A load balancing scheduling method for data center virtualized servers
By collecting microarchitecture-level operational metrics to construct a contention graph, virtual machine migration can be predicted and optimized, solving the problem of invisible microarchitecture contention in existing technologies and achieving efficient load balancing scheduling and stable performance.
Patent Information
- Application Number
- CN202611298918.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
Existing load balancing scheduling methods cannot effectively identify and predict the microarchitectural resource contention relationships between virtual machines, resulting in poor targeting of virtual machine migration scheduling, easy to cause scheduling oscillations, and inability to guarantee the performance stability of latency-sensitive services.
By collecting microarchitecture-level operational metrics through hardware performance counters, constructing a microarchitecture contention graph, and combining time-series causal verification and latency mutual information analysis, the risk of contention before scheduling is predicted, and noise-source virtual machines are prioritized for migration after scheduling. Resource isolation strategies are configured to reduce performance degradation.
It improves the load balancing scheduling accuracy of virtualized server clusters, reduces the probability of business performance degradation, avoids repeated migrations and scheduling oscillations, and ensures the performance stability of latency-sensitive businesses.
Smart Images

Figure CN123114749A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of server load scheduling, and in particular to a load balancing scheduling method for data center virtualization servers. Background Technology
[0002] With the continuous expansion of cloud computing and data center scale, load balancing scheduling in virtualized server clusters has become a key technology for ensuring resource utilization and business performance. Existing load balancing scheduling methods mainly rely on macro-level resource indicators observable at the operating system level, such as CPU utilization, memory usage, network throughput, and disk I / O queue length. They employ methods such as round-robin, minimum connection count, dynamic weighting, or threshold triggering to migrate or deploy virtual machines to hosts with relatively low loads, aiming to achieve balanced resource utilization across hosts within the cluster. Some solutions further introduce hardware performance counters to collect underlying operational data for performance monitoring, but their scheduling decisions still primarily rely on macro-level resource indicators and have not yet incorporated micro-architecture level resource contention information into the load balancing scheduling system.
[0003] However, in a virtualized environment, multiple virtual machines share the same physical host's microarchitectural resources, such as last-level cache, memory bandwidth, cache coherency protocol processing units, and TLB. Relying solely on macro-level resource metrics for scheduling fails to detect the implicit contention between virtual machines arising from shared microarchitectural resources. In actual operation, the following situation often occurs: macro-level metrics such as host CPU utilization appear normal, but one virtual machine experiences both cache-sensitive loads and high last-level cache miss rates, causing a significant increase in tail latency for another virtual machine, resulting in a hidden overload problem of "normal macro-level metrics, but degraded service performance." Existing scheduling methods cannot quantify the risk of microarchitectural contention between the target virtual machine and existing virtual machines on the candidate host before scheduling, nor can they accurately identify the noise-generating virtual machine causing performance interference after contention occurs. Typically, they can only passively migrate affected virtual machines or repeatedly attempt migration, resulting in poor scheduling targeting, easy scheduling oscillations, and an inability to effectively guarantee the performance stability of latency-sensitive services.
[0004] To address this, this application proposes a load balancing scheduling method for data center virtualized servers. This method aims to solve the technical problems of invisible, unpredictable, and unlocatable implicit contention in existing scheduling methods from the perspective of microarchitecture resource contention awareness, thereby improving the load balancing scheduling accuracy and business performance assurance capabilities of virtualized server clusters. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a load balancing scheduling method for data center virtualization servers, employing the following technical solution: A load balancing scheduling method for data center virtualized servers includes the following steps: S1. On each host machine, the micro-architecture level operation indicators of each virtual machine are collected by hardware performance counters at a preset sampling period. The micro-architecture level operation indicators include the last-level cache miss rate, memory bandwidth consumption, cache line invalidation frequency, TLB miss rate, branch predictor contention index, and instruction retirement rate degradation index. S2. Obtain the steady-state baseline of each virtual machine. The steady-state baseline is obtained by selecting a sampling window where the overall load of the host machine is lower than a preset load threshold as the steady-state window, taking a preset quantile value for each microarchitecture metric of the virtual machine within each steady-state window as the steady-state baseline; normalizing each microarchitecture metric based on the steady-state baseline, constructing a microarchitecture contention fingerprint vector for each virtual machine, and marking each virtual machine as stress-inducing or sensitive based on historical running behavior. S3. For each pair of virtual machines coexisting on a host machine, construct a microarchitecture contention graph across virtual machines based on time-series causality testing and latency mutual information analysis. The microarchitecture contention graph uses virtual machines as nodes, directed edges to represent the direction of microarchitecture resource suppression by one virtual machine on another virtual machine, and edge weights to represent the suppression intensity. S4. When it is necessary to migrate or deploy the target virtual machine, obtain the existing virtual machine set of all candidate hosts, query the existing contention edge weights between the target virtual machine and the existing virtual machines in each candidate host based on the microarchitecture contention relationship graph, and combine the existing contention edge weights with the contention fingerprint similarity to comprehensively calculate the pre-scheduling contention risk value between the target virtual machine and each candidate host; if there is no existing contention edge, the weight of the edge is recorded as zero, and the contention fingerprint similarity is used as the main calculation basis. S5. Based on the pre-scheduling contention risk value, each candidate host is divided into three levels: high risk, medium risk, and low risk. High-risk candidate hosts are directly eliminated. If the candidate host set is empty after elimination, it is processed according to the preset degradation strategy. Medium-risk candidate hosts retain the qualification to participate in the scoring under the premise of synchronously configuring resource isolation strategy. Low-risk candidate hosts participate in the scoring normally. S6. Combine the contention risk value of the candidate host after hierarchical processing with the load balancing score to form a scheduling decision function, sort the candidate host scores, and select the candidate host with the highest comprehensive score as the migration or deployment target of the target virtual machine. S7. After scheduling execution, continuously monitor the microarchitecture metrics of each virtual machine. When the monitored microarchitecture metrics meet the preset contention anomaly conditions, locate the victim virtual machine with impaired performance, locate the noise source virtual machine based on the microarchitecture contention relationship graph, prioritize the migration of the noise source virtual machine rather than the victim virtual machine, and update the edge weights of the microarchitecture contention relationship graph in reverse according to the actual contention results.
[0006] Preferably, in step S2, the preset load threshold is that the overall load of the host machine is less than 30%, and the preset percentile value is the 95th percentile; and the steady-state window also meets the following conditions: combined with the virtual machine service layer monitoring data, the fluctuation range of the service request arrival rate of the virtual machine within the steady-state window is less than the preset fluctuation threshold, and no virtual machine migration, snapshot or backup operation has occurred.
[0007] Preferably, in step S3, the temporal causality test adopts the Granger causality test, and the time delay mutual information analysis is performed in the following manner: For a pair of virtual machines (VM_i, VM_j) coexisting on the host machine, calculate the latency mutual information between the change in the last-level cache miss rate of VM_i and the instruction retirement rate degradation index of VM_j. Scan a preset latency range, take the latency that maximizes the mutual information value as the optimal latency, and use the maximum mutual information value as the initial value of the edge weight. When the p-value of the Granger causality test is lower than the preset significance threshold and the mutual information value exceeds the preset mutual information threshold, establish a directed edge from VM_i to VM_j. The edge weight is obtained by weighted normalization of the p-value and the mutual information value.
[0008] Preferably, in step S4, the contention fingerprint similarity is calculated in the following manner: When there is no existing contention edge between the target virtual machine and an existing virtual machine in the candidate host in the microarchitecture contention graph, the cosine similarity between the microarchitecture contention fingerprint vector of the target virtual machine and the microarchitecture contention fingerprint feature of the existing virtual machine marked as sensitive in the candidate host when it was suppressed in the past is calculated; if the cosine similarity exceeds a preset similarity threshold, the similarity contention risk value is calculated based on the cosine similarity and the intensity of the historical suppression of the sensitive virtual machine, and the similarity contention risk value is included in the contention risk value before scheduling; the microarchitecture contention fingerprint feature when it was suppressed in the past is the statistical feature of the microarchitecture contention fingerprint vector of the sensitive virtual machine during the period when the historical contention event occurred.
[0009] Preferably, in step S5, the medium-risk candidate host retains its eligibility to participate in the scoring process, provided that a resource isolation policy is configured synchronously. Specifically: For medium-risk candidate hosts, a resource isolation policy based on cache allocation technology is generated. The resource isolation policy includes allocating a maximum of 8 final-level cache paths and a maximum of 40% memory bandwidth to the target virtual machine. The resource isolation policy is then issued and executed along with the scheduling decision.
[0010] Preferably, when all candidate hosts are classified as high-risk after classification, scheduling is not rejected. Instead, the high-risk candidate host with the lowest contention risk value is selected as the scheduling target, and the resource isolation policy is forcibly issued to the host. The maximum number of last-level cache paths in the resource isolation policy is 4, and the maximum memory bandwidth is 20%.
[0011] Preferably, in step S4, when calculating the pre-scheduling contention risk value, the historical contention fingerprint time series of the target virtual machine is used to predict the predicted contention fingerprint of the target virtual machine in a future preset time window, the contention fingerprint similarity is calculated based on the predicted contention fingerprint, and the pre-scheduling contention risk value is calculated in combination with the existing contention edge weights.
[0012] Preferably, in step S3, the microarchitecture contention graph adopts a two-level update mechanism: The first level is real-time incremental update within the same host: After each sampling period is completed, only the edge weights between coexisting virtual machines within the same host are updated exponentially. When the number of consecutive occurrences of a new contention event exceeds a preset threshold, the new edge is officially added. When the weight of an old edge decays to below a preset removal threshold, the edge is removed. The second level is full-scale computation across the entire cluster: according to a preset long period, time-series causality checks and latency mutual information analyses are re-executed for all active virtual machines across the entire cluster, and the contention graph is reconstructed.
[0013] Preferably, in step S7, the method for locating the noise source virtual machine is as follows: In the microarchitecture contention graph, find all incoming edges pointing to the victim virtual machine. Select the source virtual machines corresponding to directed edges whose edge weights exceed a preset noise threshold to form a noise source candidate set. Calculate the migration cost for each virtual machine in the noise source candidate set. Select the noise source virtual machine with the highest comprehensive score after weighted summation of migration cost and contention intensity and perform migration. When the noise source candidate set is empty or the migration cost of all noise source virtual machines is higher than the migration cost of the victim virtual machine, migrate the victim virtual machine instead.
[0014] Preferably, in step S7, the reverse update is performed in the following manner: The actual contention results after scheduling are compared with the predicted contention risks before scheduling: when the predicted contention risk is high and contention actually occurs, the weights of the corresponding directed edges are strengthened; when the predicted contention risk is low but contention actually occurs, missed events are recorded and the corresponding contention edges are supplemented; when the predicted contention risk is high but no contention actually occurs, the weights of the corresponding directed edges are accelerated to decay; after each update, the updated edge weights are normalized to ensure that the edge weights remain within the preset range.
[0015] In summary, this application includes at least one of the following beneficial technical effects: By collecting microarchitecture-level performance metrics such as last-level cache miss rate, memory bandwidth consumption, cache line miss frequency, TLB miss rate, branch predictor contention index, and instruction retirement rate degradation index using hardware performance counters, and constructing a cross-virtual machine microarchitecture contention graph based on time-series causality testing and latency mutual information analysis, the implicit contention relationships between virtual machines at the shared microarchitecture resource level are made visible and quantifiable, solving the problem that existing technologies cannot detect the implicit overload problem of "normal macro load but conflicting micro performance" by relying only on macro indicators at the operating system level.
[0016] By querying the existing contention edge weights between the target virtual machine and existing virtual machines in the candidate host before scheduling execution, and combining the contention fingerprint similarity to calculate the contention risk value before scheduling, the microarchitectural contention risk that may occur on the target host can be predicted before virtual machine migration or deployment, rather than passively responding after overload or performance degradation has already occurred, thereby reducing the probability of business performance degradation after scheduling from the root.
[0017] By continuously monitoring and locating noise source virtual machines based on the microarchitecture contention graph after scheduling, and prioritizing the migration of noise source virtual machines rather than victim virtual machines, this approach changes the previous practice of passively migrating performance-impaired virtual machines. It avoids repeated migrations and scheduling oscillations caused by the failure to eliminate noise sources, while also reducing unnecessary migration downtime overhead for victim virtual machines.
[0018] By classifying the contention risk value into three levels—high risk, medium risk, and low risk—before scheduling, and synchronously configuring resource isolation strategies based on cache allocation technology for medium-risk candidate hosts, the performance of latency-sensitive services can still be guaranteed by limiting the number of last-level cache paths and memory bandwidth of noise sources in scenarios where contention risk cannot be completely avoided. This fills the decision gap of "either migrate or let it run" in existing scheduling methods.
[0019] By comparing the actual contention results after scheduling with the predicted contention risks before scheduling, reverse update operations such as strengthening, adding, accelerating decay, or normal decay are performed on the edge weights of the contention relationship graph. This allows the contention relationship graph to be continuously iterated and optimized as the cluster runs, gradually enhancing its ability to identify new workloads and periodic contention patterns, and reducing the false negative rate over time.
[0020] By employing a two-level update mechanism that combines real-time incremental updates on the same host machine with full-cycle computation across the entire cluster, the computational overhead required for daily maintenance is reduced from O(n²) of full-cycle computation to a linear level related to the number of active hosts, while ensuring the timeliness of the contention graph. This meets the engineering deployment requirements of large-scale virtualization clusters. Attached Figure Description
[0021] Figure 1This is a flowchart of a data center virtualization server load balancing scheduling method according to an embodiment of this application. Detailed Implementation
[0022] The following is in conjunction with the appendix Figure 1 This application will be described in further detail.
[0023] This application discloses a load balancing scheduling method for data center virtualization servers. (Refer to...) Figure 1 A load balancing scheduling method for data center virtualized servers is disclosed. This method runs in the scheduler of the virtualization cluster and works in conjunction with the performance metric collection agents and virtualization management platform on each host machine. The method includes the following steps: S1. Microarchitecture-level operation metrics collection: On each host machine, microarchitecture-level operation metrics of each virtual machine are collected through hardware performance counters at a preset sampling period. The microarchitecture-level operation metrics include last-level cache miss rate, memory bandwidth consumption, cache line miss frequency, TLB miss rate, branch predictor contention index, and instruction retirement rate degradation index. S2. Steady-state baseline acquisition and contention fingerprint construction: The steady-state baseline of each virtual machine is acquired by selecting a sampling window where the overall load of the host machine is lower than a preset load threshold as the steady-state window, and taking the preset quantile value of each microarchitectural indicator of the virtual machine within each steady-state window as the steady-state baseline; based on the steady-state baseline, each microarchitectural indicator is normalized to construct the microarchitectural contention fingerprint vector of each virtual machine, and each virtual machine is marked as stress-inducing or sensitive according to its historical running behavior; S3. Construction of Microarchitecture Contention Graph: For each pair of virtual machines coexisting on a host machine, a cross-virtual machine microarchitecture contention graph is constructed based on time-series causality test and latency mutual information analysis; the microarchitecture contention graph uses virtual machines as nodes, directed edges represent the direction of microarchitecture resource suppression by one virtual machine on another virtual machine, and edge weights represent the suppression intensity. S4. Calculation of contention risk value before scheduling: When it is necessary to migrate or deploy the target virtual machine, obtain the existing virtual machine set of all candidate hosts, query the existing contention edge weights between the target virtual machine and the existing virtual machines in each candidate host based on the microarchitecture contention relationship graph, and combine the existing contention edge weights with the contention fingerprint similarity to comprehensively calculate the pre-scheduling contention risk value between the target virtual machine and each candidate host; if there is no existing contention edge, the weight of the edge is recorded as zero, and the contention fingerprint similarity is used as the main calculation basis; S5. Candidate Host Classification and Resource Isolation Strategy Configuration: Based on the pre-scheduling contention risk value, each candidate host is divided into three levels: high risk, medium risk, and low risk. High-risk candidate hosts are directly eliminated. If the candidate host set is empty after elimination, it is processed according to the preset downgrade strategy. Medium-risk candidate hosts retain their eligibility to participate in the scoring under the premise of synchronously configuring resource isolation strategies. Low-risk candidate hosts participate in the scoring normally. S6. Joint scheduling decision: The contention risk value of the candidate host after hierarchical processing is combined with the load balancing score to form a scheduling decision function. The candidate hosts are ranked by score, and the candidate host with the highest comprehensive score is selected as the migration or deployment target of the target virtual machine. S7. Post-scheduling contention monitoring and reverse update: After scheduling is executed, the microarchitecture metrics of each virtual machine are continuously monitored. When the monitored microarchitecture metrics meet the preset contention anomaly conditions, the victim virtual machine with impaired performance is located. Based on the microarchitecture contention relationship graph, the noise source virtual machine is located. The noise source virtual machine is migrated first instead of the victim virtual machine. The edge weights of the microarchitecture contention relationship graph are updated in reverse according to the actual contention results.
[0024] The specific implementation methods for each step are explained below.
[0025] Step S1, the collection of microarchitecture-level operational metrics, specifically involves: A lightweight performance acquisition agent is deployed on each host machine. This agent reads hardware performance counter data through the vPMU interface provided by the hypervisor. In a KVM / QEMU environment, the PMU event count value corresponding to the vCPU can be obtained through the perf_event_open system call; in a VMware ESXi environment, an equivalent metric can be obtained through the vSphere Performance API; and in a Xen environment, it can be obtained through xenoprof. This embodiment uses a KVM / QEMU environment as an example, but those skilled in the art should understand that the same function can be achieved through equivalent hardware performance count events under different processor architectures such as ARM and AMD. The preset sampling period is 1 second. The agent reads the hardware performance counter value of each virtual machine once in each sampling period, calculates the increment, and divides it by the sampling period to obtain the rate-type metric.
[0026] This embodiment collects the following six microarchitecture-level operational metrics: (1) The last-level cache miss rate is obtained by reading the LLC occupancy count corresponding to each RMID through Intel CMT, or by the ratio of PMU events MEM_LOAD_RETIRED.LLC_MISS to MEM_LOAD_RETIRED.ALL_LOADS. The calculation formula is as follows: Lmiss=MEM_LOAD_RETIRED.LLC_MISS / MEM_LOAD_RETIRED.ALL_LOADS This metric reflects the degree of competition among virtual machines for the shared last-level cache capacity. The higher the last-level cache miss rate, the greater the probability that the virtual machine will be evicted of cached data by other virtual machines.
[0027] (2) Memory bandwidth consumption, which is obtained by reading the local memory bandwidth value corresponding to each RMID through Intel MBM, in MB / s. On the AMD platform, it can be obtained through the equivalent QoS Monitoring function. This indicator reflects the intensity of the virtual machine's use of memory controller bandwidth.
[0028] (3) The cache line invalidation frequency is calculated by summing the PMU events MEM_LOAD_L3_HIT_RETIRED.XSNP_HITM (loading dirty cache lines that hit other cores) and MEM_LOAD_L3_HIT_RETIRED.XSNP_HIT (loading clean cache lines that hit other cores), i.e.: Clinv=XSNP_HITM+XSNP_HIT This metric reflects the overhead of cache consistency protocols. In a virtualization environment with a multi-core shared LLC, frequent reads and writes of shared cache lines by a virtual machine can cause frequent invalidation of cache lines on other cores, generating a large amount of consistency protocol traffic.
[0029] (4) TLB miss rate is calculated by the ratio of PMU events DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK to DTLB_LOAD_MISSES.WALK_COMPLETED. Instruction TLB events can be combined if necessary. This indicator reflects the contention pressure of the address translation structure. When the TLB miss rate increases, the virtual machine needs to perform page table traversal more frequently, consuming additional memory bandwidth and clock cycles.
[0030] (5) Branch predictor contention index, calculated by the ratio of PMU events BR_MISP_RETIRED.ALL_BRANCHES to BR_INST_RETIRED.ALL_BRANCHES, and then compared with the branch misprediction rate under the steady-state baseline of the virtual machine: Bpcon=Mispredict_rate_current / Mispredict_rate_baseline The branch predictor is a critical shared resource in modern CPU front-end pipelines. When multiple virtual machines run simultaneously on the same physical core via time-division multiplexing or SMT, the branch history table and branch target buffer are frequently overwritten. An increase in Bpcon indicates that the virtual machine is more severely affected by front-end contention.
[0031] (6) Instruction retirement rate degradation index, calculated as the ratio of the sampling rate of the PMU event INST_RETIRED.ANY to the instruction retirement rate at the steady-state baseline of the virtual machine: Retdeg=Retire_rate_baseline / Retire_rate_current Retire_rate represents the number of instructions successfully retired per unit of time. This index measures the degree to which the virtual machine is affected by microarchitectural contention from the perspective of overall execution efficiency. A Retdeg greater than 1 indicates that the current execution efficiency is lower than the steady-state baseline, and the larger the value, the more severe the performance degradation.
[0032] Of the six metrics mentioned above, the last-level cache miss rate, memory bandwidth consumption, cache line miss frequency, TLB miss rate, and branch predictor contention index are used to characterize the virtual machine's occupation and competition behavior for shared resources in the microarchitecture, while the instruction retirement rate degradation index is used to measure the overall performance degradation of the virtual machine after being affected by contention. Each metric is continuously collected according to the sampling period to form time-series data, providing a data foundation for subsequent contention graph construction and scheduling decisions.
[0033] Step S2, steady-state baseline acquisition and contention fingerprint construction, specifically involves: Since virtual machines typically run on shared host machines and lack completely independent operating environments, this embodiment obtains the steady-state baseline of each virtual machine through load threshold filtering and quantile statistics. First, the overall host machine load, i.e., the total CPU utilization of all vCPUs on the host machine, is collected through the host machine monitoring interface of a virtualization management platform (such as OpenStackNova or VMware vCenter). Then, the continuous time is divided into 10-minute windows, and the average overall host machine load is calculated for each window. Window windows where the overall host machine load is lower than a preset load threshold (30% in this embodiment) are selected as candidate steady-state windows. Further, within the candidate steady-state windows, based on virtual machine service layer monitoring data, time periods where the virtual machine's service request arrival rate fluctuation exceeds a preset fluctuation threshold (±15% in this embodiment) are excluded. The service request arrival rate is obtained through virtual machine network connection counts, application layer access logs, or monitoring data reported by tenants. Time periods during which the virtual machine undergoes migration, snapshot, or backup operations are also excluded. The final retained window is the steady-state window.
[0034] For each microarchitectural metric (Lmiss, MBW, Clinv, TLBmiss, Bpcon, Retdeg) of the virtual machine within each steady-state window, sort the data in ascending order and take the value corresponding to the 95th percentile as the steady-state baseline value of that metric. Let the steady-state baseline vector be: ; The 95th percentile value was used instead of the mean to eliminate the interference of a few abnormal spikes on the baseline, so that the baseline can better represent the stable operation level of the virtual machine in a low contention environment.
[0035] The microarchitecture metrics are normalized based on the steady-state baseline. This embodiment uses ratio normalization, comparing the raw metric values collected in each sampling period with the steady-state baseline of the virtual machine to obtain the normalized microarchitecture metrics: ; in, Let be the original value of the k-th index at time t. Let k be the steady-state baseline value of the k-th index. The smoothing factor has a value of 10. -6 To avoid division by zero, the normalized metrics form the microarchitectural contention fingerprint vector of the virtual machine at time t.
[0036] This vector reflects the change in microarchitectural resource usage pressure of the virtual machine at the current moment relative to its steady-state baseline.
[0037] Based on historical operational behavior, each virtual machine is marked as either stressful or sensitive. Stressful type determination: If, during historical operation, the normalized last-level cache miss rate of virtual machine VM_i increases, and the instruction retirement rate degradation index of at least one other virtual machine VM_j on the same host increases synchronously, and the Pearson correlation coefficient between the two is greater than a preset correlation coefficient threshold (0.6 in this embodiment), and passes the significance test (p<0.05), then VM_i is marked as stressful. Sensitive type determination: If, during historical operation, the instruction retirement rate degradation index of virtual machine VM_j is sensitive to changes in the last-level cache miss rate of other virtual machines on the same host (determination condition is the same as above), then VM_j is marked as sensitive. It should be noted that a virtual machine can be marked as both stressful and sensitive simultaneously. This marking result will be used for calculating the contention fingerprint similarity in step S4.
[0038] Step S3, the construction of the microarchitecture contention graph, specifically involves: The microarchitectural contention graph G=(V,E,W) uses virtual machines as nodes. Directed edges represent the direction of microarchitectural resource suppression by one virtual machine on another, and edge weights represent the suppression intensity. For example, if there is a directed edge from VM_i to VM_j, it means that VM_i is suppressing VM_j in terms of microarchitectural resources; the larger the edge weight, the stronger the suppression.
[0039] For each pair of virtual machines (VM_i, VM_j) coexisting on the host machine, perform the following analysis to construct directed edges: First, a Granger causality test is used to determine whether changes in the microarchitectural metrics of VM_i are the cause of performance degradation in VM_j. Specifically, a vector autoregression (VAR) model is established with the change in the last-level cache miss rate ΔLmiss_i(t) of VM_i as the independent variable and the change in the instruction retirement rate degradation index ΔRetdeg_j(t) of VM_j as the dependent variable. The lag order is determined according to the AIC criterion. The null hypothesis H0 is tested: "ΔLmiss_i of VM_i is not a Granger cause of ΔRetdeg_j". If the p-value of the test is lower than the preset significance threshold α (0.05 in this embodiment), the null hypothesis is rejected, and it is considered that there is a temporal causal relationship between VM_i and VM_j.
[0040] Next, perform time delay mutual information analysis. Calculate the time delay mutual information between ΔLmiss_i of VM_i and ΔRetdeg_j of VM_j. Scan a preset delay range τ∈[0,30] sampling periods, and calculate for each τ:
[0041] The probability density is estimated using the histogram method, and the continuous variable is discretized according to a preset number of bins. In this embodiment, the default number of bins is 10. The maximum delay is taken as the optimal delay τ*, and this maximum mutual information value is used. As an indicator of correlation strength.
[0042] When the p-value of the Granger causality test is lower than the preset significance threshold (0.05) and When the mutual information threshold is exceeded (0.2 in this embodiment), a directed edge is established from VM_i to VM_j. The edge weight is obtained by weighted normalization of the p-value and the mutual information value. Since a smaller p-value indicates stronger causal significance and a larger mutual information value indicates stronger correlation, and their directions are inconsistent, they need to be unified into a positive index first. In this embodiment, the p-value is converted into a negative logarithmic form, then normalized with the mutual information value, and finally weighted summed and mapped to the (0,1) interval using the sigmoid function: ; Where σ(x) = 1 / (1 + e^{-x}), α and β are preset weighting coefficients, both of which are taken as 0.5 in this embodiment; and Calculated for the current batch The maximum value of the mutual information is used for normalization.
[0043] The contention graph is maintained by the scheduler and stored in an in-memory database (such as Redis), organized as an adjacency list. Each node maintains a list of outgoing edges and a list of incoming edges; each edge records the source VMID, destination VMID, edge weight, most recent update timestamp, and number of triggers.
[0044] This embodiment employs a two-level update mechanism. The first level is real-time incremental updates within the same host machine: After each sampling period, the acquisition agent pushes the microarchitectural metric data from the host machine to the scheduler. The scheduler only performs exponential decay updates on the existing edge weights between coexisting virtual machine pairs within the host machine. Specifically, after each sampling period, the edge weight is multiplied by a decay coefficient of 0.99; if a new contention event that meets the conditions is detected, the corresponding edge weight is strengthened (multiplied by 1.05 and truncated to the upper limit of 0.95). If the number of consecutive occurrences of contention events between a pair of VMs exceeds a preset threshold (10 times in this embodiment), a new edge is formally added to the graph; if the weight of an old edge decays to below a preset removal threshold (0.1 in this embodiment), the edge is removed.
[0045] The second level is a full-cluster periodic computation: a full computation is performed every 6 hours. The scheduler background process clears the current contention graph and, based on the microarchitectural metrics data from the last 24 hours, re-executes Granger causality tests and latency mutual information analysis on all active virtual machine pairs across the entire cluster to reconstruct the contention graph. The time complexity of the full computation is O(n²), where n is the total number of VMs in the cluster; however, through the two-level update mechanism, daily real-time maintenance only requires incremental processing of VM pairs on active host machines, reducing the daily computational complexity to a linear level.
[0046] When a virtual machine is migrated from host A to host B, the contention edges between it and the VMs on host A are not immediately deleted. Instead, they undergo accelerated exponential decay at a rate of 0.9 per hour. If the VM migrates back to host A within 72 hours, the historical contention relationship can be quickly restored; after 72 hours, the edge weight decays to below the removal threshold and is automatically deleted. This design preserves historical experience, facilitating the rapid identification of recurring contention patterns on the new host.
[0047] Step S4, the calculation of contention risk value before scheduling, is as follows: When a new virtual machine is deployed in the virtualization cluster, an existing virtual machine is migrated due to excessive load, a load balancing strategy is periodically migrated, or a virtual machine is required to be migrated due to hardware failure prediction / manual intervention, the pre-scheduling contention risk assessment in step S4 is triggered. The scheduler initially filters out host machines that meet the resource requirements of the target virtual machine based on the available resources (CPU, memory, storage, etc.) of all host machines in the cluster, forming a candidate host machine set {H_1,H_2,...,H_m}.
[0048] For the target virtual machine VM_k and each candidate host H_m, assume that H_m already hosts a set of virtual machines S_m={VM_1,VM_2,...,VM_n}. Calculate the pre-scheduling contention risk value R_total(k,H_m) as follows: (1) Existing Contest Edge Weight Query. In the microarchitecture contest graph, query the weights of all existing directed edges between the target virtual machine VM_k and each existing virtual machine in the candidate host H_m. If existing edges exist, take the maximum edge weight and the average edge weight to calculate the direct contest risk: R_direct(k,H_m)=λ_max·max_{j∈S_m}w_{k→j}+λ_avg·(1 / |S_m|)Σ_{j∈S_m}w_{k→j} Where λ_max and λ_avg are preset weight coefficients, which are 0.7 and 0.3 respectively in this embodiment. If there is no existing contention edge (i.e., the target VM has never run on the same host machine as any VM in S_m), the edge weight is recorded as zero, i.e., R_direct=0, and contention fingerprint similarity is used as the main calculation basis.
[0049] (2) Risk of contention fingerprint similarity. When there is no existing contention edge between the target virtual machine VM_k and an existing virtual machine VM_j in the candidate host machine H_m, the cosine similarity between the microarchitecture contention fingerprint vector F_k of the target virtual machine and the microarchitecture contention fingerprint feature of VM_j when it was historically suppressed is calculated. The microarchitecture contention fingerprint feature when it was historically suppressed is defined as: the statistical characteristics of the microarchitecture contention fingerprint vector of VM_j during the period when the historical contention event occurred (the judgment condition is that the instruction retirement rate degradation index of VM_j exceeds 2.0, and the LLC miss rate of at least one other VM on the same host machine increases synchronously). In this embodiment, the median vector of each indicator during this period is taken.
[0050] The formula for calculating cosine similarity is: cos_sim(k,j)=(F_k·F_victim(j)) / (||F_k||·||F_victim(j)||) Where F_victim(j) is the historical suppressed feature vector of VM_j. If the cosine similarity exceeds a preset similarity threshold (0.8 in this embodiment), then the similarity contention risk value is calculated: R_similarity(k,j)=cos_sim(k,j)×(S_victim(j) / S_max) Where S_victim(j) is the historical suppression intensity of VM_j, defined as the ratio of the peak value of the instruction retirement rate degradation index of VM_j in historical contention events to the steady-state baseline value; S_max is the maximum value of the historical suppression intensity of the entire cluster, used for normalization. The above similarity contention risk value is calculated for all sensitive VMs in candidate host H_m that meet the conditions, and the maximum value is taken as the similarity risk R_similarity(k,H_m) of that candidate host.
[0051] (3) Existing internal contention risk. There may already be contention among existing virtual machines on the candidate host H_m. If there are many contention edges or their weights are high, it indicates that the host is already at microarchitectural saturation and it is not advisable to add more tasks. Calculate the existing internal contention risk: R_internal(H_m)=(1 / |S_m|)Σ_{i,j∈S_m,i≠j}w_{i→j} (4) Comprehensive contention risk. The contention risk value before scheduling is obtained by weighted summation of the above three factors: R_total(k,H_m)=γ1·R_direct(k,H_m)+γ2·R_similarity(k,H_m)+γ3·R_internal(H_m); Wherein γ1, γ2, and γ3 are preset weight coefficients, which are set to 0.5, 0.3, and 0.2 respectively in this embodiment. The confidence level of direct historical contention relationship is the highest, so γ1 is the largest; similarity risk belongs to blind zone prediction, so its confidence level is the second highest; internal risk is an additive factor, so its weight is the smallest.
[0052] When calculating the contention risk value before scheduling, the historical contention fingerprint time series of the target virtual machine can be used to predict the predicted contention fingerprint of the target virtual machine within a future preset time window. This embodiment uses exponential smoothing with a smoothing coefficient α of 0.3. An exponential smoothing model is established for each dimension of the historical contention fingerprint of the target VM_k to predict the contention fingerprint vector F_k_pred after T minutes (10 minutes in this embodiment). If the historical data of the VM is insufficient (e.g., less than 100 sampling periods), no prediction is performed, and the current fingerprint is used directly. When using the predicted fingerprint, F_k_pred is used to replace F_k to perform the above cosine similarity calculation and existing contention edge weight query, in order to predict periodic or sudden contention risks in advance. For example, if the last-level cache miss rate of a virtual machine has a periodic peak between 2:00 AM and 4:00 AM every day, when scheduling is performed at 1:50 AM, even if the current fingerprint is normal, the predicted peak fingerprint is used for matching, thereby avoiding deploying the virtual machine that is about to experience contention to an unsuitable host.
[0053] Step S5, the configuration of candidate host machine classification and resource isolation strategy, is as follows: Candidate hosts are categorized into three levels based on the pre-scheduling contention risk value R_total. A value of R_total ≥ 0.7 indicates high risk; 0.3 ≤ R_total < 0.7 indicates medium risk; and R_total < 0.3 indicates low risk. These thresholds are example values and can be dynamically adjusted based on cluster size and historical misjudgment rates in actual deployments.
[0054] High-risk candidate hosts are directly removed from the candidate set. If the candidate host set is empty after removal, a preset degradation strategy is executed: scheduling is not rejected, but the high-risk candidate host with the lowest contention risk value is selected as the scheduling target, and a resource isolation policy is forcibly issued to that host. The upper limit of the last-level cache path and the upper limit of memory bandwidth in the resource isolation policy are configured with preset values that are more stringent than those for the medium-risk level. In this embodiment, the medium-risk scenario allocates 8 / 20 paths of LLC and a memory bandwidth limit of 40%; the high-risk degradation scenario allocates 4 / 20 paths of LLC and a memory bandwidth limit of 20%.
[0055] For medium-risk candidate hosts, the scheduler generates a resource isolation policy based on cache allocation technology. Taking IntelCAT (Cache Allocation Technology) / RDT (Resource Director Technology) as an example: LLC is divided into N paths (20 paths in this embodiment). By setting the virtual machine's COS association value, the upper limit of LLC path allocation is set for the target virtual machine; in medium-risk scenarios, 8 LLC paths are allocated to the target virtual machine. A memory bandwidth upper limit is set through Intel MBM limits or CDP; in medium-risk scenarios, the memory bandwidth upper limit for the target virtual machine is set to 40% of the total bandwidth. After this resource isolation policy is generated by the scheduler, it is sent to the target host through libvirt's cputune configuration or by directly manipulating the MSR register. On the AMD platform, the same functionality can be achieved using AMD QoS extensions or equivalent interfaces of the platform security processor. The resource isolation policy is issued and executed along with the scheduling decision. Low-risk candidate hosts participate in the scoring normally.
[0056] Load balancing scoring can be achieved using existing multidimensional load assessment methods. This embodiment employs the following calculation method: LoadScore(H_m)=w1·U_cpu+w2·U_mem+w3·U_io Where U_cpu, U_mem, and U_io are the normalized values of CPU utilization, memory utilization, and IO queue length of the candidate host machine, respectively, and w1, w2, and w3 are weighting coefficients, which are 0.4, 0.3, and 0.3 in this embodiment. The lower the LoadScore, the lighter the host machine load and the greater the load balancing benefit.
[0057] Step S6, joint scheduling decision, specifically includes: The contention risk value of the candidate host after tiered processing is combined with the load balancing score to form the scheduling decision function: Score(k,H_m)=α·(1-LoadScore(H_m))-β·R_total(k,H_m) LoadScore is normalized to the [0,1] interval, and (1-LoadScore) represents the load balancing benefit; R_total is the contention risk value before scheduling; α and β are preset weight coefficients, both set to 0.5 in this embodiment. The Score is calculated for all candidate hosts after tiered processing, and they are sorted from highest to lowest Score. The candidate host with the highest Score is selected as the migration or deployment target for the target virtual machine. The scheduler sends this decision to the virtualization management platform to execute the virtual machine migration or deployment operation.
[0058] Step S7, Post-Schedule Contention Monitoring and Reverse Update, specifically includes: After scheduling is executed, the microarchitecture metrics of each virtual machine continue to be monitored at a sampling period of 1 second. A contention anomaly is determined when any of the following preset contention anomaly conditions are met: Condition 1: The instruction retirement rate degradation index of any virtual machine exceeds a preset anomaly threshold (2.0 in this embodiment) for 5 consecutive sampling periods, i.e., Retdeg > 2.0 for 5 consecutive seconds; Condition 2: The service SLO metric (e.g., p99 latency) of any virtual machine exceeds a preset SLO threshold (e.g., p99 > 100ms) for a duration exceeding a preset duration (e.g., 3 minutes). Meeting either condition triggers the contention anomaly handling process.
[0059] Victim virtual machines are those whose instruction retirement rate degradation index exceeds the threshold or whose SLO deteriorates according to the above anomaly detection conditions. For example, if VM_j's Retdeg exceeds 2.0 for 5 consecutive seconds, then VM_j is identified as a victim virtual machine.
[0060] In the microarchitecture contention graph, all incoming edges pointing to the victim virtual machine VM_j are searched. The source virtual machines corresponding to directed edges whose edge weights exceed a preset noise threshold (0.5 in this embodiment) are selected to form a noise source candidate set. If the noise source candidate set is empty, it means that the contention graph has not yet learned the contention relationship. An online contention detection is then performed: the LLC miss rate of each virtual machine on the same host is compared, and the virtual machine with the largest change rate that matches the performance degradation time of the victim VM is selected as a temporary noise source.
[0061] Calculate the migration cost for each virtual machine in the candidate noise source set: C_i=C_downtime+C_bandwidth+C_risk C_downtime is an estimated downtime based on the virtual machine's dirty page rate and pre-migration time, and is proportional to the memory size and dirty page rate; C_bandwidth is an estimated transmission time based on the total memory divided by the available migration bandwidth; C_risk is based on the number of previous migration failures for this VM or the probability of SLO degradation after migration.
[0062] For each VM_i in the candidate noise source set, calculate the comprehensive score: SelectScore_i=w_{i→victim}×0.6+(1-C_i / C_max)×0.4 Where w_{i→victim} is the weight of the directed edge from VM_i to the victim VM, and C_max is the maximum migration cost in the noise source candidate set. The noise source virtual machine with the highest SelectScore is selected for migration. When the noise source candidate set is empty, or the migration cost of all noise source virtual machines is more than 1.2 times the migration cost of the victim virtual machine, the victim virtual machine is migrated instead. If neither the noise source nor the victim virtual machine can be migrated (e.g., migration locking exists), a resource isolation policy is forcibly issued to limit the LLC path number and memory bandwidth of the noise source to alleviate contention.
[0063] The actual contention results after scheduling are compared with the predicted contention risk before scheduling, and the following reverse update rules are executed: When the predicted contention risk is high (R_total≥0.7) and contention actually occurs, the weight of the corresponding directed edge is strengthened, w_new=min(0.95,w_old×1.2); when the predicted contention risk is low (R_total<0.3) but contention actually occurs, the missed event is recorded and the corresponding contention edge is supplemented, with the initial weight of the new edge set to 0.5; when the predicted contention risk is high but no contention actually occurs (no anomalies for 5 consecutive sampling periods), the weight of the corresponding directed edge is accelerated, w_new=w_old×0.8; in other cases, the weight of the relevant edge is normally decayed, w_new=w_old×0.95. After each update, the updated edge weights are normalized to ensure that the edge weights remain within the preset range [0,1]: if they exceed 1, they are truncated to 1; if they are below 0, they are set to 0. If the total weight of the outgoing edges of a VM exceeds the threshold, all outgoing edges are scaled proportionally.
[0064] Example: A scenario where high-defense cleaning VMs and latency-sensitive service VMs are deployed together.
[0065] A data center deployed a DDoS protection scrubbing VM and an online payment service VM on the same host machine. During the scrubbing cycle, the DDoS protection scrubbing VM consumes up to 80% of the total bandwidth and has a last-level cache miss rate exceeding 90%, making it a typical stress-intensive VM. The online payment service VM is latency-sensitive, requiring a Service Level Requirement (SLO) of p99 < 50ms, making it a typical sensitive VM.
[0066] Traditional load balancing only observes a host CPU utilization of 45%, deeming the load normal, and does not trigger any scheduling action. However, in actual operation, the high-defense cleaning VM consumes a large amount of memory bandwidth and LLC during each cleaning cycle, causing the instruction retirement rate degradation index of the online payment VM to rise to 2.3, and the p99 latency to soar to over 300ms, severely violating the SLO.
[0067] After applying the method of this invention: In S1, the agent collects the microarchitecture-level operating metrics of each virtual machine through the PMU and Intel CMT / MBM interfaces; in S2, the steady-state baselines of the high-defense cleaning VM and the online payment VM are determined, their respective microarchitecture contention fingerprint vectors are constructed, and the high-defense cleaning VM is marked as a stressor and the online payment VM as a sensitive VM; in S3, a microarchitecture contention graph is constructed, and a directed edge is established from the high-defense cleaning VM to the online payment VM, with an edge weight calculated to be 0.87; in S4, when load balancing scheduling is triggered, the pre-scheduling contention risk value between each candidate host and the target virtual machine is calculated; in S5, the candidate hosts are divided into risk levels, and resource isolation policies are synchronously configured for medium-risk hosts; in S6, the optimal target host is selected by combining the load balancing score and the contention risk value.
[0068] Specifically, the scheduler migrates the high-defense cleaning VM to a host group consisting of compute-intensive, latency-insensitive batch analysis VMs, or selects a candidate host belonging to a different NUMA node than the online payment VM. After migration, the instruction retirement rate degradation index of the online payment VM recovers from 2.3 to 1.1, and the p99 latency recovers from 300ms to 45ms. If the high-defense cleaning VM cannot be migrated immediately due to business continuity requirements, the system triggers a medium-risk resource isolation strategy, allocating a 3 / 20-way LLC cache to the high-defense cleaning VM via Intel CAT, limiting memory bandwidth to 35% of the total bandwidth. This strategy ensures that the cleaning operation performance of the high-defense cleaning VM decreases within an acceptable range (≤15%), while the performance of the online payment VM is fully restored, and the p99 latency stabilizes below 50ms.
[0069] This embodiment demonstrates that the present invention can effectively identify and avoid implicit contention at the microarchitecture level, and is especially suitable for scenarios such as high-defense cleaning VMs with sudden resource consumption and mixed with latency-sensitive services.
[0070] Compared to existing load balancing scheduling methods that rely solely on operating system-level metrics such as CPU and memory utilization, this embodiment can proactively identify and avoid implicit resource contention at the microarchitecture level, even when macroscopic metrics do not show anomalies. Through this method, this embodiment achieves the following technical effects: significantly improved scheduling accuracy; compared to scheduling methods using only OS-level metrics, microarchitecture contention prediction can reduce performance interference events between virtual machines on the same host by more than half; more stable service SLOs and significantly improved p99 latency; higher resource utilization; avoiding leaving a large safety margin for "invisible contention," resulting in significantly improved resource utilization; elimination of "false positive" scheduling; reverse migration of noise sources rather than victims, reducing the number of scheduling oscillations; and self-iterative optimization, with the contention relationship graph continuously updated with runtime data, gradually enhancing adaptability to new workloads and reducing the false negative rate over time.
[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.
Claims
1. A data center virtualization server load balancing scheduling method, characterized in that, Includes the following steps: S1. On each host machine, the micro-architecture level operation indicators of each virtual machine are collected by hardware performance counters at a preset sampling period. The micro-architecture level operation indicators include the last-level cache miss rate, memory bandwidth consumption, cache line invalidation frequency, TLB miss rate, branch predictor contention index, and instruction retirement rate degradation index. S2. Obtain the steady-state baseline of each virtual machine. The steady-state baseline is obtained by selecting a sampling window where the overall load of the host machine is lower than a preset load threshold as the steady-state window, taking a preset quantile value for each microarchitecture metric of the virtual machine within each steady-state window as the steady-state baseline; normalizing each microarchitecture metric based on the steady-state baseline, constructing a microarchitecture contention fingerprint vector for each virtual machine, and marking each virtual machine as stress-inducing or sensitive based on historical running behavior. S3. For each pair of virtual machines coexisting on a host machine, construct a microarchitecture contention graph across virtual machines based on time-series causality testing and latency mutual information analysis. The microarchitecture contention graph uses virtual machines as nodes, directed edges to represent the direction of microarchitecture resource suppression by one virtual machine on another virtual machine, and edge weights to represent the suppression intensity. S4. When it is necessary to migrate or deploy the target virtual machine, obtain the existing virtual machine set of all candidate hosts, query the existing contention edge weights between the target virtual machine and the existing virtual machines in each candidate host based on the microarchitecture contention relationship graph, and combine the existing contention edge weights with the contention fingerprint similarity to comprehensively calculate the pre-scheduling contention risk value between the target virtual machine and each candidate host. If there is no existing contested edge, the weight of the edge is recorded as zero, and the similarity of contested fingerprints is used as the main calculation basis. S5. Based on the pre-scheduling contention risk value, divide each candidate host into three levels: high risk, medium risk, and low risk. High-risk candidate hosts are directly eliminated. If the candidate host set is empty after elimination, it is processed according to the preset degradation strategy. Medium-risk candidate hosts retain their eligibility to participate in the scoring process, provided that resource isolation policies are configured synchronously; low-risk candidate hosts participate in the scoring process normally. S6. Combine the contention risk value of the candidate host after hierarchical processing with the load balancing score to form a scheduling decision function, sort the candidate host scores, and select the candidate host with the highest comprehensive score as the migration or deployment target of the target virtual machine. S7. After scheduling execution, continuously monitor the microarchitecture metrics of each virtual machine. When the monitored microarchitecture metrics meet the preset contention anomaly conditions, locate the victim virtual machine with impaired performance, locate the noise source virtual machine based on the microarchitecture contention relationship graph, prioritize the migration of the noise source virtual machine rather than the victim virtual machine, and update the edge weights of the microarchitecture contention relationship graph in reverse according to the actual contention results.
2. The method of claim 1, wherein, In step S2, the preset load threshold is that the overall load of the host machine is less than 30%, and the preset percentile value is the 95th percentile; in addition, the steady-state window also meets the following conditions: combined with the virtual machine service layer monitoring data, the fluctuation range of the service request arrival rate of the virtual machine within the steady-state window is less than the preset fluctuation threshold, and no virtual machine migration, snapshot or backup operation has occurred.
3. The method of claim 1, wherein, In step S3, the temporal causality test adopts the Granger causality test, and the time delay mutual information analysis is performed in the following manner: For a pair of virtual machines (VM_i, VM_j) coexisting on the host machine, calculate the latency mutual information between the change in the last-level cache miss rate of VM_i and the instruction retirement rate degradation index of VM_j. Scan a preset latency range, take the latency that maximizes the mutual information value as the optimal latency, and use the maximum mutual information value as the initial value of the edge weight. When the p-value of the Granger causality test is lower than the preset significance threshold and the mutual information value exceeds the preset mutual information threshold, establish a directed edge from VM_i to VM_j. The edge weight is obtained by weighted normalization of the p-value and the mutual information value.
4. The method of claim 1, wherein, In step S4, the contention fingerprint similarity is calculated as follows: When there is no existing contention edge between the target virtual machine and an existing virtual machine in the candidate host in the microarchitecture contention graph, the cosine similarity between the microarchitecture contention fingerprint vector of the target virtual machine and the microarchitecture contention fingerprint feature of the existing virtual machine marked as sensitive in the candidate host when it was suppressed in the past is calculated; if the cosine similarity exceeds a preset similarity threshold, the similarity contention risk value is calculated based on the cosine similarity and the intensity of the historical suppression of the sensitive virtual machine, and the similarity contention risk value is included in the contention risk value before scheduling; the microarchitecture contention fingerprint feature when it was suppressed in the past is the statistical feature of the microarchitecture contention fingerprint vector of the sensitive virtual machine during the period when the historical contention event occurred.
5. The method of claim 1, wherein, In step S5, the medium-risk candidate host retains its eligibility to participate in the scoring process, provided that a resource isolation policy is configured synchronously. Specifically: For medium-risk candidate host machines, a resource isolation policy based on cache allocation technology is generated. The resource isolation policy includes allocating a maximum of 8 paths for the last-level cache and a maximum of 40% of the memory bandwidth to the target virtual machine. The resource isolation strategy will be issued and executed along with the scheduling decision.
6. The data center virtualization server load balancing and scheduling method of claim 5, wherein, When all candidate hosts are classified as high-risk, scheduling is not rejected. Instead, the high-risk candidate host with the lowest contention risk value is selected as the scheduling target, and the resource isolation policy is forcibly issued to that host. The maximum number of last-level cache paths in the resource isolation policy is 4, and the maximum memory bandwidth is 20%.
7. The method of claim 1, wherein, In step S4, when calculating the pre-scheduling contention risk value, the historical contention fingerprint time series of the target virtual machine is used to predict the predicted contention fingerprint of the target virtual machine in the future preset time window. Based on the predicted contention fingerprint, the contention fingerprint similarity is calculated, and the pre-scheduling contention risk value is calculated in combination with the existing contention edge weights.
8. The method of claim 1, wherein, In step S3, the microarchitecture contention graph adopts a two-level update mechanism: The first level is real-time incremental update within the same host: After each sampling period is completed, only the edge weights between coexisting virtual machines within the same host are updated exponentially. When the number of consecutive occurrences of a new contention event exceeds a preset threshold, the new edge is officially added. When the weight of an old edge decays to below a preset removal threshold, the edge is removed. The second level is full-scale computation across the entire cluster: according to a preset long period, time-series causality checks and latency mutual information analyses are re-executed for all active virtual machines across the entire cluster, and the contention graph is reconstructed.
9. The method of claim 1, wherein, In step S7, the method for locating the noise source virtual machine is as follows: In the microarchitecture contention graph, find all incoming edges pointing to the victim virtual machine, and select the source virtual machines corresponding to the directed edges whose edge weights exceed the preset noise threshold to form a noise source candidate set; calculate the migration cost for each virtual machine in the noise source candidate set, and select the noise source virtual machine with the highest comprehensive score after weighted summation of migration cost and contention intensity to perform migration. When the candidate set of noise sources is empty or the migration cost of all noise source virtual machines is higher than the migration cost of the victim virtual machine, the victim virtual machine is migrated instead.
10. The method of claim 1, wherein, In step S7, the reverse update is performed as follows: The actual contention results after scheduling are compared with the predicted contention risks before scheduling: when the predicted contention risk is high and contention actually occurs, the weights of the corresponding directed edges are strengthened; when the predicted contention risk is low but contention actually occurs, missed events are recorded and the corresponding contention edges are supplemented; when the predicted contention risk is high but no contention actually occurs, the weights of the corresponding directed edges are accelerated to decay; after each update, the updated edge weights are normalized to ensure that the edge weights remain within the preset range.
