparison mechanism. This judgment method rigorously guarantees cluster availability and task anti-false kill requirements at the mathematical level, and can immediately determine and release idle nodes with extremely obvious characteristics, greatly shortening the idle time of ineffective computing power.
[0028] S3. For nodes to be recycled, the time within the preset time window after the judgment is discretized into multiple consecutive time slots, and the expected benefit value of performing the recycling operation in each time slot is calculated. The expected benefit value is the resource release benefit value minus the task migration cost value and the task interruption penalty value. Furthermore, step S3, which discretizes the time within the predetermined time window after determination into multiple consecutive time slots, includes: Extract the scheduler clock frequency and the average task arrival rate to determine the minimum time granularity for capturing cluster state abrupt changes; Based on the minimum time granularity, the future time span from when a node is determined to be reclaimed to the preset mandatory timeout period is divided into multiple continuous time slots; The current network congestion index and the degree of computing power fragmentation are mapped to the environmental state matrix under each time slot, and a discretized inference space of the Markov decision process is constructed.
[0029] Further, step S3, the step of calculating the expected revenue value of performing the recovery operation in each time slot, includes: Evaluate the matching degree of high-priority tasks to be processed with the computing power to be released from the nodes to be recycled, and convert it into resource release revenue value; By combining cross-node bandwidth margin and the number of dirty pages in node memory, the task migration cost and the task interruption penalty for violating the service agreement are quantitatively calculated. Using a linear summation operator, the expected return value is output by subtracting the task migration cost and task interruption penalty from the resource release benefit value.
[0030] Specifically, after identifying the nodes to be recycled, the system needs to discretize their time within a preset future time window. The system first extracts the current clock tick frequency and the average arrival rate of the task queue from the global scheduler to determine the minimum time granularity used to capture sudden changes in cluster state. Let the scheduler clock frequency be... The average arrival rate of the task was Minimum time granularity The calculation formula is: ; Based on the determined minimum time granularity The system will determine the current moment when a node is identified as needing to be reclaimed. From the date specified in the manual timeout period until the preset mandatory timeout period. The future time span between these points is strictly divided into multiple consecutive time slots of equal length. Let the total number of discretized consecutive time slots be... The calculation formula is as follows: ; The system then collects the network congestion index and global computing power fragmentation level within the data center in real time, mapping them to the environmental state matrix of a Markov decision process in each discrete time slot. Let the... The network congestion index for each time slot is The degree of computing power fragmentation is The corresponding constructed environment state matrix for: ; This step provides a discretized numerical space for subsequent state-based optimal stopping time deduction.
[0031] After constructing the discretized derivation space, the system calculates the expected benefit value of performing a reclamation operation in each discrete time slot. The system first scans the high-priority tasks currently suspended in the scheduling queue, reads the number of logical CPU cores and GPU memory capacity requested by the tasks to be processed, calculates the ratio of this ratio to the number of physical cores and physical GPU memory capacity to be released by the node to be reclaimed, and converts this ratio into a resource release benefit value. Let the th... Within the first time slot The matching fit of the high-priority tasks is The task system weight is The resource release benefit value for that time slot for: ; The system combines the cross-node physical bandwidth margin monitored by the underlying network interface card (NIC) with the total number of dirty memory pages for existing long-lived connections on the node to be reclaimed to quantitatively calculate the network time overhead required for task hot migration as the task migration cost value. Let the cross-node bandwidth margin of the current time slot be... The total number of dirty pages in memory is The protocol conversion migration resistance coefficient is Then the task migration cost value The calculation formula is: ; Simultaneously, it is assessed that if the recovery operation results in the forced termination of existing critical tasks, the system must compensate for breach of contract as stipulated in the Service Level Agreement. Duration of business interruption Calculate the task interruption penalty value Specifically: ; The system uses a linear summation operator, treating the resource release benefit as a positive incentive, and subtracting the task migration cost and task interruption penalty as negative resistance, to output the final result. Net expected revenue from performing recovery operations per time slot : ; Through the above steps, the system transforms the complex cluster operating environment and continuous time dimension into a quantifiable discrete benefit evaluation model. This method accurately calculates the business gains brought about by the release of computing power and the system oscillation overhead caused by cross-node migration, providing a numerical decision-making basis for the scheduler to capture the optimal execution time.
[0032] S4. For multiple consecutive time slots, compare the expected revenue value of each time slot with the expected revenue value of waiting until the next time slot to perform the recycling operation, and determine the time slot with the largest expected revenue value as the target execution time slot. Further, step S4, which involves sequentially comparing the expected revenue value of each time slot with the expected revenue value of waiting until the next time slot to perform the recovery operation, includes: A dynamic value function is constructed based on the Bellman equation, defining the total value of the current time slot system as the maximum value between immediate execution and continued waiting; The reverse derivation algorithm is used to perform reverse recursion from the end of the preset time window to solve for the expected mathematical benefit of continuing to wait until the next time slot; Perform a numerical comparison calculation. If the expected benefit of immediate execution is greater than the expected mathematical benefit of continuing to wait, lock the current time slot as the target execution time slot.
[0033] Specifically, after obtaining the expected revenue values for performing the recovery operation in each consecutive time slot, the system sequentially performs optimal stopping time deduction based on a Markov decision process for the discretized consecutive time slots. The system first constructs a dynamic value function to evaluate the overall execution strategy based on the Bellman equation. This function defines that when facing a recovery decision, the total system value of the current time slot is equal to the maximum of the actual expected revenue obtained by immediately performing the recovery operation and the discounted value of the expected mathematical revenue obtained by continuing to wait and delaying execution to a subsequent time slot. Let the th... The environmental state matrix corresponding to each time slot is: The expected return value for immediate execution in this state is The time series discount factor is Dynamic value function The mathematical expression is constructed as follows: ; in This represents the environmental state matrix composed of the current network congestion index and the degree of computing power fragmentation. Next, the system transitions to the next time slot environment state matrix. The expected value of the overall mathematical return that can be obtained at any given time. Through the above value function, the scheduler incorporates the short-term resource release utility and the long-term system oscillation risk into a unified evaluation operator.
[0034] To solve the aforementioned dynamic value function and obtain the globally optimal decision, the system employs a backward derivation algorithm to perform rigorous reverse recursive calculations from the end of a preset time window. The system sets the final time slot corresponding to the preset mandatory timeout period. At this point, the scheduling system must unconditionally enforce the resource clearing operation. Therefore, the state value of the end time slot is directly assigned to the expected immediate execution benefit value of that time slot, i.e., assigned as... After these boundary conditions are established, the algorithm starts from the time slot. The calculation proceeds recursively along the reverse timeline up to the currently occurring time slot. In each step of the reverse inference process, the system retrieves historical network traffic fluctuation curves and scheduler logs recorded by the cluster's underlying monitoring module, and uses a Markov chain to fit the probability distribution matrix of the transition from the current state to the next stage state. Then, the expected mathematical return for continuing to wait until the next time slot is calculated. The corresponding mathematical formula is: ; After obtaining the expected mathematical benefit of continuing to wait in the current time slot through the above reverse derivation process, the system executes real-time numerical comparison and judgment logic. The underlying orchestration engine extracts the expected benefit of immediately executing the recycling in the current time slot. And compare it with the expected mathematical return value of continuing to wait until the next time slot, calculated from the results. Perform a scalar comparison. When the condition is satisfied... Strictly greater than This conditional branch indicates that the network congestion costs and task service agreement default risks caused by continuing to delay the recovery action have, mathematically, completely outweighed the potential buffer gains from delaying the release of computing power. The system then directly interrupts subsequent state deductions, locking the time slot mapped to the current physical time point as the absolute target execution time slot for performing computing power isolation and task migration.
[0035] This implementation step transforms the selection of the timing for cloud resource release actions into a state-based optimal solution process, avoiding the computing power avalanche caused by blindly performing data migration during periods of high network concurrency, and ensuring a smooth transition for high-density cluster reclamation tasks.
[0036] S5. During the target execution time slot, perform recycling operations, including task migration, on the node to be recycled, and update the computing resource parameters released by the node to be recycled to the global available resource pool.
[0037] Furthermore, step S5, which involves performing recycling operations, including task migration, on the node to be recycled, includes: Issue a scheduling isolation command to cut off the new task route of the node to be reclaimed, and simultaneously capture a snapshot of the memory context of the existing tasks within the node; Establish a data transmission channel for the target node, use the hot migration protocol to transfer the memory context snapshot and restore business operation to complete the task migration; The process of reclaiming nodes is terminated and the environment is reset, and residual routing and storage mount points are cleaned up to release the underlying computing resources.
[0038] Specifically, once the global scheduler enters the target execution slot, the system immediately issues a scheduling isolation command to the node to be reclaimed via the cluster orchestration engine. This isolation command modifies the status of the node to be reclaimed to unschedulable, thereby cutting off the routing of new computing task requests to the node at the network load balancing level. Simultaneously, the system calls the underlying virtual machine monitor or container runtime daemon to synchronously capture a snapshot of the memory context of existing tasks within the node. This snapshot includes the task's current process running state parameters, processor register values, and the baseline memory page data initially allocated to the task.
[0039] After isolation and snapshot generation are completed, the system selects target nodes with sufficient load from the cluster and establishes a high-speed encrypted data transmission channel between the nodes to be reclaimed and the target nodes through the underlying virtual switch. The system utilizes a hot migration protocol based on a pre-copy mechanism to begin a full transfer of the captured memory context snapshots. Since existing tasks on the nodes to be reclaimed continue to run and generate new dirty pages during the transfer, the system needs to perform multiple rounds of iterative synchronous transfer of the newly generated dirty pages. To ensure a smooth business transition, the system monitors the total amount of remaining dirty pages to be transferred and the available physical bandwidth of the current data transmission channel in real time, and calculates the estimated business downtime accordingly. Let the current total amount of remaining dirty pages be... The available physical bandwidth of the data transmission channel is Expected downtime The calculation formula is: ; When the calculated estimated business downtime When the time interruption is significantly less than the system's preset maximum tolerable service interruption threshold, the system immediately suspends task execution on the node to be reclaimed and sends the final processor state and the remaining minimal amount of dirty memory pages to the target node in one go. Subsequently, the system immediately resumes service operation on the target node based on the received complete snapshot data, thus completing the task migration seamlessly for the service.
[0040] After confirming the successful resumption of the task on the target node, the system performs a complete forced process termination operation on the node to be reclaimed, sending system termination signals to all remaining user-mode and kernel-mode processes associated with the original task. Subsequently, the system performs a low-level environment reset, thoroughly cleaning up any remaining virtual network interfaces and routing table entries within the node to be reclaimed, and forcibly unbinding the distributed block storage mount point, thus completely releasing the underlying computing resources. Finally, the system tallies the number of CPU cores, GPU memory capacity, and system physical memory capacity completely released on the node, and synchronously registers and updates these computing resource parameters to the cluster's globally available resource pool for subsequent high-priority task scheduling.
[0041] This step completed the physical removal of idle nodes without interrupting the operation of existing critical business operations, achieving lossless release of cloud computing resources and closed-loop update of the cluster's global capacity status.
[0042] Example 2 Traditional resource reclamation methods often rely on static thresholds and fixed waiting time windows for judgment. This mechanism lacks flexibility when facing complex and variable cloud workloads, easily leading to wasted computing power due to idle nodes waiting for the window, or erroneous triggering of reclamation due to short-term load fluctuations, thus mistakenly killing normal critical services. To solve the above problems, this invention provides a resource reclamation system for cloud computing servers, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this system is as follows: The data acquisition and calculation module is used to acquire resource utilization data sequences of computing server nodes at a preset frequency, and calculate the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequences. The state determination module is used to compare the log-likelihood ratio with a preset decision upper boundary. If the log-likelihood ratio is greater than the decision upper boundary, the corresponding node is determined to be a node to be recycled. The benefit assessment module is used to discretize the time within the preset time window after the judgment of the node to be recycled into multiple consecutive time slots, and calculate the expected benefit value of performing the recycling operation in each time slot. The expected benefit value is the value obtained by subtracting the task migration cost value and the task interruption penalty value from the resource release benefit value. The timing determination module is used to compare the expected benefit value of each time slot with the expected benefit value of waiting until the next time slot to perform the recycling operation for the multiple consecutive time slots, and determine the time slot with the largest expected benefit value as the target execution time slot; The execution and update module is used to perform the recycling operation, including task migration, on the node to be recycled within the target execution time slot, and update the computing resource parameters released by the node to be recycled to the global available resource pool.
[0043] Specifically, the system first acquires denoised multi-dimensional hardware indicators in real time through the data acquisition and calculation module to construct a resource utilization data sequence. It then uses a sequential test algorithm to calculate the logarithm of the ratio of the conditional probability of a node being in an idle or normal operating state at the current moment, and superimposes this ratio with historical cumulative values to generate a dynamic log-likelihood ratio. Subsequently, the state determination module abandons the traditional rigid fixed waiting window and compares this likelihood ratio with the decision upper boundary based on the system's misjudgment tolerance rate in real time. If an upward boundary is exceeded, the corresponding node is immediately determined as a node to be reclaimed. Next, the benefit evaluation module extracts the scheduler frequency parameters to determine the minimum time granularity, strictly dividing the node's future preset time span into multiple continuous discrete time slots. It then comprehensively quantifies environmental factors such as high-priority task suitability, cross-node bandwidth margin, and the number of dirty memory pages to calculate the resource release within each time slot. The net expected return value is obtained after deducting task migration costs and business interruption penalties from the released revenue. Based on this, the timing determination module constructs a dynamic value function based on the Bellman equation and uses a reverse recursive algorithm to derive the expected mathematical return value of continuing to wait until the next time slot. By comparing the difference in returns between immediate execution and continued delay in each time slot, the target execution time slot that maximizes the overall net utility of the system is accurately locked. Finally, when the target time slot is reached, the execution and update module issues an isolation command to cut off the new task route and uses a hot migration protocol to seamlessly transfer the memory snapshot of the existing critical business to the target node with sufficient load. Under the premise of ensuring smooth and uninterrupted business flow, the original node is forced to perform process termination and environment reset, and the underlying computing power parameters that have been completely freed up are re-registered back to the globally available resource pool. This achieves elastic and lossless release of high-value computing power while avoiding cluster network oscillations.
[0044] The following is a description with reference to specific embodiments: An internet company operates a large, hybrid load data center containing 5,000 computing servers. The cluster contains a mix of high-priority AI model training tasks and low-priority e-commerce recommendation microservices.
[0045] During the early morning hours when e-commerce traffic is low, the traffic on computing node A, where a certain e-commerce recommendation microservice is located, plummeted. The underlying probe collected data once per second and found that its CPU utilization dropped to 8%, and GPU memory usage was 0%. The system used a Gaussian mixture model to calculate conditional probabilities and combined this with a sequential test algorithm to continuously accumulate the log-likelihood ratio.
[0046] Traditionally, a node needs to run at low load for 15 minutes before it can be considered idle. However, in this system, because the node's idle characteristics are extremely obvious, its log-likelihood ratio broke through the upper boundary of the decision based on the extremely low leak tolerance rate after only 45 seconds of sampling and accumulation. Therefore, the system decisively identified node A as a node to be reclaimed.
[0047] At this moment, there is a high-priority AI image rendering task suspended in the global scheduling queue. The system discretizes the next 5 minutes into 300 second-level time slots. The assessment found that releasing node A would yield extremely high resource release benefits; at the same time, it detected that the current internal network bandwidth is sufficient, there are few dirty pages in memory, and the calculated task migration cost is extremely low.
[0048] The system uses the Bellman equation to deduce in reverse that if the recycling is executed immediately while the network is currently open, the net expected benefit is far greater than continuing to wait. Therefore, the system directly locks the current time slot as the target execution time slot.
[0049] The system issues an isolation command to node A to cut off external request routing and immediately establishes a data channel with the low-load node B. Using a pre-copy hot migration protocol, the remaining microservice memory snapshot on node A is seamlessly migrated to node B and restored to operation in 2.5 seconds. Subsequently, the system forcibly wipes the underlying processes and network configuration of node A, updates its 32-core CPU and GPU computing power to the global resource pool in real time, and allocates all of it to waiting AI rendering tasks within one second.
[0050] Table 1 shows a comparison of the application effects of traditional resource recycling schemes using fixed thresholds and fixed time windows for eviction with the technical solution of this invention.
[0051] Table 1 Comparison of Application Effects between Traditional Resource Recycling Solutions and the Technical Solution of This Invention Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for resource recycling of a cloud computing server, characterized in that, include: S1. Collect resource utilization data sequences of computing server nodes at a preset frequency, and calculate the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequences. S2. Compare the log-likelihood ratio with a preset decision upper boundary. If the log-likelihood ratio is greater than the decision upper boundary, then the corresponding node is determined to be a node to be recycled. S3. For the node to be recycled, the time within the preset time window after the judgment is discretized into multiple consecutive time slots, and the expected benefit value of performing the recycling operation in each time slot is calculated. The expected benefit value is the value obtained by subtracting the task migration cost value and the task interruption penalty value from the resource release benefit value. S4. For the multiple consecutive time slots, compare the expected benefit value of each time slot with the expected benefit value of waiting until the next time slot to perform the recycling operation, and determine the time slot with the largest expected benefit value as the target execution time slot. S5. During the target execution time slot, perform the recycling operation, including task migration, on the node to be recycled, and update the computing resource parameters released by the node to be recycled to the global available resource pool.
2. The resource recycling method for a cloud computing server according to claim 1, characterized in that, Step S1, the step of collecting resource utilization data sequences of computing server nodes at a preset frequency, includes: Deploy monitoring probes to synchronously collect underlying multi-dimensional hardware status indicators, including CPU utilization, GPU memory usage, and memory bandwidth. The collected multidimensional hardware status indicators are denoised and smoothed using filtering operators to remove transient abnormal peaks and obtain stable time-domain indicator data. The denoised time-domain index data are aligned by timestamps and spliced together by dimensions to construct a discrete time matrix to generate the resource utilization data sequence.
3. The resource recycling method for a cloud computing server according to claim 1, characterized in that, In step S1, the step of calculating the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequence includes: Construct a probabilistic model and define the normal operation assumptions that represent nodes maintaining high load, and the idle state assumptions that represent nodes being suspended under low load. Using the probability density function, calculate the conditional probability values of the resource utilization data at the current sampling time under the assumptions of normal operation and idle state, respectively; By combining the sequential test algorithm, the logarithm of the ratio of the conditional probability value under the idle state assumption to the conditional probability value under the normal operation state assumption is calculated, and then superimposed with the historical cumulative value to generate the log-likelihood ratio.
4. The resource recycling method for a cloud computing server according to claim 1, characterized in that, Step S2, the step of comparing the log-likelihood ratio with a preset upper decision boundary, includes: Obtain the preset system misjudgment tolerance rate and the missed release tolerance rate, and map them to the decision upper boundary and decision lower boundary respectively through the logarithmic transformation formula; The log-likelihood ratio is extracted in each sampling period and compared with the decision upper boundary. If an upward overshoot occurs, the corresponding node is determined as the node to be recycled. If the log-likelihood ratio is between the upper and lower decision boundaries, then adaptive dynamic sampling observation is maintained until either boundary is exceeded.
5. The resource recycling method for a cloud computing server according to claim 1, characterized in that, In step S3, the step of discretizing the time within the preset time window after determination into multiple consecutive time slots includes: Extract the scheduler clock frequency and the average task arrival rate to determine the minimum time granularity for capturing cluster state abrupt changes; Based on the minimum time granularity, the future time span from when a node is determined to be recycled to the preset mandatory timeout period is divided into the multiple consecutive time slots; The current network congestion index and the degree of computing power fragmentation are mapped to the environmental state matrix under each time slot, and a discretized inference space of the Markov decision process is constructed.
6. The resource recycling method for a cloud computing server according to claim 1, characterized in that, Step S3, the step of calculating the expected revenue value of performing the recovery operation in each time slot, includes: Evaluate the compatibility between the high-priority tasks to be processed and the computing power to be released from the nodes to be reclaimed, and convert it into the resource release benefit value. By combining cross-node bandwidth margin and the number of dirty pages in node memory, the task migration cost and the task interruption penalty for violating the service agreement are quantitatively calculated. Using a linear summation operator, the expected profit value is output by subtracting the task migration cost value and the task interruption penalty value from the resource release profit value.
7. The resource recycling method for a cloud computing server according to claim 1, characterized in that, Step S4, the step of sequentially comparing the expected return value of each time slot with the expected return value of waiting until the next time slot to perform the recovery operation, includes: A dynamic value function is constructed based on the Bellman equation, defining the total value of the current time slot system as the maximum value between immediate execution and continued waiting; The reverse derivation algorithm is used to perform reverse recursion from the end of the preset time window to solve for the expected mathematical benefit of continuing to wait until the next time slot; Perform a numerical comparison calculation. When the expected benefit value for immediate execution is greater than the expected mathematical benefit value for continued waiting, lock the current time slot as the target execution time slot.
8. The resource recycling method for a cloud computing server according to claim 1, characterized in that, Step S5, the step of performing the recycling operation, including task migration, on the node to be recycled includes: Issue a scheduling isolation command to cut off the new task route of the node to be reclaimed, and simultaneously capture a snapshot of the memory context of the existing tasks within the node; Establish a data transmission channel for the target node, use a hot migration protocol to transfer the memory context snapshot and restore business operation, thereby completing the task migration; The process of the nodes to be recycled is terminated and the environment is reset to clean up the remaining routing and storage mount points in order to release the underlying computing resources.
9. A resource recycling system for a cloud computing server, characterized in that, A resource recycling method for a cloud computing server according to any one of claims 1-8, the system comprising: The data acquisition and calculation module is used to acquire resource utilization data sequences of computing server nodes at a preset frequency, and calculate the log-likelihood ratio of the corresponding node at the current sampling time based on the resource utilization data sequences. The state determination module is used to compare the log-likelihood ratio with a preset decision upper boundary. If the log-likelihood ratio is greater than the decision upper boundary, the corresponding node is determined to be a node to be recycled. The benefit assessment module is used to discretize the time within the preset time window after the judgment of the node to be recycled into multiple consecutive time slots, and calculate the expected benefit value of performing the recycling operation in each time slot. The expected benefit value is the value obtained by subtracting the task migration cost value and the task interruption penalty value from the resource release benefit value. The timing determination module is used to compare the expected benefit value of each time slot with the expected benefit value of waiting until the next time slot to perform the recycling operation for the multiple consecutive time slots, and determine the time slot with the largest expected benefit value as the target execution time slot; The execution and update module is used to perform the recycling operation, including task migration, on the node to be recycled within the target execution time slot, and update the computing resource parameters released by the node to be recycled to the global available resource pool.