A multi-layer network state perception and decision method for cloud computing distributed system

Abstract

The application discloses a kind of multi-layer network state perception and decision method for cloud computing distributed system, comprising: collecting each layer network original state index data and environmental factor data;Environment factor deviation calculation is carried out, and standard environment equivalent value is generated by environment normalization correction, the environmental attribution degree for representing the probability of explaining anomaly by environmental factors is calculated, and recoverable expected label and judgment confidence are generated;Establish cross-layer mapping relationship;From hierarchical, space-time, causality, granularity and confidence five dimensions, decision resolution is generated, and hierarchical decision scheme is generated;The corresponding operation and tracking feedback are executed, the environmental factor is included in hardware index correction calculation by the application, and environment stress type and ontology essential anomaly are effectively distinguished, permanent property and transient property are differentiated, and the problem that performance reduction caused by environmental factors is misjudged as hardware failure is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a multi-layer network state perception and decision-making method for cloud computing distributed systems. Background Technology

[0002] Cloud computing distributed systems have evolved from single clusters to complex, multi-layered architectures spanning data centers and regions. The network layer exhibits a topology interwoven with four layers: physical network, virtual network, service network, and application network. Network state awareness and decision-making in cloud computing distributed systems primarily involve: eBPF-based observability technology, which uses kernel-mode program injection to achieve network event tracking and deep metric collection; causal inference-based root cause localization technology, which achieves accurate root cause analysis by modeling service and resource dependencies and considering event temporal characteristics; deep reinforcement learning-based intelligent decision-making technology, applied to computing power routing and resource scheduling; and data center environmental monitoring technology, which uses sensor networks to achieve real-time monitoring and threshold alarms for multi-dimensional environmental parameters such as temperature, humidity, air pressure, and harmful gases in the computer room, ensuring the normal operation of infrastructure such as UPS and precision air conditioning.

[0003] While data center monitoring systems deploy sensors for temperature, humidity, and air pressure, this data is not used to correct IT hardware performance metrics. For example, high altitudes can cause disk seek times to increase, high temperatures can lead to SSD frequency reduction, and thunderstorms can cause optical module error rates to spike. These environmental impacts are not considered as preconditions for network status perception data in practical applications. Furthermore, these environmental impacts are categorized into permanent and transient effects, making it difficult to make differentiated decisions for different types of impacts. This results in overreaction to permanent impacts or insufficient response to transient disturbances. Therefore, this paper proposes a multi-layered network status perception and decision-making method for cloud computing distributed systems to address these issues. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-layer network state perception and decision-making method for cloud computing distributed systems, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for multi-layer network state awareness and decision-making in cloud computing distributed systems includes the following steps:

[0007] Step 1: Collect raw state index data and environmental factor data of each network layer by deploying a layered perception agent at the physical network layer, virtual network layer, service network layer, and application network layer; the environmental factor data includes at least one or more of atmospheric pressure, ambient temperature, relative humidity, electromagnetic field strength, mechanical vibration, and altitude.

[0008] Step 2: Input the original state index data and the environmental factor data into the environmental assessment layer. The environmental assessment layer performs the following operations:

[0009] Based on the preset sensitivity coefficients of each hardware type to different environmental factors, calculate the deviation of at least one environmental factor from the calibrated environmental value.

[0010] Based on the deviation and the preset sensitivity coefficient, the original state index data is normalized and corrected to generate a standard environmental equivalent value.

[0011] Based on the difference between the original state index data and the standard environmental equivalent value, the environmental attribution degree is calculated; the environmental attribution degree is used to characterize the probability that the anomalies in the original state index data are explained by environmental factors.

[0012] Based on the attribute type of the environmental factors, generate recoverable expected labels; the attribute types include permanent attributes and transient attributes;

[0013] Based on the health status and data freshness of the environmental sensor network, a confidence score is generated.

[0014] Step 3: Input the standard environmental equivalent value corrected by the environmental assessment layer, the environmental attribution degree, the recoverable expected label, and the assessment confidence into the fusion analysis layer to establish cross-layer mapping relationships and correlation analysis between network states of each layer;

[0015] Step 4: Based on the output of the fusion analysis layer, make decisions about the network state from five dimensions: hierarchy, spatiotemporal, causality, granularity, and confidence, and generate a hierarchical decision scheme.

[0016] Step 5: Perform corresponding network operation and maintenance operations according to the hierarchical decision-making scheme, and track and provide feedback on the execution results to optimize subsequent decisions.

[0017] Preferably, the environmental normalization correction adopts a Bayesian multi-factor compensation model, and the standard environmental equivalent value SE is calculated in the following way:

[0018]

[0019] in, The original state index data, For the first The main effect sensitivity coefficient of each environmental factor For the first The deviation of an environmental factor from its calibrated environmental value For the first The and the first The coupling effect coefficient between environmental factors.

[0020] Preferably, the environmental attribution EA is calculated in the following manner:

[0021] when hour, ;

[0022] when hour, ;

[0023] According to the above The values ​​categorize exception types as follows: This is an environmental stress-related abnormality. This is a complex abnormality. This is an ontological substantial anomaly.

[0024] Preferably, the evaluation confidence level JC is calculated in the following way:

[0025]

[0026] in, For the health of the environmental sensor itself, For the freshness of environmental data, To compensate for the model's level of confidence in the current combination of hardware and environmental factors, , 2. These are the corresponding weighting coefficients;

[0027] when At this time, the environmental assessment layer degenerates into a pass-through mode, and the original state index data is directly transmitted to the fusion analysis layer without correction.

[0028] Preferably, the recoverable expected label determines the decision-making strategy based on the type of environmental factor attribute:

[0029] When the environmental factor is altitude or latitude, it is marked as a permanent attribute, and the corresponding decision strategy is to permanently reduce the load balancing weight of the node without triggering a hardware replacement operation.

[0030] When the environmental factor is a thunderstorm or electromagnetic interference, it is marked as a short-term transient attribute, and the corresponding decision strategy is to temporarily enable forward error correction enhancement without triggering hardware alarms.

[0031] When the environmental factor is ambient temperature, it is marked as a long-term transient attribute, and the corresponding decision strategy is to activate temporary cooling enhancement or assess whether the load needs to be migrated.

[0032] Preferably, the environmental factor data is collected through an environmental sensor network, which includes: temperature and humidity sensors deployed on each rack; pressure sensors deployed in each computer room; vibration sensors deployed on each rack; electromagnetic monitoring instruments deployed at the grounding terminals of each computer room; and lightning location receivers deployed outside the data center; the temperature and humidity sensors are deployed with two redundancy options out of three.

[0033] Preferably, the hierarchical differentiation in the decision differentiation step adopts a hierarchical voting mechanism:

[0034] A state fingerprint database is pre-established for each network layer, which contains the characteristic patterns of each network layer under abnormal states.

[0035] When an anomaly is detected, the anomaly features are matched with the fingerprints of each layer, and the time start point and slope steepness of the anomaly are compared.

[0036] The network layer that first experiences a sharp change and lasts the longest is identified as the originating layer, and the remaining layers are marked as transit layers.

[0037] The hierarchical decision-making scheme generates radical cure decisions only for the source layer and only diversion decisions for the transmission layer.

[0038] Preferably, the causal discrimination in the decision discrimination step adopts the perturbation propagation delay difference method:

[0039] Pre-configure the inherent processing latency parameters for each network layer, including physical forwarding latency, virtual switch processing latency, service mesh proxy processing latency, and application thread processing latency;

[0040] When abnormal indicators appear simultaneously in a multi-layer network, calculate the time difference Δt between the occurrence of abnormal peaks in each layer.

[0041] If Δt is equal to or approximately equal to the inherent processing delay of the lower layer, then the fault is determined to be propagating from the lower layer to the upper layer in a positive direction, and the root cause lies in the lower layer.

[0042] If Δt is equal to zero or negative, the fault is determined to be reverse coupling, and the root cause is in the upper layer.

[0043] The hierarchical decision-making scheme generates repair decisions only for the network layer identified as the root cause, and only generates dredging decisions for other layers.

[0044] Among them, when the environmental attribution degree output by the environmental assessment layer At that time, abnormal peaks in the corresponding network layer indicators are not included in the calculation of the disturbance propagation delay difference.

[0045] Preferably, the granularity discrimination in the decision discrimination step divides the decision action into the following levels:

[0046] Level L0: Micro-management level decisions, affecting a single Pod or a single link, including adjusting the Sidecar retry timeout or adjusting the physical port weight;

[0047] Level L0.5: Weight downgrade decision, affecting a single node, used to respond to environmental degradation where the recoverable expected label is a permanent attribute, and to perform load balancing weight downgrade;

[0048] Level L1: Local isolation level decisions, affecting a single node or rack, including removing nodes from the service discovery registry or adjusting the hash factor of the TOR switch;

[0049] Level 2: Regional traffic redirection decisions, affecting a single availability zone, including adjusting DNS weights or GSLB routes to redirect traffic within the availability zone to other availability zones;

[0050] Level L3: Global rate limiting or circuit breaker level decisions, affecting the entire cluster, including triggering global backup mode and enabling the global rate limiter;

[0051] The decision discrimination step determines the decision level based on the entropy value of the abnormal distribution of the network state. The lower the entropy value, the lower the decision level, and the higher the entropy value, the higher the decision level.

[0052] Preferably, the confidence level discrimination in the decision discrimination step adopts the following strategy:

[0053] Overall confidence level Consistency is a measure of cross-layer state consistency, and similarity is a measure of similarity to historical failure modes. The confidence level is defined as the assessment level, and the confirmation level is defined as the result of active confirmation by the real-time probe. , , , These are the corresponding weighting coefficients;

[0054] when At that time, fully automated decision-making is executed and rollback strategies are automatically mounted;

[0055] when When necessary, a canary decision is made, which means that changes are only issued to a preset proportion of instances in the cluster, and after observation and verification, a decision is made to issue changes to all instances or roll back based on the improvement of the status.

[0056] when At that time, the decision is suspended and an operation and maintenance work order is pushed.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] 1. In this invention, by setting up an independent environmental assessment layer, external physical environmental factors such as atmospheric pressure, ambient temperature, relative humidity, electromagnetic field strength, mechanical vibration, and altitude are incorporated into the correction calculation of hardware performance indicators. By using a Bayesian multi-factor compensation model to remove the combined effects of multiple environmental factors, a standard environmental equivalent value is generated. This enables the system to accurately distinguish between environmental anomalies and intrinsic anomalies in network status, solving the problem of hardware performance degradation being misjudged as hardware failure under environments such as high altitude, high temperature, and thunderstorms, and reducing false alarm rates and ineffective maintenance operations.

[0059] 2. In this invention, environmental factors are classified into permanent attributes and transient attributes by recoverable expected tags. Permanent attributes trigger a permanent reduction in L0.5 level load balancing weight, which reduces the load on nodes but keeps them online without triggering hardware replacement. Transient attributes trigger temporary forward error correction enhancement or heat dissipation enhancement, which are automatically deactivated after the environment recovers. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall process flow of the present invention;

[0061] Figure 2 This is a schematic diagram of the data acquisition process of the present invention;

[0062] Figure 3 This is a schematic diagram of the environmental assessment process of the present invention;

[0063] Figure 4 This is a schematic diagram of the fusion analysis process of the present invention;

[0064] Figure 5 This is a schematic diagram of the decision-making process of the present invention;

[0065] Figure 6 This is a schematic diagram of the decision-making and execution process of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Please see Figures 1-6This invention provides a multi-layer network state perception and decision-making method for cloud computing distributed systems. Taking a cloud computing distributed system deployed in a data center at an altitude of 3050 meters as an example, the method first collects data by using layered perception agents deployed at the physical network layer, virtual network layer, service network layer, and application network layer to collect raw state indicator data for each layer of the network.

[0068] The physical network layer sensing agent collects the transmit / receive traffic, packet loss rate, and CRC check error rate of the physical port connected to the node once per second through the Telemetry function of the switch; and collects the received optical power, transmitted optical power, temperature, and bias current of the optical module through the DDMI interface of the optical module once every 10 seconds.

[0069] The virtual network layer awareness agent is mounted to the kernel's vxlan_socket function through the eBPF program to collect the CPU time for encapsulation / decapsulation of the VXLAN tunnel and the number of tunnel MTU fragments, with a collection frequency of once per second;

[0070] The service network layer-aware proxy collects P99 latency, request success rate, and active connection count of each service on the node through the indicator interface of the service mesh control plane, with a collection frequency of once every 5 seconds.

[0071] The application network layer-aware proxy collects RPC call success rate, retry rate, and circuit breaker status between microservices on the node through sampling data from the distributed tracing system, with a sampling frequency of once every 10 seconds.

[0072] Secondly, regarding environmental assessment, the specific method for collecting environmental factor data using the environmental sensor network is as follows: Temperature and humidity sensors are deployed at the center of the front and rear doors of each physical rack, employing a 2-out-of-3 redundancy logic. In this embodiment, the rack temperature sensor readings are 42.1℃, 41.8℃, and 42.3℃, with an average of 42.0℃. Atmospheric pressure sensors are deployed at opposite corners of each server room, with readings of 70.1kPa and 70.3kPa in this embodiment, and an average of 70.2kPa. A vibration sensor is deployed at the bottom of each rack, with a sampling frequency of 100Hz. An electromagnetic monitoring instrument is installed at the grounding terminal of each server room, linked to an external lightning location receiver; in this embodiment, the thunderstorm level is level 2. The altitude of this node is obtained as 3050 meters via a GPS module.

[0073] After receiving the data packet reported by the perception layer, the environmental assessment layer queries the sensitivity coefficient library for the preset sensitivity coefficients of each hardware type of the node to different environmental factors. In this embodiment, the SSD of the data center has a temperature sensitivity coefficient. This means it takes effect at temperatures above 40°C and has a sensitivity coefficient to air pressure. Coupling effect coefficient Calculate the deviation of each environmental factor in this data center from the calibrated environmental value, including the temperature deviation. Air pressure deviation .

[0074] A Bayesian multifactor compensation model was used to perform environmental normalization correction on the raw 4K random read IOPS values ​​of SSD, where: the sum of main effect terms Coupling effect term Total correction factor = 1 + 0.173 + (-0.159) = 1.014; Standard environmental equivalent value .

[0075] The environmental attribution of the SSD's 4K random read IOPS was calculated. , , ,at this time This means that only about 1.4% of the SSD's IOPS decrease was due to environmental factors, while the remaining 98.6% was due to hardware performance degradation. The anomaly was classified as a "substantial anomaly".

[0076] The environmental assessment layer generates recoverable expectation labels based on the attribute types of environmental factors:

[0077] The node's altitude is 3050m > 2000m. The altitude is a "permanent attribute" and is marked as such. The expected recovery is "never recoverable". The corresponding decision strategy is to permanently reduce the load balancing weight of this node without triggering a hardware replacement operation.

[0078] The current thunderstorm level is 2, marked as short-term transient, with an expected recovery time of 45 minutes. The corresponding decision-making strategy is to temporarily activate forward error correction enhancement without triggering hardware alarms.

[0079] The current temperature of 42℃ is considered seasonally high and is expected to last for about 4 hours, which is a long-term transient condition. The corresponding decision-making strategy is to activate temporary cooling enhancement.

[0080] The environmental assessment layer generates assessment confidence scores based on the health status and data freshness of the environmental sensor network. In this embodiment, the health scores are: temperature sensor 0.99, pressure sensor 0.98, humidity sensor 0.97, thunderstorm monitor 0.92, and vibration sensor 0.88. The total health score is set to the minimum of 0.88, and the overall data freshness is 0.92. The compensation model's cognitive certainty is set to 0.95. , , ,but The environmental assessment layer is operating normally, and the corrected data and labels are transmitted to the fusion analysis layer.

[0081] Next, the fusion analysis layer performs fusion analysis. This layer receives corrected data and labels from the environment assessment layer and performs cross-layer mapping and correlation analysis. Specifically: cross-layer topology mapping obtains the physical network topology through the northbound interface of the SDN controller, recording the information of port P3 on the physical switch where the node resides; it obtains VXLAN tunnel information through the virtualization platform API, recording the VXLAN tunnel T carried by port P3; it obtains Pod information through the Kubernetes API, recording the set of Pods connected to tunnel T; and it obtains Service information through the service mesh control plane, recording the Service S to which the Pod belongs. The fusion analysis layer aligns the data from each layer with microsecond-level timestamps within a 100ms time window.

[0082] Decision discrimination before implementation is divided into hierarchy discrimination, causal discrimination, granularity discrimination, and confidence discrimination, among which:

[0083] Layer-based discrimination employs a hierarchical voting mechanism, pre-establishing a state fingerprint database for each network layer: Physical layer fingerprints include sudden increases in bit error rate (>1e-6), sudden increases in CRC check failure rate (>0.1%), and optical power fluctuations (>±2dBm); Virtual layer fingerprints include sudden increases in VXLAN encapsulation CPU utilization (>80%) and sudden increases in MTU fragmentation size (>200% of baseline); Service layer fingerprints include sudden increases in the number of active connections in the connection pool (>80% of maximum capacity) and increases in TLS handshake timeout rate; Application layer fingerprints include normal service TPS, sudden increases in RT (>200% of baseline), and no network packet loss.

[0084] In this embodiment, a "sudden increase in physical layer CRC error rate" and a "surge in service layer P99 latency" were detected. The time start points of the anomalies were compared: the physical layer anomaly started at 14:32:10.123456, and the service layer anomaly started at 14:32:15.200123. According to the fingerprint matching of each layer, the physical layer fingerprint matching was successful, and the service layer fingerprint matching was successful. The physical layer that first experienced a sudden change and lasted the longest was determined to be the source layer, and the service layer was marked as the transmission layer. The decision scheme only generates a radical solution for the physical layer and only generates a mitigation solution for the service layer.

[0085] Causal analysis employs the disturbance propagation delay difference method, pre-configuring inherent processing delay parameters for each network layer: physical forwarding delay 0.1ms, virtual switch processing delay 1ms, service mesh sidecar processing delay 5ms, and application thread processing delay 50ms. The time difference of abnormal peak occurrences at each layer is calculated. The sidecar's inherent processing latency is close to 5ms, indicating the fault is propagating from the physical layer to the upper layers, with the root cause at the physical layer. The abnormal EA value of the physical layer CRC error rate output by the environment assessment layer was checked. This abnormal peak participates in PDD calculation. Decision scheme: only repair decisions are generated for the physical layer, and only diversion decisions are generated for the service layer.

[0086] In granular resolution, anomalies are concentrated in a single physical switch port. The anomaly distribution entropy value is calculated as: the total number of anomalous nodes. , , If the entropy value H≈0, an L0-level decision is triggered: the weight of the physical link is reduced by 50%, and traffic is switched to the backup link. In this embodiment, the anomaly is caused by the aging of the optical module itself, which does not trigger an L0.5-level decision. If the anomaly is caused by the degradation of permanent environmental attributes, then an L0.5-level decision is triggered to reduce the load balancing weight.

[0087] The confidence level calculation in the confidence level differentiation yields the following results: current cross-layer state consistency = 0.88, historical pattern similarity = 0.85, environmental assessment confidence level JC = 0.913, and real-time probe active confirmation level = 0.90. Taking α = 0.3, β = 0.3, γ = 0.2, and δ = 0.2, then:

[0088]

[0089] because The "Canary Decision" mode is triggered: 5% of instances are randomly selected from the cluster as the canary group. A link switching decision is issued to the canary group, and the group is continuously observed for 30 seconds, monitoring changes in error rate and latency. After 30 seconds, the error rate of the canary group decreases from 2.3% to 0.15%, and the latency decreases from 85ms to 52ms, indicating an improvement in status. The decision is then issued to all instances, and a rollback policy is automatically implemented, meaning that automatic rollback occurs when the error rate increases by more than 20%.

[0090] Finally, the decision execution and tracking feedback are implemented. The decision execution layer receives the "L0-level link switching decision", calls the SDN controller's API, reduces the weight of physical switch port P3 from 100 to 50, and switches the traffic to the backup link port P7.

[0091] After the decision is executed, the system starts a feedback tracking timer preset to 10 seconds to continuously monitor the CRC error rate, link packet loss rate, and P99 latency of the corresponding service on the port. After the thunderstorm ends, the system automatically assesses the status of the link, confirms that the CRC error rate has returned to normal, and automatically performs a switchback operation to switch the traffic back to the original link.

[0092] Contents not described in detail in this specification are existing technologies known to those skilled in the art. Standard parts used in this invention can be purchased commercially, and irregularly shaped parts can be custom-made according to the description and drawings. The specific connection methods for each part all employ conventional methods such as bolts, rivets, and welding, which are already mature technologies. The machinery, parts, and equipment all use conventional models from the prior art, and the circuit connections also employ conventional connection methods from the prior art, which will not be detailed here.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-layer network state perception and decision-making method for cloud computing distributed systems, characterized in that, Includes the following steps: Step 1: Collect raw state index data and environmental factor data of each network layer by deploying a layered perception agent at the physical network layer, virtual network layer, service network layer, and application network layer; the environmental factor data includes at least one or more of atmospheric pressure, ambient temperature, relative humidity, electromagnetic field strength, mechanical vibration, and altitude. Step 2: Input the original state index data and the environmental factor data into the environmental assessment layer. The environmental assessment layer performs the following operations: Based on the preset sensitivity coefficients of each hardware type to different environmental factors, calculate the deviation of at least one environmental factor from the calibrated environmental value. Based on the deviation and the preset sensitivity coefficient, the original state index data is normalized and corrected to generate a standard environmental equivalent value. Based on the difference between the original state index data and the standard environmental equivalent value, the environmental attribution degree is calculated; the environmental attribution degree is used to characterize the probability that the anomalies in the original state index data are explained by environmental factors. Based on the attribute type of the environmental factors, generate recoverable expected labels; The attribute types include permanent attributes and transient attributes; Based on the health status and data freshness of the environmental sensor network, a confidence score is generated. Step 3: Input the standard environmental equivalent value corrected by the environmental assessment layer, the environmental attribution degree, the recoverable expected label, and the assessment confidence into the fusion analysis layer to establish cross-layer mapping relationships and correlation analysis between network states of each layer; Step 4: Based on the output of the fusion analysis layer, make decisions on the network state from five dimensions: hierarchy, spatiotemporal, causality, granularity, and confidence, and generate a hierarchical decision scheme; Step 5: Perform corresponding network operation and maintenance operations according to the hierarchical decision-making scheme, and track and provide feedback on the execution results to optimize subsequent decisions.

2. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The environmental normalization correction employs a Bayesian multi-factor compensation model, and the standard environmental equivalent value SE is calculated in the following manner: in, The original state index data, For the first The main effect sensitivity coefficient of each environmental factor For the first The deviation of an environmental factor from its calibrated environmental value. For the first The and the first The coupling effect coefficient between environmental factors.

3. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The environmental attribution score (EA) is calculated in the following manner: when hour, ; when hour, ; According to the above The values ​​categorize exception types as follows: This is an environmental stress-related abnormality. This is a complex abnormality. This is a substantial anomaly.

4. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The confidence level JC is calculated in the following way: in, For the health of the environmental sensor itself, For the freshness of environmental data, To compensate for the model's level of confidence in the current combination of hardware and environmental factors, , 2. These are the corresponding weighting coefficients; when At this time, the environmental assessment layer degenerates into a pass-through mode, and the original state index data is directly transmitted to the fusion analysis layer without correction.

5. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The recoverable expected label determines the decision-making strategy based on the type of environmental factor attributes: When the environmental factor is altitude or latitude, it is marked as a permanent attribute, and the corresponding decision strategy is to permanently reduce the load balancing weight of the node without triggering a hardware replacement operation. When the environmental factor is a thunderstorm or electromagnetic interference, it is marked as a short-term transient attribute, and the corresponding decision strategy is to temporarily enable forward error correction enhancement without triggering hardware alarms. When the environmental factor is ambient temperature, it is marked as a long-term transient attribute, and the corresponding decision strategy is to activate temporary cooling enhancement or assess whether the load needs to be migrated.

6. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The environmental factor data is collected through an environmental sensor network, which includes: Temperature and humidity sensors deployed in each rack; Pressure sensors deployed in various computer rooms; Vibration sensors deployed on each rack; Electromagnetic monitoring instruments deployed at the grounding terminals of each computer room; Lightning location receivers deployed outside the data center; The temperature and humidity sensors are deployed with two out of three redundancy.

7. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The hierarchical differentiation in the decision-making process employs a tiered voting mechanism: A state fingerprint database is pre-established for each network layer, which contains the characteristic patterns of each network layer under abnormal states. When an anomaly is detected, the anomaly features are matched with the fingerprints of each layer, and the time start point and slope steepness of the anomaly are compared. The network layer that first experiences a sharp change and lasts the longest is identified as the originating layer, and the remaining layers are marked as transit layers. The hierarchical decision-making scheme generates radical cure decisions only for the source layer and only diversion decisions for the transmission layer.

8. The method for multi-layer network state perception and decision-making in a cloud computing distributed system according to claim 1, characterized in that, The causal discrimination in the decision discrimination step adopts the perturbation propagation delay difference method: Pre-configure the inherent processing latency parameters for each network layer, including physical forwarding latency, virtual switch processing latency, service mesh proxy processing latency, and application thread processing latency; When abnormal indicators appear simultaneously in a multi-layer network, calculate the time difference Δt between the occurrence of abnormal peaks in each layer. If Δt is equal to or approximately equal to the inherent processing delay of the lower layer, then the fault is determined to be propagating from the lower layer to the upper layer in a positive direction, and the root cause lies in the lower layer. If Δt is equal to zero or negative, the fault is determined to be reverse coupling, and the root cause is in the upper layer. The hierarchical decision-making scheme generates repair decisions only for the network layer identified as the root cause, and only generates dredging decisions for other layers. Among them, when the environmental attribution degree output by the environmental assessment layer At that time, abnormal peaks in the corresponding network layer indicators are not included in the calculation of the disturbance propagation delay difference.

9. A multi-layer network state perception and decision-making method for cloud computing distributed systems according to claim 1, characterized in that, The granularity discrimination step in the decision discrimination process divides decision actions into the following levels: Level L0: Micro-management level decisions, affecting a single Pod or a single link, including adjusting the Sidecar retry timeout or adjusting the physical port weight; Level L0.5: Weight downgrade decision, affecting a single node, used to respond to environmental degradation where the recoverable expected label is a permanent attribute, and to perform load balancing weight downgrade; Level L1: Local isolation level decisions, affecting a single node or rack, including removing nodes from the service discovery registry or adjusting the hash factor of the TOR switch; Level 2: Regional traffic redirection decisions, affecting a single availability zone, including adjusting DNS weights or GSLB routes to redirect traffic within the availability zone to other availability zones; Level L3: Global rate limiting or circuit breaker level decisions, affecting the entire cluster, including triggering global backup mode and enabling the global rate limiter; The decision discrimination step determines the decision level based on the entropy value of the abnormal distribution of the network state. The lower the entropy value, the lower the decision level, and the higher the entropy value, the higher the decision level.

10. A multi-layer network state perception and decision-making method for cloud computing distributed systems according to claim 1, characterized in that, The confidence level discrimination in the decision discrimination step adopts the following strategy: Overall confidence level Consistency is a measure of cross-layer state consistency, and similarity is a measure of similarity to historical failure modes. The confidence level is defined as the assessment level, and the confirmation level is defined as the result of active confirmation by the real-time probe. , , , These are the corresponding weighting coefficients; when At that time, fully automated decision-making is executed and rollback strategies are automatically mounted; when When necessary, a canary decision is made, which means that changes are only issued to a preset proportion of instances in the cluster, and after observation and verification, a decision is made to issue changes to all instances or roll back based on the improvement of the status. when At that time, the decision is suspended and an operation and maintenance work order is pushed.

Images