IaaS and cloud platform monitoring management method and system

Abstract

The invention discloses an IaaS and cloud platform monitoring management method and system, and relates to the field of cloud computing operation and maintenance, and the method comprises the steps: collecting original data, generating a uniform resource key, and constructing a resource topological graph; guardrail alarm judgment is carried out on the resource topological graph, a baseline deviation event is generated through one-step forward baseline judgment of a cold-start and hot-start long-short-term memory network, and a guardrail alarm and the baseline deviation event are combined to generate an effective alarm; judging a freezing mark based on the effective alarm, constructing an event graph, determining a root cause path of the event graph, extracting a disposal decision chain based on the freezing mark and the root cause path, and summarizing unfrozen and undisposed uniform resource keys to form an allowable arrangement set; and selecting non-real-time loads which are allowed to arrange the set, judging the resource topological graph to obtain candidate target hosts, and arranging the candidate target hosts to complete non-real-time load placement. According to the invention, an array server IaaS monitoring management closed loop is constructed.

Description

Technical Field

[0001] This invention relates to the field of cloud computing operations and maintenance, and in particular to a method and system for monitoring and managing IaaS and cloud platforms. Background Technology

[0002] The array server supports resource provisioning for IaaS and cloud platforms. Operations and maintenance typically collect host, network, and service metrics through out-of-band management and control plane interfaces, combine them with configuration management database mapping to form object relationships, and write them to time series and log systems for monitoring and processing through threshold alarms, trend dashboards, and work order flow. Conventional practices emphasize unified data collection standards, metric visualization, and process compliance, and use orchestration systems to perform operations and maintenance actions such as migration, scaling, and restart to ensure the continuity and security of multi-tenant environments.

[0003] For large-scale heterogeneous arrays, conventional methods rely heavily on static rules and manual closed loops for alarm triggering and handling linkage, making it difficult to form a consistent and verifiable triggering standard. At the same time, in terms of data acquisition end adaptation and version management, full-scale switching or manual release is often used, lacking a formalized path for canary verification of data acquisition end and consistent distribution of parameters across sites. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an IaaS and cloud platform monitoring and management method to solve the problems of consistent and verifiable alarm triggering and handling linkage in large-scale heterogeneous array scenarios, as well as the governance of canary verification and cross-site parameter consistent distribution governance for the data acquisition end.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an IaaS and cloud platform monitoring and management method, comprising,

[0008] Collect raw data and generate unified resource keys to construct a resource topology graph;

[0009] The resource topology map is subjected to guardrail alarm determination. The baseline deviation event is generated by the one-step forward baseline determination of the cold start and warm start long short-term memory network. The guardrail alarm and the baseline deviation event are merged to generate a valid alarm.

[0010] Based on valid alarms, determine the freeze flag, construct an event graph, determine the root cause path of the event graph, extract the handling decision chain based on the freeze flag and root cause path, and summarize the unfrozen and unhandled unified resource keys to form an allowed orchestration set.

[0011] Select non-real-time loads that are allowed to be orchestrated and determine the resource topology to obtain candidate target hosts. Then, place the non-real-time loads by arranging the candidate target hosts.

[0012] Determine whether the heterogeneous array server acquisition terminal using the canary deployment is allowed to increase the throughput, calculate the global parameter vector that allows the throughput to increase as the monitoring baseline version configuration and distribute it to the acquisition terminal site, and update the monitoring baseline configuration.

[0013] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the specific steps for collecting raw data and generating a unified resource key are as follows:

[0014] The out-of-band interface and control plane interface obtain resource identifiers and attributes as raw data, and obtain normalized records through field specification;

[0015] The normalized records are concatenated into a string, and an irreversible Uniform Resource Key is calculated using a hash function.

[0016] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the resource topology graph includes tenant nodes, resource nodes, host nodes, ownership edges, resident edges, and resource relationship edges.

[0017] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the resource topology map is used for guardrail alarm determination, specifically through the following steps:

[0018] Based on the unified resource key, and with power consistency, memory consistency, temperature limit and link error correction rate limit as prerequisites, guardrail alarms are determined and generated.

[0019] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the steps of generating a baseline deviation event by determining the one-step forward baseline of the cold start and warm start long short-term memory network, and merging guardrail alarms and baseline deviations to generate a valid alarm, are as follows:

[0020] To unify the monitoring standards by binding indicator sequences to a unified resource key, the raw readings of the unified resource key are captured using the unified resource key and indicator names as index dimensions. The observations are then aligned to the time window. The indicator sequences are then de-shaken, smoothed, and quantile-based to obtain a comprehensive representative value.

[0021] When the number of samples available for training is lower than the minimum number of samples required to enter the learning branch, a cold start baseline is generated by combining representative values.

[0022] Once the number of available training samples is greater than or equal to the minimum number of samples to enter the learning branch, the observations aligned by the time window are used to generate a warm-start baseline using a long short-term memory network. The cold-start baseline and the warm-start baseline are merged and output as a unified baseline in the same time grid. Baseline deviation is then determined, and baseline deviation events are generated. Guardrail alarms and baseline deviation events are merged in the same time grid to generate a valid alarm.

[0023] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the specific steps of determining the freeze flag based on valid alarms are as follows:

[0024] Based on valid alarms, the scope of freezing is determined to be the unified resource key of the valid alarm and all unified resource keys on the host node of the unified resource key of the valid alarm.

[0025] Set a freeze window time. During the freeze window time, the scope is frozen and migration and scaling actions are prohibited. Only isolation and logging actions are allowed to enter the execution queue.

[0026] The freezing flag is determined based on the freezing scope and the freezing window time.

[0027] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the steps of constructing an event graph, determining root cause paths, extracting the handling decision chain, and summarizing unfrozen and unhandled Uniform Resource Keys to form an allowed orchestration set are as follows:

[0028] Construct an event graph with effective alarms as event nodes, determine the root cause path based on minimum hop count and time delay consistency, perform action matching and record based on the root cause path and freeze flag, accumulate records to obtain the handling decision chain, and summarize the unfrozen and unhandled unified resource keys to obtain the allowed orchestration set.

[0029] As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the steps of selecting non-real-time loads from an allowed orchestration set and determining candidate target hosts based on the resource topology graph, and then arranging the candidate target hosts to complete the placement of non-real-time loads, are as follows:

[0030] Non-real-time loads are filtered out based on the allowed orchestration set, and candidate target hosts are obtained by performing migration compliance and feasibility judgments on the resource topology graph;

[0031] Candidate target hosts are arranged in a fixed order to obtain a target host queue. Non-real-time loads are then sequentially placed on the target host queue using an initial adaptive decreasing approach until all loads are placed and the single-batch orchestration limit is reached. As a preferred embodiment of the IaaS and cloud platform monitoring and management method described in this invention, the steps of determining whether the heterogeneous array server acquisition terminal using canary deployment is allowed to increase load, calculating the global parameter vector allowing for increased load as the monitoring baseline version configuration, distributing it to the acquisition terminal site, and updating the monitoring baseline configuration are as follows:

[0032] If canary deployment is used, the defect count and sampling alarm rate are statistically analyzed using a uniform sampling step size and a fixed window, and compared with those not using canary deployment. If the defect count is zero and the difference in sampling alarm rates does not exceed the alarm rate difference threshold, the deployment is considered successful.

[0033] The global parameter vector is generated based on the robust standardized value and the length of the forward input sequence of the learning branch. The local parameter vectors obtained through local training are combined with the local parameter vectors of each data acquisition station to form the global parameter vector. The global parameter vector is then distributed to the data acquisition stations as the monitoring baseline version configuration to update the monitoring baseline configuration.

[0034] Secondly, this invention provides an IaaS and cloud platform monitoring and management system, including,

[0035] The data acquisition and mapping module collects raw data, generates unified resource keys, and constructs a resource topology map.

[0036] The guardrail baseline module performs guardrail alarm determination on the resource topology map, generates baseline deviation events by using the one-step forward baseline determination of cold start and warm start long short-term memory networks, and merges guardrail alarms and baseline deviation events to generate valid alarms.

[0037] The root cause freezing module determines the freezing flag based on valid alarms, constructs an event graph, determines the root cause path of the event graph, extracts the handling decision chain based on the freezing flag and root cause path, and summarizes the unfrozen and unhandled unified resource keys to form an allowed orchestration set.

[0038] The orchestration and placement module selects non-real-time loads that are allowed to be orchestrated and determines candidate target hosts by judging the resource topology. It then completes the placement of non-real-time loads by arranging the candidate target hosts.

[0039] The acceptance and distribution module determines whether the heterogeneous array server acquisition terminal using the canary deployment is allowed to increase the volume. It calculates the global parameter vector that allows for increased volume and distributes it as the monitoring baseline version configuration to the acquisition terminal site, and updates the monitoring baseline configuration.

[0040] The beneficial effects of this invention are as follows: By generating baseline deviation events through a one-step forward baseline determination using a cold-start and warm-start long short-term memory network, and merging guardrail alarms and baseline deviations to generate valid alarms, a unified trigger is achieved, serving as the entry point for freeze flags, event graphs, and handling decision chains, thus achieving rapid location and steady-state control; based on canary release acceptance, it determines whether the newly connected heterogeneous array server acquisition end is allowed to increase the volume, calculates the global parameter vector that allows the volume increase as the new baseline parameter package to be distributed to the site, and achieves parameter governance and cross-site consistent updates, thus achieving continuous evolution and low-risk release. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart for monitoring and management methods for IaaS and cloud platforms.

[0043] Figure 2 This is a schematic diagram of an IaaS and cloud platform monitoring and management system.

[0044] Figure 3 A flowchart for unified resource key generation and resource topology graph construction.

[0045] Figure 4 This is a flowchart for obtaining guardrail alarms and valid alarms. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figures 1-4As one embodiment of the present invention, this embodiment provides an IaaS and cloud platform monitoring and management method, including the following steps:

[0050] S1. Collect raw data and generate a unified resource key to construct a resource topology graph.

[0051] Collect raw records, obtain host identifiers, blade tag numbers and port status based on out-of-band inventory, pull projects, tenants, quotas, compute hosts, virtual machines, ports, routers, volumes (CinderVolume) and images (GlanceImage) through the OpenStack control plane interface, and pull nodes, namespaces, container groups and deployment controllers through the Kubernetes control plane interface.

[0052] The original records are normalized by using field normalization functions to achieve lowercase conversion, removal of leading and trailing whitespace, unit standardization (e.g., capacity is standardized to GiB, speed is standardized to Mb / s), timestamp conversion to UTC time, and deduplication of identifiers, resulting in normalized records.

[0053] The resource key generation function is used to concatenate the normalized records into a string, and the hash function SHA-256 is used to calculate the unified resource key. The unified resource key is then output in hexadecimal lowercase and is used as a unique identifier for the four links of data collection, alarm, handling and metering.

[0054] The resource key generation function concatenates the following into a string: tenant identifier, project identifier, region identifier, availability zone identifier, host identifier, resource type, and control plane native ID.

[0055] Furthermore, a resource topology graph is constructed, with nodes divided into tenant nodes, resource nodes, and host nodes, and edges divided into ownership edges, resident edges, and resource relationship edges.

[0056] Tenant nodes represent tenant entities and are used for end-to-end referencing based on a unified resource key. Resource nodes represent cloud resource instances that can be measured and managed. One cloud resource instance represents one resource node. Host nodes represent the array server host machine that actually carries the resources.

[0057] Ownership edges point from tenant nodes to resource nodes, indicating the ownership relationship of resources belonging to tenants. Residency edges point from resource nodes to host nodes, indicating the residency relationship of resources currently running and mounted on the array server host. Resource relationship edges connect two resource nodes, indicating the direct reference and management relationship on the control plane. Resource relationship edges point from the caller and manager to the dependent and managed parties.

[0058] Resource relationship edges are distinguished into weak and strong relationships, with three criteria: same tenant, same host, and direct reference relationship to the control plane (e.g., port binding to router, virtual machine volume mounting, and container group belonging to deployment controller). For any two resource nodes, the number of times the three criteria are true is counted. When two or more of the three criteria are true, the resource relationship edge is marked as a strong relationship. When only one of the three criteria is true, it is marked as a weak relationship. When none of the three criteria are true, no resource relationship edge is established. If multiple candidate edges with the same strength appear, the edge containing the direct reference relationship to the control plane is selected first, followed by the edge with the same host.

[0059] The resource topology graph is stored in a node table and an edge table. The node table includes a unified resource key, node type, node ownership information, node display name, and node record timestamp. The edge table includes the edge start point unified resource key, edge end point unified resource key, edge type, relationship strength, and record timestamp.

[0060] Furthermore, a list of monitoring objects is formed based on the node table of the resource topology graph. The original records of the monitoring objects are retrieved using the unified resource key as an index, and a monitoring baseline configuration is established.

[0061] S2. Perform guardrail alarm determination on the resource topology graph, generate baseline deviation events by determining the one-step forward baseline of the cold start and warm start long short-term memory network, and merge guardrail alarms and baseline deviation events to generate valid alarms.

[0062] Based on the resource topology map, guardrail alarms are generated by comparing the product of power supply voltage and current with the power meter reading. The upper limit of the power meter's accuracy is used as the error value. If the product of power supply voltage and current and the power meter reading are inconsistent, it is judged as invalid data and hardware abnormality, and a guardrail alarm is generated directly. The used memory of the resource node is compared with the total memory of the resource node. The used memory of the resource node must not exceed the total memory of the resource node. If it exceeds, it is regarded as memory count abnormality and memory count leakage, and an alarm is generated directly, blocking further judgment. The device temperature on the host node is read and compared with the temperature upper limit given by the manufacturer. If the device temperature on the host node reaches the temperature upper limit given by the manufacturer, it is considered as thermal risk, and an alarm is generated directly, along with suggestions for load reduction and migration. The error rate of the device link on the host node is read and compared with the upper limit of the manufacturer's specifications. If it exceeds the upper limit, it is judged as device link quality abnormality and media aging on the host node, and an alarm is generated directly, along with suggestions for switching links and replacing components.

[0063] If any item fails, the unified resource key of the guardrail alarm is stored in the alarm record table. If all items pass, a forward baseline and deviation ratio are generated for the same resource key and the same caliber indicator.

[0064] The same caliber metric is based on the Uniform Resource Key (URI). The URI is bound to the corresponding metric. The URI and the metric name are used as index dimensions. The uniform sampling step size and sliding window length are set. At each sampling time, the raw readings corresponding to the URI are captured from the Kubernetes metric interface and out-of-band interface. The raw point series is obtained by arranging them in chronological order. Any sampling time is mapped to a uniform time grid. Missing and conflict are handled to obtain the observations after time window alignment.

[0065] The uniform sampling step size is the time granularity of all sampling surfaces on the same time axis. The value is not less than the capture period of any sampling end and is a common multiple of the capture period of the sampling ends. The value range is 1–60 seconds, with a default of 10 seconds.

[0066] The sliding window length is set to cover the minimum judgment time and filter out instantaneous jitter; the default value is 60 seconds.

[0067] Furthermore, to reduce random jitter, a simple moving average is used on the original point series sorted by time. That is, the arithmetic mean of the existing w sample points is calculated, and if there are less than w, the arithmetic mean of all sample points is calculated. Window smoothing is performed on each curve to obtain the current smoothed value of each index for each unified resource key.

[0068] w is the number of nearest samples used for smoothing and quantile calculations, obtained by the ratio of the sliding window length to the uniform sampling step size, with a value range of [3, 30].

[0069] The comprehensive representative value is calculated based on the aligned observations within the window, and the expression is as follows:

[0070]

[0071] Where i is the window sample index, j is the window sample index after i, and t i It represents the i-th sampling time within the window, where k is the Uniform Resource Key and m is the index name. At time t i For aligned observations of the index m of the Uniform Resource Key, median is the median value. The median slope coefficient is obtained by fitting a linear trend to the observations after the time window is aligned using Theil–Sen robust linear regression, resulting in a robust regression line. The median slope coefficient is then calculated from the median slope of the robust regression line. It is the median intercept coefficient. The median intercept coefficient is obtained by calculating the median intercept of the robust regression line. k,m (t) is the comprehensive representative value.

[0072] Set a minimum sample size threshold for entering the learning branch, which is the number of time steps required for ten predictions. The number of supervised samples can grow linearly, and the sample pairs and input dimensions should be at or above the same order of magnitude to avoid overfitting with small samples.

[0073] If the number of samples available for training is lower than the minimum sample size threshold for entering the learning branch, a cold start value is used for determination. The arithmetic mean of the comprehensive representative value of the most recent sample points is calculated to obtain the cold baseline value at the current time. If the number of most recent sample points is not met, the average of the available points is calculated.

[0074] When the number of samples available for training exceeds the minimum sample size threshold for entering the learning branch, a warm-start value is used for determination, expressed as:

[0075]

[0076] Where τ is the time index in the input sequence of the learning branch, τ∈{t2-L,t2-T} s}, t2 is the current decision point in the learning branch, t2≥t * +T s , t * The first trainable moment is when the number of samples available for training equals the minimum sample size threshold required to enter the learning branch. This is determined using the cold start value, T. s It is the uniform sampling step size, L is the length of the forward input sequence in one step of the learning branch, and z k,m (τ) is a robust standardized value within the time window. Inside, put Dimensionless transformation is performed using the median and the absolute median difference. It is a window The median within It is a window The absolute median difference within the range, ε is a standardized positive number that prevents zero, and its value range is

[10] . -6 10 -3 ], It is a standardized window. It is an anti-normalization mapping. It is the baseline value of the learning branch, i.e., the original dimensional value of the one-step forward prediction, f k,m It is the LSTM prediction function.

[0077] Furthermore, based on the cold start value determination and the hot start value determination, a unified baseline is obtained, expressed as:

[0078]

[0079] in, This is the cold baseline value, N k,m (t) is the number of valid samples up to t, N min It is the minimum sample size threshold for entering the learning branch. It is a unified baseline.

[0080] Furthermore, the deviation ratio is calculated based on the unified baseline, expressed as:

[0081]

[0082] Where, δ k,m (t) is the deviation ratio, where a positive observation indicates that the observation is above the baseline and a negative observation indicates that the observation is below the baseline.

[0083] Baseline deviation is determined based on the deviation ratio. For example, when the deviation from the unified baseline exceeds ±30% during hot start, a baseline deviation event is triggered, local self-healing is activated, and a baseline deviation event is generated. Guardrail alarms and baseline deviation events are merged within the same time grid to generate a valid alarm, which is recorded in the alarm record table. A work order is generated and the alarm record table is included in the work order.

[0084] S3. Based on valid alarms, determine the freeze flag, construct an event graph, determine the root cause path of the event graph, extract the handling decision chain based on the freeze flag and root cause path, and summarize the unfrozen and unhandled unified resource keys to form an allowed orchestration set.

[0085] Based on valid alarms in the alarm log table, a freeze flag is determined. The freeze scope includes the unified resource key of the valid alarm and all unified resource keys on the host node of the valid unified resource key. A freeze window time is set. During the freeze window time, migration and scaling actions are prohibited in the freeze scope. Only isolation and recording actions are allowed to enter the execution queue. When the unified resource key no longer triggers valid alarms and continues to not trigger valid alarms for more than the release retention time, the freeze flag is automatically cleared, or it is cleared by a manual release command, and the operator is recorded.

[0086] The freeze window time is the strategy duration determined based on the manufacturer's upper temperature limit, error rate limit, and uniform sampling step size for guardrail alarm alignment.

[0087] The hold-off duration is the minimum duration of no alarms required to confirm the stability of effective alarm fading. The hold-off duration is obtained by reading the hold-off duration values ​​of 1-5 minutes from the same deployment parameter file.

[0088] Furthermore, based on the resource topology graph, valid alarms are extracted as event nodes, and an event graph consisting of a set of nodes and a set of directed edges is constructed using linear scanning, time sorting, and conditional edge construction.

[0089] Status events are deterministic events generated at a uniform sampling time by comparing real-time status fields such as port status, temperature, and correction error rate with thresholds such as manufacturer's upper temperature limit and manufacturer's upper error rate limit.

[0090] A node set is a collection of event instances consisting of a Uniform Resource Key, a timestamp, and an event type.

[0091] A set of directed edges is a set of directed connections between event pairs that satisfy the conditions of temporal order and the first hop and time delay of events falling into the window on the same exchange.

[0092] Furthermore, based on the event graph, the root cause path is selected in lexicographical order using the shortest number of hops as the first criterion and the time delay deviation as the second criterion.

[0093] Starting from all event nodes with an in-degree of 0, a breadth-first search algorithm is used to obtain the shortest hop path to all terminal nodes. If multiple shortest hop paths exist, the path with the fewest hops is selected first. If hop counts are tied, the time delay deviation of the shortest path is calculated, and the path with the smallest time delay deviation is selected. The expression is:

[0094]

[0095] Among them, h a It is event node u a The alignment timestamp, Δ0 is the reference propagation delay constant, taken as a value equivalent to the sampling step size, P is a directed path on the event graph, (u a →u j )∈P represents event u a Prior to u j The occurrence and satisfaction of the first hop and time delay falling into the window constraints of the same switch, u a It is the a-th event node in the event graph, Δ dev (P) is the maximum time delay deviation of path P.

[0096] If time delay deviations are parallel, the path with the earliest start time is selected, and the starting event node of the root cause path is denoted as the root cause event.

[0097] Furthermore, the node set of the event graph prioritizes root cause events based on the freeze flag to perform restricted action matching, completes isolation and loss mitigation, shortens the average repair time and avoids secondary failures. After the constraints are lifted, it resumes to full action. If there is no freeze flag, it performs full action matching, generates and executes them sequentially according to the root cause starting point priority and alarm level, and records them, accumulating to obtain the handling decision chain.

[0098] Based on valid alarms and root cause path determination, freeze flags are identified and a handling decision chain is executed. Work orders are generated and synchronously written to the audit table. Unfrozen and unhandled Uniform Resource Keys are aggregated into an allowed orchestration set and included in the work order.

[0099] S4. Select non-real-time loads that are allowed to be orchestrated and determine the resource topology to obtain candidate target hosts. Arrange the candidate target hosts to complete the placement of non-real-time loads.

[0100] Based on the allowed orchestration set of unified resource keys, the unified resource keys of non-real-time loads are filtered out, and candidate target action pairs are divided according to the alarm source. For example, host hotspots correspond to migrating virtual machines and evict container groups, platform service degradation corresponds to restarting services, network degradation corresponds to switching routers, and resource shortages correspond to creating elastic computing instances.

[0101] Furthermore, the candidate target hosts are screened to determine the compliance and feasibility of migration, resulting in a set of feasible target hosts.

[0102] The compliance assessment for migration involves comparing each constraint of the candidate target host in terms of isolation domain, availability zone, affinity, anti-affinity, and license list. Only if all constraints are met will the capacity check proceed.

[0103] The feasibility assessment is based on whether the CPU capacity is sufficient, whether the bandwidth capacity is sufficient, and whether the total allocation of video memory does not exceed 4 times the physical video memory, and the allocation of each graphics card does not exceed 4 times the physical video memory of that graphics card.

[0104] The feasible target host set is deterministically selected using lexicographical order rules. The host with the most remaining video memory is selected first. If there is a tie, the host with the most remaining CPU is selected from the ties. If there is a second tie, the host with the most remaining bandwidth is selected. If all three are exactly the same, the feasible target host with the fewest instances to be migrated in this round of orchestration is selected to obtain an ordered target host queue.

[0105] Furthermore, candidate target action pairs are extracted to obtain a list of objects to be placed. Using the first-adaptive decreasing method, after determining the bottleneck dimension, the non-real-time loads are sorted from largest to smallest. Feasibility is determined one by one according to the ordered target host queue and the non-real-time loads are placed. After each placement, the host's used resources and remaining resources are updated in real time. This process is repeated until all are placed or the single batch arrangement limit is reached.

[0106] The single-batch orchestration limit is the maximum number of non-real-time load instances that can be placed concurrently within a single execution cycle.

[0107] The bottleneck dimension is the resource dimension with the highest total demand for a given resource dimension for the object to be placed.

[0108] S5. Determine whether the heterogeneous array server acquisition terminal released by the canary is allowed to increase the volume, calculate the global parameter vector that allows the volume to increase as the monitoring baseline version configuration and send it to the site, and update the monitoring baseline configuration.

[0109] The canary release method is used to verify the newly connected heterogeneous array server acquisition terminal. The canary release extracts 5% and 10% by default. This invention extracts 5% of the unified resource key and switches to the new acquisition terminal. Under the premise of not changing the guardrail alarm, baseline deviation trigger, freeze constraint and allowed orchestration set, the migration compliance judgment, feasibility judgment, canary release statistical field defects are 0, the alarm rate difference per 10,000 points of sampling is lower than the alarm rate difference threshold and traceable work orders are met. If any one of them is not met, it will be rolled back. If all of them are met, the canary release is judged to be approved and the volume can be increased.

[0110] The defects in the statistical fields published by the canary are aggregated according to a uniform sampling step size and observed by summing within a window to count the defects in the statistical fields published by the canary.

[0111] The difference in alarm rate per 10,000 points is obtained by comparing the alarm rate per 10,000 points obtained by aggregating with a uniform sampling step size and summing by window with the original alarm rate per 10,000 points before the canary release was adopted.

[0112] The alarm rate difference threshold is obtained by averaging the differences between adjacent windows based on the most recent alarm rate sequences per 10,000 points before the canary deployment was implemented.

[0113] Furthermore, the robust standardized value and the length of the forward input sequence of the learning branch are used to form a supervised pair. Several rounds of local training are performed using mean squared error to obtain the local parameter vector. The local parameter vectors and sample sizes of the data collection stations are summarized and arithmetically averaged according to the sample size ratio to obtain the global parameter vector. The global parameter vector is then distributed to the data collection stations as the monitoring baseline version configuration to update the monitoring baseline configuration.

[0114] This embodiment also provides an IaaS and cloud platform monitoring and management system, including:

[0115] The data acquisition and mapping module collects raw data, generates unified resource keys, and constructs a resource topology map.

[0116] The guardrail baseline module performs guardrail alarm determination on the resource topology map, generates baseline deviation events by using the one-step forward baseline determination of cold start and warm start long short-term memory networks, and merges guardrail alarms and baseline deviation events to generate valid alarms.

[0117] The root cause freezing module determines the freezing flag based on valid alarms, constructs an event graph, determines the root cause path of the event graph, extracts the handling decision chain based on the freezing flag and root cause path, and summarizes the unfrozen and unhandled unified resource keys to form an allowed orchestration set.

[0118] The orchestration and placement module selects non-real-time loads that are allowed to be orchestrated and determines candidate target hosts by judging the resource topology. It then completes the placement of non-real-time loads by arranging the candidate target hosts.

[0119] The acceptance and distribution module determines whether the heterogeneous array server acquisition terminal using the canary deployment is allowed to increase the volume. It calculates the global parameter vector that allows for increased volume and distributes it as the monitoring baseline version configuration to the acquisition terminal site, and updates the monitoring baseline configuration.

[0120] This embodiment also provides a computer device applicable to IaaS and cloud platform monitoring and management methods, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the IaaS and cloud platform monitoring and management method proposed in the above embodiment.

[0121] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0122] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the IaaS and cloud platform monitoring and management method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0123] In summary, this invention achieves rapid location and steady-state control by: generating baseline deviation events through a one-step forward baseline determination using cold-start and warm-start long short-term memory networks; merging guardrail alarms and baseline deviations to generate valid alarms; unifying triggering and serving as the entry point for freeze flags, event graphs, and the handling decision chain; and verifying whether newly accessed heterogeneous array server acquisition terminals using canary deployments are allowed to increase throughput, calculating the global parameter vector that allows for throughput increases as the new baseline parameter package for distribution to sites, thus achieving parameter governance and consistent updates across sites, and enabling continuous evolution and low-risk deployment.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A monitoring and management method for IaaS and cloud platforms, characterized in that: include, Collect raw data and generate unified resource keys to construct a resource topology graph; The resource topology map is subjected to guardrail alarm determination. The baseline deviation event is generated by the one-step forward baseline determination of the cold start and warm start long short-term memory network. The guardrail alarm and the baseline deviation event are merged to generate a valid alarm. Based on valid alarms, determine the freeze flag, construct an event graph, determine the root cause path of the event graph, extract the handling decision chain based on the freeze flag and root cause path, and summarize the unfrozen and unhandled unified resource keys to form an allowed orchestration set. Select non-real-time loads that are allowed to be orchestrated and determine the resource topology to obtain candidate target hosts. Then, place the non-real-time loads by arranging the candidate target hosts. Determine whether the heterogeneous array server acquisition terminal using the canary deployment is allowed to increase the throughput, calculate the global parameter vector that allows the throughput to increase as the monitoring baseline version configuration and distribute it to the acquisition terminal site, and update the monitoring baseline configuration.

2. The IaaS and cloud platform monitoring and management method as described in claim 1, characterized in that: The specific steps for collecting raw data and generating a unified resource key are as follows: The out-of-band interface and control plane interface obtain resource identifiers and attributes as raw data, and obtain normalized records through field specification; The normalized records are concatenated into a string, and an irreversible Uniform Resource Key is calculated using a hash function.

3. The IaaS and cloud platform monitoring and management method as described in claim 2, characterized in that: The resource topology graph includes tenant nodes, resource nodes, host nodes, ownership edges, resident edges, and resource relationship edges.

4. The IaaS and cloud platform monitoring and management method as described in claim 3, characterized in that: The resource topology map is used to determine guardrail alarms. The specific steps are as follows: Based on the unified resource key, and with power consistency, memory consistency, temperature limit and link error correction rate limit as prerequisites, guardrail alarms are determined and generated.

5. The IaaS and cloud platform monitoring and management method as described in claim 4, characterized in that: The process involves using a cold-start and warm-start long short-term memory network to determine the forward baseline and generate a baseline deviation event. This is then combined with guardrail alarms and baseline deviations to generate a valid alarm. The specific steps are as follows: To unify the monitoring standards by binding indicator sequences to a unified resource key, the raw readings of the unified resource key are captured using the unified resource key and indicator names as index dimensions. The observations are then aligned to the time window. The indicator sequences are then de-shaken, smoothed, and quantile-based to obtain a comprehensive representative value. When the number of samples available for training is lower than the minimum number of samples required to enter the learning branch, a cold start baseline is generated by combining representative values. Once the number of available training samples is greater than or equal to the minimum number of samples to enter the learning branch, the observations aligned by the time window are used to generate a warm-start baseline using a long short-term memory network. The cold-start baseline and the warm-start baseline are merged and output as a unified baseline in the same time grid. Baseline deviation is then determined, and baseline deviation events are generated. Guardrail alarms and baseline deviation events are merged in the same time grid to generate a valid alarm.

6. The IaaS and cloud platform monitoring and management method as described in claim 5, characterized in that: The specific steps for determining the freeze flag based on valid alarms are as follows: Based on valid alarms, the scope of freezing is determined to be the unified resource key of the valid alarm and all unified resource keys on the host node of the unified resource key of the valid alarm. Set a freeze window time. During the freeze window time, the scope is frozen and migration and scaling actions are prohibited. Only isolation and logging actions are allowed to enter the execution queue. The freezing flag is determined based on the freezing scope and the freezing window time.

7. The IaaS and cloud platform monitoring and management method as described in claim 6, characterized in that: The specific steps for constructing the event graph, determining the root cause path, extracting the action decision chain, and summarizing the unfrozen and unprocessed Uniform Resource Keys to form an allowed orchestration set are as follows: Construct an event graph with effective alarms as event nodes, determine the root cause path based on minimum hop count and time delay consistency, perform action matching and record based on the root cause path and freeze flag, accumulate records to obtain the handling decision chain, and summarize the unfrozen and unhandled unified resource keys to obtain the allowed orchestration set.

8. The IaaS and cloud platform monitoring and management method as described in claim 7, characterized in that: The steps involve selecting non-real-time loads that can be orchestrated from the allowed set and determining candidate target hosts based on the resource topology. The placement of non-real-time loads is then completed by arranging these candidate target hosts. Non-real-time loads are filtered out based on the allowed orchestration set, and candidate target hosts are obtained by performing migration compliance and feasibility judgments on the resource topology graph; Candidate target hosts are arranged in a fixed order to obtain a target host queue. Non-real-time loads are placed sequentially on the target host queue through first-fit decrement until all are placed and the single-batch orchestration limit is reached.

9. The IaaS and cloud platform monitoring and management method as described in claim 8, characterized in that: The determination of whether the heterogeneous array server data acquisition terminal deployed by the canary server is allowed to increase throughput, the calculation of the global parameter vector allowing for increased throughput as the monitoring baseline version configuration, the distribution of this configuration to the data acquisition terminal sites, and the updating of the monitoring baseline configuration are as follows: If canary deployment is used, the defect count and sampling alarm rate are statistically analyzed using a uniform sampling step size and a fixed window, and compared with those not using canary deployment. If the defect count is zero and the difference in sampling alarm rates does not exceed the alarm rate difference threshold, the deployment is considered successful. The global parameter vector is generated based on the robust standardized value and the length of the forward input sequence of the learning branch. The local parameter vectors obtained through local training are combined with the local parameter vectors of each data acquisition station to form the global parameter vector. The global parameter vector is then distributed to the data acquisition stations as the monitoring baseline version configuration to update the monitoring baseline configuration.

10. An IaaS and cloud platform monitoring and management system, based on the IaaS and cloud platform monitoring and management method according to any one of claims 1 to 9, characterized in that: include, The data acquisition and mapping module collects raw data, generates unified resource keys, and constructs a resource topology map. The guardrail baseline module performs guardrail alarm determination on the resource topology map, generates baseline deviation events by using the one-step forward baseline determination of cold start and warm start long short-term memory networks, and merges guardrail alarms and baseline deviation events to generate valid alarms. The root cause freezing module determines the freezing flag based on valid alarms, constructs an event graph, determines the root cause path of the event graph, extracts the handling decision chain based on the freezing flag and root cause path, and summarizes the unfrozen and unhandled unified resource keys to form an allowed orchestration set. The orchestration and placement module selects non-real-time loads that are allowed to be orchestrated and determines candidate target hosts by judging the resource topology. It then completes the placement of non-real-time loads by arranging the candidate target hosts. The acceptance and distribution module determines whether the heterogeneous array server acquisition terminal using the canary deployment is allowed to increase the volume. It calculates the global parameter vector that allows for increased volume and distributes it as the monitoring baseline version configuration to the acquisition terminal site, and updates the monitoring baseline configuration.

Images