An intelligently designed cloud service management method and system

Abstract

This invention relates to the field of cloud service technology and provides an intelligent cloud service management method and system, comprising the following steps: collecting hardware performance counter data on each physical node, predicting the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and determining the interference fingerprint based on the number of instruction cycles per second; calculating the interference potential coefficient between each pair of applications, calculating the global co-location interference entropy of nodes, and classifying nodes into different levels based on the interference entropy; when a node is at a low or medium interference level, identifying aggressive applications, and dynamically adjusting the last-level cache quota, memory bandwidth limit, and CPU scheduling priority of aggressive applications; when a node is at a high interference level, determining candidate migration plans: selecting target nodes through a distributed node negotiation mechanism and executing application migration. This invention achieves the optimal balance between performance assurance and system overhead through a two-level strategy of local resource elastic adjustment and proactive migration.

Description

Technical Field

[0001] This invention relates to the field of cloud service technology, specifically to an intelligent cloud service management method and system. Background Technology

[0002] In cloud computing data centers, to improve the utilization of physical resources, multiple virtual machines or containers are often co-located on the same physical server. However, these co-located applications compete for shared hardware resources, especially the last-level cache (LLC) and memory bandwidth. When multiple compute-intensive and memory-intensive applications run simultaneously, they interfere with each other, causing unpredictable performance fluctuations in critical applications and making it difficult to guarantee service quality.

[0003] Existing cloud service management technologies primarily address co-location interference issues through several methods: First, static resource partitioning, such as allocating a fixed portion of cache to each application using Intel CAT technology. However, this method cannot adapt to dynamic changes in application load. Second, threshold-based interference detection, such as monitoring CPI (cycles per instruction) or cache miss rate to see if it exceeds a certain threshold. However, these methods suffer from slow response times and cannot distinguish the source and severity of interference, easily leading to false positives or false negatives. Therefore, an intelligently designed cloud service management method and system are needed to solve these problems. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent cloud service management method and system to solve the problems existing in the above-mentioned background technology.

[0005] This invention is implemented as follows: an intelligent cloud service management method, the method comprising the following steps: Hardware performance counter data is collected on each physical node to predict the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and interference fingerprints are determined by combining the number of instruction cycles per instruction. Based on the interference fingerprints of all applications on the same node, the interference potential coefficient between each pair of applications is calculated. The global co-location interference entropy of the node is calculated according to the interference potential coefficient. The node is then classified into low-interference, medium-interference, or high-interference levels according to the interference entropy. When a node is at a low or medium level of interference, identify the aggressive application and dynamically adjust its last-level cache quota, memory bandwidth limit, and CPU scheduling priority to reduce interference entropy. When a node is at a high level of interference, or when the interference entropy is still higher than the safety threshold after dynamic adjustment, a candidate migration plan is determined: the target node is selected through a distributed node negotiation mechanism, the application migration is executed, and a migration cooling-off period and random backoff mechanism are set to avoid oscillation.

[0006] As a further aspect of this invention, when calculating the global co-location interference entropy of nodes, the formula is used: Hc=- Where K is the total number of applications co-located on the node. To apply the interference potential coefficient for (a,b), The interference potential coefficient is the sum of the interference potential coefficients of all applications on the node, and Hc is the co-location interference entropy; the interference potential coefficient It is calculated by combining the predicted memory bandwidth contention intensity, predicted last-level cache contention intensity, and the mean and standard deviation of the number of instruction cycles for applications a and b.

[0007] As a further aspect of the present invention, the step of determining the interfering fingerprint specifically includes: On each physical core, hardware performance counters are read at 1-millisecond intervals to collect the number of last-level cache misses, memory bandwidth read / write bytes, number of instruction cycles per cycle, and number of remote memory access requests. Maintain a circular buffer with a length of s sampling points for each performance counter, calculate the first and second differences of each sequence, and extract the mean, variance, and peak difference over the past t1 milliseconds as feature vectors; Input the feature vector into the online gradient boosting decision tree model, and output the predicted values ​​of memory bandwidth contention intensity and last-level cache contention intensity within the next t2 milliseconds, where s, t1, and t2 are constants; Simultaneously calculate the average number of instruction cycles per instruction for the current application, and combine the predicted memory bandwidth contention intensity, the predicted last-level cache contention intensity, and the average number of instruction cycles per instruction into a three-dimensional interference fingerprint vector.

[0008] As a further aspect of the present invention, the steps of identifying aggressive applications and dynamically adjusting the last-level cache quota, memory bandwidth limit, and CPU scheduling priority of the aggressive applications specifically include: Based on the interference potential coefficient of each pair of applications, the top three pairs of applications with the greatest contribution are selected, and those with higher predicted competitive intensity are marked as aggressive applications, while those with lower predicted competitive intensity are marked as sensitive applications. For aggressive applications, the last-level cache quota is gradually reduced, with each reduction being m% of the current quota, until the co-location interference entropy drops to a preset value of one. At the same time, the cache of sensitive applications is guaranteed not to be lower than the minimum guaranteed value calculated from their historical last-level cache miss rate. Memory bandwidth allocation technology is used to set a bandwidth cap for aggressive applications. The initial cap is set at 80% of the current bandwidth usage. If the interference entropy still increases in the next cycle, it will be reduced by 10%, with a minimum of no less than 40% of the original demand. When aggressive and sensitive applications are running on the same physical core, the CPU scheduling priority of the sensitive application is increased, giving it more execution time.

[0009] As a further aspect of the present invention, the step of determining the candidate migration plan specifically includes: The application most severely affected by the current node is selected as the application to be migrated. This application is the one with the highest number of instruction cycles per instruction and the lowest predicted last-level cache contention intensity. The node to be migrated broadcasts a migration query message to its neighboring nodes. The message contains the interference fingerprint of the application to be migrated and the co-location interference entropy of the current node. When a neighboring node receives a query, it simulates and calculates its own new co-location interference entropy after accepting the application. If the new entropy value is less than the preset value of 2 and there are sufficient remaining resources, it will provide an acceptance feedback. Collect all feedback and select the target node that minimizes the sum of global co-location interference entropy as the final migration target.

[0010] As a further aspect of the present invention, the step of calculating the interference potential coefficient between each pair of applications is as follows: obtaining the predicted last-level cache contention intensity PL(a) and PL(b), the predicted memory bandwidth contention intensity PB(a) and PB(b), the mean per instruction cycle μCPI(a) and μCPI(b), and the standard deviation per instruction cycle σCPI(a) and σCPI(b) for applications a and b; calculating the common aggression factor = min(PL(a), PL(b)) × min(PB(a), PB(b)); calculating the sensitivity factor = max(μCPI(a), μCPI(b)) + 1; calculating the volatility consistency factor = 1 / (1 + |σCPI(a) - σCPI(b)|); dividing the common aggression factor by the sensitivity factor and then multiplying by the volatility consistency factor to obtain the interference potential coefficient. .

[0011] Another object of the present invention is to provide an intelligent cloud service management system, the system comprising: The interference fingerprint module is used to collect hardware performance counter data on each physical node, predict the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and determine the interference fingerprint by combining the number of instruction cycles per instruction. The interference calculation module is used to calculate the interference potential coefficient between each pair of applications based on the interference fingerprint of all applications on the same node, calculate the global co-location interference entropy of the node based on the interference potential coefficient, and classify the node into low interference, medium interference or high interference levels according to the interference entropy. The dynamic adjustment module is used to identify aggressive applications when a node is in a low or medium level of interference, and dynamically adjust the last-level cache quota, memory bandwidth limit, and CPU scheduling priority of the aggressive application to reduce interference entropy. The application migration module is used to determine candidate migration plans when a node is at a high level of interference or when the interference entropy is still higher than the safety threshold after dynamic adjustment: the target node is selected through a distributed node negotiation mechanism, the application migration is executed, and a migration cooling-off period and random backoff mechanism are set to avoid oscillation.

[0012] Compared with the prior art, the beneficial effects of the present invention are: By introducing interference fingerprints and co-located interference entropy, it is possible to distinguish whether interference is concentrated on a single pair of applications or dispersed across multiple pairs, thus providing a more accurate basis for subsequent decision-making. A two-tiered strategy of local resource elastic adjustment and proactive migration is employed. For low- to medium-level interference, local solutions are prioritized through dynamic adjustments to cache quotas, memory bandwidth limits, and CPU scheduling priorities, avoiding the performance overhead and data transfer costs associated with frequent application migrations. Migration is only triggered in cases of severe interference or ineffective suppression, achieving an optimal balance between performance guarantees and system overhead. The model predicts the resource contention intensity of applications in the near future, making interference assessment forward-looking and enabling action to be taken before or at the early stages of interference, overcoming the lag problem of traditional methods that rely solely on current or historical data. Attached Figure Description

[0013] Figure 1 A flowchart for an intelligent cloud service management method.

[0014] Figure 2 A flowchart for identifying interfering fingerprints in a cloud service management method for intelligent design.

[0015] Figure 3 This is a flowchart illustrating the dynamic adjustment of quotas in an intelligent cloud service management method.

[0016] Figure 4 This is a flowchart for determining candidate migration plans in a cloud service management method based on intelligent design.

[0017] Figure 5 A schematic diagram of the structure of an intelligent cloud service management system. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0020] like Figure 1As shown in the figure, this embodiment of the invention provides an intelligent cloud service management method, the method comprising the following steps: The S100 collects hardware performance counter data on each physical node to predict the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and determines the interference fingerprint by combining the number of instruction cycles per instruction.

[0021] In this embodiment of the invention, a quantifiable interference fingerprint is generated for each running application to describe its tendency to compete for shared resources (last-level cache and memory bandwidth). Specifically, on each physical server node, the CPU's internal hardware performance counter (PMC) is read at millisecond intervals to extract metrics related to memory access behavior, such as cache misses and memory read / write bytes. Based on this historical data, a lightweight predictive model is used to estimate the application's competition intensity for memory bandwidth and last-level cache in the next few tens of milliseconds. Simultaneously, the application's cycle per instruction (CPI) is recorded; a higher CPI indicates that the application is more constrained by memory access latency, i.e., more sensitive to interference. The predicted competition intensities and CPI values ​​are combined to form a three-dimensional vector, yielding the interference fingerprint, which serves as the basis for all subsequent interference quantification and scheduling decisions.

[0022] S200 calculates the interference potential coefficient between each pair of applications based on the interference fingerprint of all applications on the same node, calculates the global co-location interference entropy of the node based on the interference potential coefficient, and classifies the node into low interference, medium interference or high interference levels based on the interference entropy.

[0023] In this embodiment of the invention, firstly, for each pair of applications on a node, an interference potential coefficient is calculated using their interference fingerprints. The larger the coefficient, the more likely the two applications are to experience severe resource contention when co-located. When calculating the interference potential coefficient for application a and application b, the predicted last-level cache contention intensity PL(a) and PL(b), the predicted memory bandwidth contention intensity PB(a) and PB(b), the mean per instruction cycle μCPI(a) and μCPI(b), and the standard deviation per instruction cycle σCPI(a) and σCPI(b) are obtained. The co-aggression factor is calculated as min(PL(a), PL(b)) × min(PB(a), PB(b)). The larger this factor, the higher the simultaneous demand of the two applications on two key resources, and the greater the potential for resource contention when co-located. The sensitivity factor is calculated as max(μCPI(a), μCPI(b)) + 1. A larger factor indicates that at least one application has a high CPI (i.e., its performance is already poor and it is sensitive to interference). In this case, the interference potential should be amplified because the sensitive application is more vulnerable. The volatility consistency factor is calculated as 1 / (1 + |σCPI(a) - σCPI(b)|). If the CPI volatility of two applications is similar, their interference behavior is more likely to be stable, so the factor is close to 1. If the volatility difference is large, the interference behavior is more random, and the factor is smaller. Finally, the common aggression factor is divided by the sensitivity factor and then multiplied by the volatility consistency factor to obtain the interference potential coefficient. The larger the value, the higher the risk of serious performance interference when application a and application b are co-located.

[0024] Then, using the interference potential coefficients of all application pairs as input, the global co-location interference entropy of the nodes is calculated, at which point the formula: Hc = - Where K is the total number of applications co-located on the node. To apply the interference potential coefficient for (a,b), The sum of interference potential coefficients for all application pairs on the node is used as a normalization factor to ensure that the weight of each application pair is between 0 and 1. Hc is the co-location interference entropy. This entropy value reflects the complexity and severity of the interference: a low entropy value indicates that the interference is mainly concentrated on one application pair, making it easy to locate and handle; a high entropy value indicates that there is varying degrees of interference between multiple application pairs, resulting in a high overall performance risk. Finally, based on the numerical range of the entropy value, the node state is divided into three levels: low interference, medium interference, and high interference, to determine the subsequent intervention strategy.

[0025] When a node is in a low or medium interference level, the S300 identifies aggressive applications and dynamically adjusts the last-level cache quota, memory bandwidth limit, and CPU scheduling priority of the aggressive applications to reduce interference entropy.

[0026] In this embodiment of the invention, when a node is at a low or medium level of interference, the interference is not yet severe, and interference is suppressed through fine-grained resource allocation. First, based on the ranking of interference potential coefficients, the applications on the node that compete most fiercely for shared resources are identified, termed aggressive applications, and the applications most severely affected by interference are designated as sensitive applications. Then, three restrictive measures are implemented for aggressive applications: first, gradually reducing their last-level cache allocation quota; second, limiting their usable memory bandwidth limit through memory bandwidth allocation technology; and third, when aggressive and sensitive applications are running concurrently on the same CPU core, proactively increasing the scheduling priority of sensitive applications to allow them to run more. These adjustments are made dynamically, and changes in co-location interference entropy are observed in real time until the entropy value drops to an acceptable range.

[0027] S400: When a node is at a high interference level, or the interference entropy is still higher than the safety threshold after dynamic adjustment, a candidate migration plan is determined: the target node is selected through a distributed node negotiation mechanism, the application migration is executed, and a migration cooling-off period and random backoff mechanism are set to avoid oscillation.

[0028] In this embodiment of the invention, when a node is at a high interference level, or when its entropy value is still higher than a preset safety threshold after adjustment in step S300, application migration is triggered. The migration plan includes: selecting the application most severely affected by interference as the target to be migrated. Then, the source node (the node to be migrated) broadcasts a migration request to its neighboring nodes. The neighboring nodes simulate and calculate their own entropy value change if they accept the application, and provide feedback on whether it is suitable as a target. The source node selects the node that minimizes the total global entropy value as the migration target and performs online migration of the application. To avoid system turbulence caused by frequent migrations, a cooling-off period is set after each application migration before it can be migrated again; at the same time, when multiple source nodes simultaneously request migration to the same target node, the target node only accepts the first request, rejects other requests, and suggests alternative nodes, thus achieving random backoff.

[0029] like Figure 2 As shown, in a preferred embodiment of the present invention, the step of determining the interfering fingerprint specifically includes: S101 reads hardware performance counters on each physical core at 1-millisecond cycles, collecting the number of last-level cache misses, memory bandwidth read / write bytes, number of instruction cycles per cycle, and number of remote memory access requests. S102 maintains a circular buffer with a length of s sampling points for each performance counter, calculates the first-order difference and second-order difference of each sequence, and extracts the mean, variance and peak difference of the past t1 milliseconds as feature vectors. S103, input the feature vector into the online gradient boosting decision tree model, and output the predicted values ​​of memory bandwidth contention intensity and last-level cache contention intensity within the next t2 milliseconds, where s, t1 and t2 are constants; S104, synchronously calculate the average number of instruction cycles per instruction for the current application, and combine the predicted memory bandwidth contention intensity, the predicted last-level cache contention intensity, and the average number of instruction cycles per instruction into a three-dimensional interference fingerprint vector.

[0030] In this embodiment of the invention, a physical node contains multiple physical cores. The physical core is the basic unit for load execution and performance data collection. On each physical core, four types of key hardware events are read in 1-millisecond cycles: last-level cache misses (reflecting cache contention), memory bandwidth read / write bytes (reflecting bandwidth contention), number of instruction cycles per second (reflecting memory access latency sensitivity), and number of remote memory access requests (reflecting cross-node memory access behavior). This data is directly obtained from the CPU's performance monitoring unit. A circular buffer with 200 sampling points is maintained for each performance counter. First-order differences (the amount of change between adjacent sampling points) and second-order differences (the rate of change) are calculated for each sequence to capture sudden trends in load. Then, three statistics from the past 10 milliseconds—mean, variance, and peak difference—are extracted to form a feature vector describing the current load behavior. This feature vector is input into an online gradient boosting decision tree model. This model is a lightweight machine learning model that can be incrementally updated. The model outputs two predicted values ​​for the next 20 milliseconds: a predicted memory bandwidth contention intensity (unit: GB / s) and a predicted last-level cache contention intensity. Simultaneously calculate the average number of instruction cycles (CPI) for the current application within the most recent time window. Combine the predicted memory bandwidth contention intensity, the predicted last-level cache contention intensity, and the average CPI into a three-dimensional vector.

[0031] like Figure 3 As shown, in a preferred embodiment of the present invention, the steps of identifying an aggressive application and dynamically adjusting the last-level cache quota, memory bandwidth limit, and CPU scheduling priority of the aggressive application specifically include: S301. Based on the interference potential coefficient of each pair of applications, select the top three pairs of applications with the greatest contribution, and mark the pair with the higher predicted competitive intensity as aggressive applications and the pair with the lower predicted competitive intensity as sensitive applications. S302, for aggressive applications, the last-level cache quota is gradually reduced, each time by m% of the current quota, until the co-location interference entropy drops to a preset value of one, while ensuring that the cache of sensitive applications is not lower than the minimum guaranteed value calculated from its historical last-level cache miss rate, where m is a constant value. S303 uses memory bandwidth allocation technology to set a bandwidth limit for aggressive applications. The initial limit is set at 80% of the current bandwidth usage. If the interference entropy still increases in the next cycle, it will be reduced by 10%, with a minimum of no less than 40% of the original demand. S304 prioritizes the CPU scheduling of the sensitive application when it is running on the same physical core as both the aggressive and sensitive applications, giving it more execution time.

[0032] In this embodiment of the invention, based on the calculated interference potential coefficient of each pair of applications, they are sorted from largest to smallest, and the top three pairs of applications with the greatest contribution are selected. For each selected pair, the predicted contention intensity of the two applications (including the combined score of last-level cache and memory bandwidth) is compared. The application with higher predicted contention intensity is marked as an aggressive application, and the other is marked as a sensitive application. Aggressive applications are the main initiators of resource contention, and sensitive applications are the main victims. For aggressive applications, hardware-supported cache allocation technology (such as Intel CAT) is used to reduce their last-level cache quota by 5% each time. After each adjustment, the co-location interference entropy of the node is recalculated. This process is repeated until the co-location interference entropy drops to a preset value of one (e.g., 0.8). At the same time, it must be ensured that the cache capacity of sensitive applications is not lower than a lower limit, which is calculated based on the historical last-level cache miss rate of sensitive applications to guarantee the basic performance of sensitive applications. Next, memory bandwidth allocation technology (such as Intel MBA) is used to set an upper limit on the memory bandwidth of aggressive applications. The initial upper limit is set to 80% of the application's current bandwidth usage. If, after setting the limit, the interference entropy still increases in the next evaluation cycle, the upper limit is further reduced by 10%. This reduction can be done multiple times, but it must never fall below 40% of the application's original requirements to avoid rendering the application completely unusable. Finally, at the operating system CPU scheduler level, when aggressive and sensitive applications are assigned to the same physical core for round-robin execution, the scheduling priority of the sensitive application is proactively increased by two levels (e.g., from normal priority to a higher priority). This allows the sensitive application to obtain more CPU execution time, reducing queuing latency caused by waiting for shared resources (such as cache and memory bus).

[0033] like Figure 4 As shown, in a preferred embodiment of the present invention, the step of determining the candidate migration plan specifically includes: S401, Select the application that is most severely disturbed on the current node as the application to be migrated. This application is the one with the highest number of instruction cycles and the lowest predicted last-level cache contention intensity. S402, broadcast a migration query message to neighboring nodes through the node to be migrated. The message contains the interference fingerprint of the application to be migrated and the co-location interference entropy of the current node. S403: When a neighboring node receives a query, it simulates and calculates its own new co-location interference entropy after accepting the application. If the new entropy value is less than the preset value of 2 and the remaining resources are sufficient, it will provide an acceptance feedback. S404: Collect all feedback and select the target node that minimizes the sum of global co-location interference entropy as the final migration target.

[0034] In this embodiment of the invention, the application most severely affected by interference is identified on the current node. The criteria for judgment are: the highest Cycles Per Instruction (CPI) and the lowest predicted last-level cache contention intensity. A high CPI indicates that the application's execution efficiency is low, while a low cache contention intensity indicates that it is not actively contending for the cache, suggesting that the application's performance degradation is caused by interference from other aggressive applications. Such applications are most suitable for migration to nodes with less interference. Then, the node to be migrated (the source node) broadcasts a migration query message to its neighboring nodes through a distributed communication mechanism based on the Gossip protocol. The message content includes at least: the complete interference fingerprint of the application to be migrated, and the current co-location interference entropy value of the source node. In this way, neighboring nodes can assess the impact on themselves after accepting the application. Each neighboring node that receives the query uses the interference fingerprint of its currently co-located application to simulate and calculate the new local co-location interference entropy if the application to be migrated is added to its local machine. If the simulated new entropy value is less than a preset value of 2 (e.g., 1.5), and the remaining resources (CPU, memory, etc.) of the node are sufficient to run the application to be migrated, then the neighboring node returns an acceptable feedback to the source node, along with the predicted new entropy value. Otherwise, no feedback is given or the feedback is rejected. The source node collects all acceptable feedback and calculates the sum of the global co-located interference entropy for each candidate target node, which is the current entropy value of the source node minus the new entropy value after the application to be migrated, plus the new entropy value after the target node accepts it. The target node that minimizes this global sum is selected as the final migration target, aiming to minimize the overall interference level of the entire system.

[0035] like Figure 5 As shown in the figure, this embodiment of the invention also provides an intelligent cloud service management system, the system comprising: The interference fingerprint module 100 is used to collect hardware performance counter data on each physical node, predict the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and determine the interference fingerprint by combining the number of instruction cycles. The interference calculation module 200 is used to calculate the interference potential coefficient between each pair of applications based on the interference fingerprint of all applications on the same node, calculate the global co-location interference entropy of the node based on the interference potential coefficient, and classify the node into low interference, medium interference or high interference levels according to the interference entropy. The dynamic adjustment module 300 is used to identify aggressive applications when a node is in a low or medium level of interference, and dynamically adjust the last-level cache quota, memory bandwidth limit and CPU scheduling priority of the aggressive application to reduce interference entropy. The application migration module 400 is used to determine candidate migration plans when a node is at a high interference level or the interference entropy is still higher than the safety threshold after dynamic adjustment: the target node is selected through a distributed node negotiation mechanism, the application migration is executed, and a migration cooling-off period and random backoff mechanism are set to avoid oscillation.

[0036] In a preferred embodiment of the present invention, the interference fingerprint module 100 includes: The data acquisition unit is used to read the hardware performance counters on each physical core at 1-millisecond cycles, and collect the number of last-level cache misses, the number of bytes read and written in memory bandwidth, the number of instruction cycles per cycle, and the number of remote memory access requests. The feature extraction unit is used to maintain a circular buffer of length s for each performance counter, calculate the first and second differences of each sequence, and extract the mean, variance, and peak difference over the past t1 milliseconds as feature vectors. The information prediction unit is used to input the feature vector into the online gradient boosting decision tree model and output the predicted values ​​of memory bandwidth contention intensity and last-level cache contention intensity within the next t2 milliseconds, where s, t1, and t2 are constants. The fingerprint vector unit is used to synchronously calculate the average number of instruction cycles per instruction for the current application, and combines the predicted memory bandwidth contention intensity, the predicted last-level cache contention intensity, and the average number of instruction cycles per instruction into a three-dimensional interference fingerprint vector.

[0037] In a preferred embodiment of the present invention, the dynamic adjustment module 300 includes: The application labeling unit is used to sort each pair of applications according to their interference potential coefficient, select the top three pairs of applications with the greatest contribution, and label the pair with the higher predicted competitive intensity as aggressive applications and the pair with the lower predicted competitive intensity as sensitive applications. The quota adjustment unit is used to gradually reduce the last-level cache quota for aggressive applications, reducing the current quota by m% each time until the co-location interference entropy drops to a preset value of one, while ensuring that the cache of sensitive applications is not lower than the minimum guaranteed value calculated from their historical last-level cache miss rate. The bandwidth setting unit is used to set a bandwidth limit for aggressive applications using memory bandwidth allocation technology. The initial limit is set to 80% of the current bandwidth usage. If the interference entropy still increases in the next cycle, it will be reduced by 10%, with a minimum of no less than 40% of the original demand. The priority adjustment unit is used to increase the CPU scheduling priority of the sensitive application when the aggressive application and the sensitive application are rotating on the same physical core, so that the sensitive application can get more execution time.

[0038] In a preferred embodiment of the present invention, the application migration module 400 includes: The migration determination unit is used to select the application that is most severely disturbed on the current node as the application to be migrated. This application is the one with the highest number of instruction cycles and the lowest predicted last-level cache contention intensity. The message broadcasting unit is used to broadcast migration query messages to neighboring nodes through the node to be migrated. The message contains the interference fingerprint of the application to be migrated and the co-location interference entropy of the current node. The simulation calculation unit is used to simulate and calculate the new co-located interference entropy of a neighboring node after it receives a query. If the new entropy value is less than a preset value of 2 and there are sufficient remaining resources, then an acceptance feedback is given. The migration target unit is used to collect all feedback and select the target node that minimizes the sum of global co-location interference entropy as the final migration target.

[0039] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0040] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0041] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A cloud service management method with intelligent design, characterized in that, The method includes the following steps: Hardware performance counter data is collected on each physical node to predict the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and interference fingerprints are determined by combining the number of instruction cycles per instruction. Based on the interference fingerprints of all applications on the same node, the interference potential coefficient between each pair of applications is calculated. The global co-location interference entropy of the node is calculated according to the interference potential coefficient. The node is then classified into low-interference, medium-interference, or high-interference levels according to the interference entropy. When a node is at a low or medium level of interference, identify the aggressive application and dynamically adjust its last-level cache quota, memory bandwidth limit, and CPU scheduling priority to reduce interference entropy. When a node is at a high level of interference, or when the interference entropy is still higher than the safety threshold after dynamic adjustment, a candidate migration plan is determined: the target node is selected through a distributed node negotiation mechanism, the application migration is executed, and a migration cooling-off period and random backoff mechanism are set to avoid oscillation.

2. The cloud service management method for intelligent design according to claim 1, characterized in that, When calculating the global co-location interference entropy of nodes, the formula is: Hc = - Where K is the total number of applications co-located on the node. To apply the interference potential coefficient for (a,b), Hc is the sum of the interference potential coefficients of all applications on the node, where Hc is the co-location interference entropy; the interference potential coefficients are... It is calculated by combining the predicted memory bandwidth contention intensity, predicted last-level cache contention intensity, and the mean and standard deviation of the number of instruction cycles for applications a and b.

3. The cloud service management method for intelligent design according to claim 1, characterized in that, The step of determining the interfering fingerprint specifically includes: On each physical core, hardware performance counters are read at 1-millisecond intervals to collect the number of last-level cache misses, memory bandwidth read / write bytes, number of instruction cycles per cycle, and number of remote memory access requests. Maintain a circular buffer with a length of s sampling points for each performance counter, calculate the first and second differences of each sequence, and extract the mean, variance, and peak difference over the past t1 milliseconds as feature vectors; Input the feature vector into the online gradient boosting decision tree model, and output the predicted values ​​of memory bandwidth contention intensity and last-level cache contention intensity within the next t2 milliseconds, where s, t1, and t2 are constants; Simultaneously calculate the average number of instruction cycles per instruction for the current application, and combine the predicted memory bandwidth contention intensity, the predicted last-level cache contention intensity, and the average number of instruction cycles per instruction into a three-dimensional interference fingerprint vector.

4. The cloud service management method for intelligent design according to claim 1, characterized in that, The steps for identifying aggressive applications and dynamically adjusting their last-level cache quota, memory bandwidth limit, and CPU scheduling priority include: Based on the interference potential coefficient of each pair of applications, the top three pairs of applications with the greatest contribution are selected, and those with higher predicted competitive intensity are marked as aggressive applications, while those with lower predicted competitive intensity are marked as sensitive applications. For aggressive applications, the last-level cache quota is gradually reduced, with each reduction being m% of the current quota, until the co-location interference entropy drops to a preset value of one. At the same time, the cache of sensitive applications is guaranteed not to be lower than the minimum guaranteed value calculated from their historical last-level cache miss rate. Memory bandwidth allocation technology is used to set a bandwidth cap for aggressive applications. The initial cap is set at 80% of the current bandwidth usage. If the interference entropy still increases in the next cycle, it will be reduced by 10%, with a minimum of no less than 40% of the original demand. When aggressive and sensitive applications are running on the same physical core, the CPU scheduling priority of the sensitive application is increased, giving it more execution time.

5. The cloud service management method for intelligent design according to claim 1, characterized in that, The step of determining the candidate migration plan specifically includes: The application most severely affected by the current node is selected as the application to be migrated. This application is the one with the highest number of instruction cycles per instruction and the lowest predicted last-level cache contention intensity. The node to be migrated broadcasts a migration query message to its neighboring nodes. The message contains the interference fingerprint of the application to be migrated and the co-location interference entropy of the current node. When a neighboring node receives a query, it simulates and calculates its own new co-location interference entropy after accepting the application. If the new entropy value is less than the preset value of 2 and there are sufficient remaining resources, it will provide an acceptance feedback. Collect all feedback and select the target node that minimizes the sum of global co-location interference entropy as the final migration target.

6. The cloud service management method for intelligent design according to claim 1, characterized in that, The steps to calculate the interference potential coefficient between each pair of applications are as follows: obtain the predicted last-level cache contention intensity PL(a) and PL(b), the predicted memory bandwidth contention intensity PB(a) and PB(b), the mean per instruction cycle μCPI(a) and μCPI(b), and the standard deviation per instruction cycle σCPI(a) and σCPI(b) for application a and application b; calculate the common aggression factor = min(PL(a), PL(b)) × min(PB(a), PB(b)); Calculate the sensitivity factor = max(μCPI(a), μCPI(b)) + 1; calculate the volatility consistency factor = 1 / (1 + |σCPI(a) - σCPI(b)|); divide the common aggression factor by the sensitivity factor and multiply by the volatility consistency factor to obtain the interference potential coefficient. .

7. A cloud service management system with intelligent design, characterized in that, The system includes: The interference fingerprint module is used to collect hardware performance counter data on each physical node, predict the memory bandwidth contention intensity and last-level cache contention intensity of each application in the near future, and determine the interference fingerprint by combining the number of instruction cycles per instruction. The interference calculation module is used to calculate the interference potential coefficient between each pair of applications based on the interference fingerprint of all applications on the same node, calculate the global co-location interference entropy of the node based on the interference potential coefficient, and classify the node into low interference, medium interference or high interference levels according to the interference entropy. The dynamic adjustment module is used to identify aggressive applications when a node is in a low or medium level of interference, and dynamically adjust the last-level cache quota, memory bandwidth limit, and CPU scheduling priority of the aggressive application to reduce interference entropy. The application migration module is used to determine candidate migration plans when a node is at a high level of interference or when the interference entropy is still higher than the safety threshold after dynamic adjustment: the target node is selected through a distributed node negotiation mechanism, the application migration is executed, and a migration cooling-off period and random backoff mechanism are set to avoid oscillation.

8. The intelligent cloud service management system according to claim 7, characterized in that, The interference fingerprint module includes: The data acquisition unit is used to read the hardware performance counters on each physical core at 1-millisecond cycles, and collect the number of last-level cache misses, the number of bytes read and written in memory bandwidth, the number of instruction cycles per cycle, and the number of remote memory access requests. The feature extraction unit is used to maintain a circular buffer of length s for each performance counter, calculate the first and second differences of each sequence, and extract the mean, variance, and peak difference over the past t1 milliseconds as feature vectors. The information prediction unit is used to input the feature vector into the online gradient boosting decision tree model and output the predicted values ​​of memory bandwidth contention intensity and last-level cache contention intensity within the next t2 milliseconds, where s, t1, and t2 are constants. The fingerprint vector unit is used to synchronously calculate the average number of instruction cycles per instruction for the current application, and combines the predicted memory bandwidth contention intensity, the predicted last-level cache contention intensity, and the average number of instruction cycles per instruction into a three-dimensional interference fingerprint vector.

9. The intelligent cloud service management system according to claim 7, characterized in that, The dynamic adjustment module includes: The application labeling unit is used to sort each pair of applications according to their interference potential coefficient, select the top three pairs of applications with the greatest contribution, and label the pair with the higher predicted competitive intensity as aggressive applications and the pair with the lower predicted competitive intensity as sensitive applications. The quota adjustment unit is used to gradually reduce the last-level cache quota for aggressive applications, reducing the current quota by m% each time until the co-location interference entropy drops to a preset value of one, while ensuring that the cache of sensitive applications is not lower than the minimum guaranteed value calculated from its historical last-level cache miss rate. The bandwidth setting unit is used to set a bandwidth limit for aggressive applications using memory bandwidth allocation technology. The initial limit is set to 80% of the current bandwidth usage. If the interference entropy still increases in the next cycle, it will be reduced by 10%, with a minimum of no less than 40% of the original demand. The priority adjustment unit is used to increase the CPU scheduling priority of the sensitive application when the aggressive application and the sensitive application are rotating on the same physical core, so that the sensitive application can get more execution time.

10. The intelligent cloud service management system according to claim 7, characterized in that, The application migration module includes: The migration determination unit is used to select the application that is most severely disturbed on the current node as the application to be migrated. This application is the one with the highest number of instruction cycles and the lowest predicted last-level cache contention intensity. The message broadcasting unit is used to broadcast migration query messages to neighboring nodes through the node to be migrated. The message contains the interference fingerprint of the application to be migrated and the co-location interference entropy of the current node. The simulation calculation unit is used to simulate and calculate the new co-located interference entropy of a neighboring node after it receives a query. If the new entropy value is less than a preset value of 2 and there are sufficient remaining resources, then an acceptance feedback is given. The migration target unit is used to collect all feedback and select the target node that minimizes the sum of global co-location interference entropy as the final migration target.

Images