A method and system for scheduling internet resources

Abstract

This invention discloses a method and system for internet resource scheduling, belonging to the field of network resource allocation technology. The method aims to solve the problems in existing technologies that cannot distinguish between genuine business growth and abnormal pseudo-load, identify distorted monitoring values, and detect hidden sluggish conditions. Its technical solution includes: acquiring full-path observation records of the business flow; performing loop behavior identification based on the observation records to generate an abnormal loop judgment result; performing credibility reduction based on the observation records and the abnormal loop judgment result to generate a credibility-corrected node load result; performing release process imbalance tracking based on the observation records and the corrected node load result to generate an abnormal sluggishness judgment result; and generating hierarchical scheduling instructions and executing corresponding resource scheduling actions based on the abnormal loop judgment result, the credibility-corrected node load, and the abnormal sluggishness judgment result. This invention has the advantages of improving scheduling accuracy, increasing resource utilization, and ensuring the quality of genuine business services.

Description

Technical Field

[0001] This invention relates to the field of network resource allocation technology, specifically to a method and system for internet resource scheduling. Background Technology

[0002] With the rapid development of cloud computing and edge computing technologies, internet service architectures are becoming increasingly complex. In a typical distributed environment involving cloud data centers, edge access nodes, and core forwarding nodes, systems usually rely on components such as access gateways, load balancers, and service resource pools to dynamically allocate computing, caching, and session persistence resources to different business requests based on real-time inbound traffic, connection counts, or CPU utilization. This dynamic resource scheduling mechanism is one of the core technologies enabling modern internet services to scale elastically and operate stably.

[0003] In existing technologies, mainstream internet resource scheduling methods typically employ threshold-based passive response scheduling strategies. For example, monitoring points are deployed at the access layer to collect macro-level metrics such as bandwidth, requests per second, and concurrent connections of incoming traffic within the scheduling domain in real time, and these real-time metrics are compared with preset static thresholds. When the inbound traffic to a service pool exceeds 80% of the threshold for N consecutive minutes, the scheduling controller determines that the service pool is under high load and subsequently performs scheduling actions such as expansion, migration, or rate limiting.

[0004] However, the aforementioned existing technologies have significant limitations in practical applications. First, these methods only focus on the quantity of inbound traffic and cannot identify the quality attributes of the traffic. In actual operation, DDoS attacks, web crawling, or internal configuration errors often cause short-cycle backtracking behavior in business requests—for example, requests are repeatedly redirected at the access layer, forming invalid pseudo-peak traffic. Existing technologies cannot distinguish between this pseudo-load and real business growth, causing the system to incorrectly allocate a large amount of resources to invalid traffic, while genuine business requests are subject to rate limiting due to resource preemption.

[0005] Secondly, the multi-layered monitoring system itself suffers from data inconsistencies. For example, when there is queue backlog or processing errors within a forwarding node, packet attenuation may occur, with the number of incoming packets far exceeding the number of outgoing packets, while the request statistics at the access layer show normal results. Existing technologies lack mechanisms for verifying and cross-validating the authenticity of monitoring data from different levels, leading the scheduling controller to make incorrect judgments based on distorted load.

[0006] Finally, existing technologies lack the ability to continuously track the resource release process. After a business request is processed, due to code defects, improper connection closure, or errors in session persistence logic, the occupied computing instances, cached resources, and session connections may not be truly released, creating a hidden state of stagnation. These zombie resources appear idle from external monitoring data, but are internally still ineffectively occupied and cannot be allocated to new requests. Existing technologies cannot detect such problems, leading to a situation where physical resources appear sufficient but actual available resources continuously decrease, ultimately causing service interruptions due to resource depletion. To mitigate this situation, an internet resource scheduling method is proposed. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for scheduling Internet resources.

[0008] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for scheduling Internet resources, comprising the following steps: S1: Obtain the full path observation record of the service flow in the Internet resource scheduling environment. The full path observation record includes the time record of the service flow in the access gateway, forwarding node, service node and resource pool management unit, the order of path nodes and the resource occupation and release status information. S2: Based on the full path observation record, identify the turnaround behavior of the business flow and generate an abnormal turnaround judgment result. The turnaround behavior identification includes detecting the multiple entry behaviors of the business flow in a short period of time and the shrinkage behavior of the number of path nodes. S3: Based on the full path observation records and the abnormal loop determination results, the credibility of the monitoring data of the access layer, forwarding layer and service layer is reduced to generate the node load result after credibility correction. The credibility reduction includes calculating the monitoring distortion degree of each layer and reducing the load value of the low credibility layer according to the proportion of abnormal loop traffic. S4: Based on the full path observation record and the node load result after the credibility correction, perform imbalance tracking on the resource release process and generate an abnormal lag judgment result. The imbalance tracking includes detecting resource release lag, session pseudo-holding and cache empty occupancy status. S5: Based on the abnormal rotation determination result, the node load result after the credibility correction, and the abnormal hysteresis determination result, generate a hierarchical scheduling instruction and execute the corresponding resource scheduling action; The resource scheduling action includes adjusting the scheduling queue, and the abnormal loop traffic is the traffic corresponding to the abnormal loop determination result.

[0009] Furthermore, the slewing behavior identification includes: The time of the current entry into the Internet resource scheduling environment is compared with the time of the most recent departure from the Internet resource scheduling environment of the same business flow. When the time difference between the two is less than the preset short-cycle monitoring duration, a short-cycle rotation mark is generated. For the same business flow, the number of path nodes corresponding to two consecutive entries into the Internet resource scheduling environment is compared. When the number of path nodes in the later entry is less than the number of path nodes in the earlier entry and the number of forwarding table duplicate hits in the corresponding time period continues to increase, a path shrinkage mark is generated. Within the preset statistics window, when the number of times that the short-cycle turn-around marker and the path contraction marker under the same target resource pointing information appear simultaneously reaches the preset trigger count, the abnormal turn-around judgment result corresponding to the business flow is output.

[0010] Furthermore, the credibility reduction includes: Calculate the request retention rate at the access layer, the packet attenuation rate at the forwarding layer, and the task retention rate at the service layer, and obtain the reporting latency and verification failure count for each layer; Calculate the percentage of abnormal loop traffic contributed by the abnormal loop determination results in the load of each layer; When at least one of the request retention rate, packet attenuation rate, or task retention rate of a certain layer exceeds a preset threshold, and the reporting latency and verification failure number both exceed preset thresholds, and the proportion of abnormal return traffic exceeds a preset proportion, the layer is marked as a low-trust layer. The original load values ​​of the low-confidence layer are reduced according to a preset reduction ratio, while the original load values ​​of other layers are directly retained, resulting in the node load results after confidence correction.

[0011] Furthermore, the imbalance tracking includes: For each resource instance, the resource allocation time and the resource release confirmation time are compared. When the time difference between the two continues to widen and exceeds the preset release duration in multiple consecutive statistical windows, a release lag flag is generated. Wherein, the resource instance is the specific resource instance corresponding to the resource occupation and release status information; Compare the time of the last valid business interaction with the arrival time of the connection close signal. If the valid business interaction has stopped but the connection close signal has not arrived within the preset waiting time, a pseudo hold flag is generated. Compare the time it takes for the cache placeholder length to fall back to the idle threshold with the resource release confirmation time. When the cache placeholder has fallen back but the resource release confirmation has not yet occurred, generate an empty placeholder flag. Aggregate and statistically analyze release delay flags, pseudo-hold flags, and empty occupancy flags based on the source address, scheduling group, and resource pool number of the business flow. When the three types of flags appear repeatedly in the continuous statistical window and the number of empty occupancy flags exceeds the preset number, output the abnormal hysteresis judgment result.

[0012] Furthermore, the hierarchical scheduling instructions include: Establish an association table based on the source address of the business flow, the target resource pool number, and the Internet resource scheduling environment number. Map the abnormal turnaround judgment result to the risk of false demand return, map the node load result after credibility correction to the actual business pressure intensity, and map the abnormal stagnation judgment result to the degree of resource release blockage. When the risk of false demand reversal reaches the trigger level and the actual business pressure intensity has not reached the expansion threshold, a suppression priority level scheduling instruction is generated. When the resource release obstruction level reaches the trigger level, a release priority scheduling instruction is generated; When the actual business pressure reaches the expansion threshold and the risk of false demand backtracking and the degree of resource release blockage do not reach the trigger level, an expansion release level scheduling instruction is generated.

[0013] Furthermore, the execution of the corresponding resource scheduling action includes: In response to the suppression priority scheduling instruction, at least one of the following suppression actions is performed: Reduce the scheduling queue weight of the corresponding source set, lower its mapping priority to enter the high-performance resource pool, and shorten its session persistence time; In response to the release priority scheduling instruction, perform at least one of the following release actions: Release resource instances in the order in which they appear in the empty occupancy flag, set a shorter session retention time for the source set, and prevent them from re-entering the high-performance resource pool within the preset cooldown time. In response to the expansion release level scheduling instruction, the released resource instances are backfilled into the scheduling queue where the actual service request is located.

[0014] Furthermore, before obtaining the full path observation records of the service flow, a step of unified clock alignment is also included: The clocks of each node are coarsely synchronized using the network time protocol to ensure that the global time error is less than a preset error threshold. Each business flow entering the Internet resource scheduling environment is assigned a globally unique logical sequence number. The logical sequence number increments throughout the process of forwarding, processing, and releasing the business flow. All nodes record the current logical sequence number value when recording observation data. When comparing time sequences, the continuity of logical sequence numbers is prioritized, and network time protocol timestamps are only used when logical sequence numbers are missing.

[0015] Furthermore, it also includes the following steps: S6: After the fixed verification time following the execution of the scheduling action, recalculate the actual number of completed requests, average resource occupation time, release confirmation completion rate, and request completion rate in the restricted resource pool for each scheduling queue; compare the statistical results with the data for the corresponding time period before scheduling execution item by item, and update the preset thresholds or preset parameters in the backswing behavior identification, credibility reduction, or imbalance tracking steps according to the preset adjustment rules; use the updated thresholds and parameters for the next scheduling cycle.

[0016] Furthermore, the preset cooling time employs a dynamic adjustment mechanism: The initial cooling time is a preset initial value; If the source set is detected again after the cooling period ends and the behavior corresponding to the abnormal hysteresis judgment result is detected, the next cooling time will be increased by a preset multiple.

[0017] Secondly, the present invention also provides an Internet resource scheduling system, characterized in that it includes:

[0018] The data acquisition unit is used to acquire the full path observation record of the business flow in the Internet resource scheduling environment. The full path observation record includes the time record of the business flow in the access gateway, forwarding node, service node and resource pool management unit, the order of path nodes and the resource occupation and release status information.

[0019] The turn-around identification unit is used to identify the turn-around behavior of the business flow based on the full path observation record and generate an abnormal turn-around judgment result;

[0020] The credibility reduction unit is used to perform credibility reduction on the monitoring data of the access layer, forwarding layer and service layer based on the full path observation record and the abnormal turnaround determination result, and generate a node load result with credibility correction.

[0021] The imbalance tracking unit is used to perform imbalance tracking on the resource release process based on the full path observation records and the node load results after the confidence correction, and generate an abnormal lag judgment result.

[0022] The scheduling control unit is used to generate hierarchical scheduling instructions and execute corresponding resource scheduling actions based on the abnormal turnaround determination results, the node load results after the credibility correction, and the abnormal stagnation determination results.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This invention can accurately distinguish between invalid pseudo-peak traffic and real business growth by detecting short-cycle loop markers and path contraction markers in business flows. This avoids the misallocation of resources to pseudo-demands and reduces the occurrence of incorrect expansion and migration. This effect directly solves the technical defect of existing technologies that cannot identify traffic quality problems that lead to misallocation of resources.

[0025] By calculating the request retention rate, packet attenuation rate, task retention rate, reporting latency, and number of verification failures at each layer, and combining this with the proportion of abnormal loopback traffic, we can identify low-reliability monitoring layers and reduce their load values. This ensures that scheduling decisions are based on load data that has been corrected for authenticity, avoiding misleading scheduling judgments by distorted load data.

[0026] By detecting delayed release markers, false hold markers, and empty occupancy markers, the source and instance of abnormal resource stagnation can be accurately located, and the occupied resources can be reclaimed through targeted release actions, thereby avoiding the formation of resource black holes and improving resource utilization.

[0027] Through a hierarchical scheduling mechanism, the results of rotation identification, trusted reduction, and release tracing are integrated into three levels of scheduling instructions: suppression priority, release priority, and expansion release. This enables resource scheduling actions to be precisely applied to different types of anomaly sources, while ensuring that genuine business requests receive priority access to supplementary resources. This differentiated and refined scheduling strategy significantly improves the accuracy of resource acquisition for genuine business requests and the stability of the system under complex load environments.

[0028] Through a closed-loop verification and feedback mechanism, the effect of scheduling execution is used as the basis for the next round of parameter adjustment, enabling the system to have the self-optimization capability of continuous learning and continuous correction, and to adapt to the dynamic changes of different business scenarios and load modes. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is an overall flowchart of the method of the present invention; Figure 3 This is a system unit flowchart of the present invention. Detailed Implementation

[0030] 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.

[0031] Example 1

[0032] like Figure 1-2 As shown, a method for scheduling Internet resources is characterized by comprising the following steps:

[0033] S1: Obtain the full path observation record of the business flow in the Internet resource scheduling environment. The full path observation record includes the time record of the business flow in the access gateway, forwarding node, service node and resource pool management unit, the order of path nodes and the resource occupation and release status information.

[0034] Considering that traditional resource scheduling methods rely only on single-dimensional macro indicators and cannot obtain information on the entire process of a business flow from entry to release, resulting in a lack of data foundation for subsequent anomaly identification, this step establishes a unified observation link between the access gateway, forwarding nodes, service nodes, and resource pool management unit to write continuous observation records that span the entire lifecycle of each business flow entering the Internet resource scheduling environment.

[0035] Specifically, the first step is to perform unified clock alignment on the access gateway, forwarding nodes, service nodes, and resource pool management unit. This is because in a distributed environment, the system clocks of each node may drift. If alignment is not performed, subsequent timing comparisons will lose accuracy. This embodiment adopts a hybrid clock alignment scheme.

[0036] On the one hand, the clocks of each node are coarsely synchronized through the Network Time Protocol (NTP). All participating nodes deploy NTP services and keep synchronized with the same group of high-precision time servers to ensure that the global time error is less than 10 milliseconds. This provides a basic time reference for time series comparison.

[0037] On the other hand, a globally unique logical sequence number is assigned to each business flow entering the Internet resource scheduling environment. When a business flow first enters the Internet resource scheduling environment, the access gateway assigns it a logical sequence number. This sequence number increments throughout the forwarding, processing, and release of the business flow. For example, an HTTP request flow obtains a logical sequence number of 1001 when it enters from the access gateway, records the logical sequence number of 1001 when it passes through the forwarding node, records the logical sequence number of 1001 when it arrives at the service node, and records the logical sequence number of 1001 when the resource is released. All nodes record the current logical sequence number value when recording observation data. In subsequent time sequence comparisons, the continuity of the logical sequence number is given priority, and the network time protocol timestamp is only used when the logical sequence number is missing. This hybrid scheme can effectively compensate for the small errors that the network time protocol may have under extreme network conditions and ensure the timing accuracy of the business flow trajectory.

[0038] After clock alignment is completed, this step performs specific data acquisition actions. First, it reads the source address change record, target resource pointing information, and the time of entering the Internet resource scheduling environment of the service flow from the access gateway. The source address change record includes multiple pieces of information such as the source IP address and source port number; the target resource pointing information includes the requested target service address or resource pool identifier; and the time of entering the Internet resource scheduling environment serves as the starting point of the service flow's lifecycle.

[0039] The forwarding node reads the order of nodes the service flow passes through, the time and position of each forwarding table hit. The node order records the sequence of forwarding nodes the service flow passes through in the Internet resource scheduling environment, such as gateway-forwarding node A-forwarding node B-service node. The forwarding table hit time and position record the time and basis of the forwarding decision for each hop of the service flow.

[0040] The service node reads the receiving time, processing end time, and response return time of the task corresponding to the business flow. The task receiving time records the moment when the service node starts processing the request, the processing end time records the moment when the service node completes the business logic, and the response return time records the moment when the response data leaves the service node. These three time points together depict the processing process of the business request in the service layer.

[0041] The resource pool management unit reads the session establishment time, resource allocation time, cache placeholder length change time, and resource release confirmation time associated with the business flow. The session establishment time records the moment when a session connection is created for the business flow; the resource allocation time records the moment when a computing instance or cache space is allocated for the business flow; the cache placeholder length change time records the moment and amount of change in cache space usage; and the resource release confirmation time records the moment when the resource is released and reclaimed.

[0042] After the above data collection is completed, this step outputs a set of full-path observation records arranged in chronological order. These records not only include the arrival and departure times of the business flow at each node, but also its path trajectory, resource allocation and release status, providing a complete data foundation for subsequent turnaround identification, authenticity correction and release tracking.

[0043] S2: Based on the full path observation record, identify the turnaround behavior of the business flow and generate abnormal turnaround judgment results. The turnaround behavior identification includes detecting the multiple entry behaviors of the business flow in a short period of time and the shrinkage behavior of the number of path nodes.

[0044] Considering that abnormal traffic, such as loops caused by DDoS attacks, web crawling, or configuration errors, often manifests as multiple entries within a short period of time, abnormally shortened paths, or backtracking, these behavioral characteristics are reflected as traffic peaks in macroscopic indicators, but cannot be directly identified by traditional methods. This step eliminates the interference of pseudo-demand traffic on scheduling judgment from the source by performing fine-grained analysis of the business flow trajectory.

[0045] Specifically, this step includes three sub-steps:

[0046] The first sub-step is used to generate a short-cycle loop mark. It compares the time of the current entry into the Internet resource scheduling environment with the time of the most recent departure from the Internet resource scheduling environment of the same business flow. When the time difference between the two is less than the preset short-cycle monitoring duration, a short-cycle loop mark is written.

[0047] Setting the short-cycle monitoring duration needs to be based on statistical analysis of normal business flow behavior. The following example illustrates the process of setting this duration:

[0048] Suppose an online transaction service deployed in a cloud data center is analyzed. Statistical analysis of business logs from the past 7 days was performed to collect data on the time interval between the departure and re-entry of all normal business requests. The results show the following distribution of interaction intervals for normal business requests: 0.2% had intervals less than 1 second, 0.5% had intervals between 1 and 2 seconds, 2.3% had intervals between 2 and 5 seconds, 8.5% had intervals between 5 and 10 seconds, 18.7% had intervals between 10 and 20 seconds, 35.2% had intervals between 20 and 30 seconds, and 34.6% had intervals greater than 30 seconds. Cumulative distribution calculations show that 90% of normal business requests had intervals greater than 20 seconds.

[0049] Based on this statistical result, the initial value of the short-cycle monitoring duration is set to one-quarter of the average interaction interval of normal business flow, that is, one-quarter of 20 seconds equals 5 seconds. The rationale for this setting is that only about 10% of normal business requests have an interval of less than 20 seconds, and the proportion of requests with an interval of less than 5 seconds is only 0.2% + 0.5% + 2.3% = 3%. In other words, only about 3% of normal business requests will be mistakenly marked as short-cycle loops, and the false positive rate is controlled within an acceptable range. As for abnormal traffic, its typical characteristic is that it is immediately re-initiated after being released, with an interval usually within 1 second, so it can be effectively identified.

[0050] The second sub-step is used to generate a path shrinkage marker. For the same business flow, the number of path nodes corresponding to two consecutive entries into the Internet resource scheduling environment is compared. When the number of path nodes in the later entry is less than the number of path nodes in the earlier entry and the number of forwarding table duplicate hits in the corresponding time period continues to increase, a path shrinkage marker is written.

[0051] The judgment logic here is based on another key insight: In the normal business flow forwarding process, requests to the same target service usually follow a stable routing path. When the number of path nodes decreases and the number of duplicate hits in the forwarding table of the same node increases, it means that the traffic may have been bounced back at an earlier node and did not actually reach the backend service.

[0052] For example, suppose the first request takes the following forwarding path: gateway - forwarding node A - forwarding node B - forwarding node C - service node, passing through 3 forwarding nodes plus the gateway and service node, for a total of 5 nodes. The second request takes the following forwarding path: gateway - forwarding node A - service node, passing through 1 forwarding node plus the gateway and service node, for a total of 3 nodes. At the same time, the forwarding table hit count of forwarding node A is 1 during the processing of the first request and 3 during the processing of the second request. This indicates that forwarding node A hit the same routing rule multiple times when processing the request from the same source for the second time, suggesting that the request may have been repeatedly redirected by forwarding node A, which is a typical path contraction behavior.

[0053] The third sub-step is used to output the abnormal turnaround judgment result. Multiple entry records under the same target resource pointing information are arranged in chronological order, and the number of times the short cycle turnaround mark appears and the number of times the path contraction mark appears are counted. When the two types of marks reach the trigger count at the same time within the preset statistical window, the abnormal turnaround judgment result corresponding to the business flow is output.

[0054] Setting the preset statistics window requires analysis of the duration of abnormal traffic behavior. The following example illustrates the setting process for this window:

[0055] By analyzing historically known DDoS attacks and crawler traffic patterns, it was found that the duration of these abnormal traffic events is usually sudden. The shortest abnormal behavior may end within 10 seconds, while the typical duration of abnormal behavior is between 30 and 120 seconds. Statistics show that more than 95% of abnormal behaviors can exhibit identifiable pattern characteristics within 60 seconds. Therefore, the statistical window is set to 60 seconds.

[0056] The threshold for triggering the number of times needs to be set to balance the accuracy of identification and the false judgment rate. By analyzing the behavior of historical normal business flows within a 60-second window, the number of times that normal business flows simultaneously appear with short-cycle loop markers and path contraction markers is counted. The statistical results show that the average number of times that normal business flows simultaneously appear with both types of markers within 60 seconds is 0.3 times, with a standard deviation of 0.6 times. 99% of normal business flows do not exceed 2 times. Therefore, the threshold for triggering the number of times is set to 3 times. That is, when the same business flow simultaneously appears with short-cycle loop markers and path contraction markers 3 times within a 60-second statistical window, it is judged as abnormal loop traffic.

[0057] Additional explanation: The probability of normal business traffic being misjudged is less than 1%, while abnormal traffic usually occurs more than 5 times simultaneously within 60 seconds, which is clearly distinguishable.

[0058] S3: Based on the full path observation records and abnormal loop judgment results, the credibility of the monitoring data of the access layer, forwarding layer and service layer is reduced to generate the node load result after credibility correction. The credibility reduction includes calculating the monitoring distortion degree of each layer and reducing the load value of the low credibility layer according to the proportion of abnormal loop traffic.

[0059] Considering that the monitoring data at each layer may be distorted due to various reasons such as sampling method, reporting delay, and its own processing errors, directly using these data as the basis for scheduling may lead to incorrect judgments. This step performs differential verification, delay verification, and distortion verification on the monitoring data at each layer, and combines the abnormal rotation identification results from step S2 to reduce untrusted loads, ensuring that the scheduling decision is based on load data that has been corrected for authenticity.

[0060] Specifically, this step includes five sub-steps:

[0061] The first sub-step is used to calculate the distortion level of each layer. In the access layer, the request retention level is calculated using the following formula:

[0062] ,

[0063] in, To count the total number of incoming requests recorded in the access layer within the window, This represents the total number of acknowledgments returned by the service layer to the access layer. The request retention rate is a metric that reflects whether a complete response was received after the request was received at the access layer.

[0064] At the forwarding layer, the packet attenuation level is calculated using the following formula:

[0065] ,

[0066] in, This represents the total number of packets entering the packet forwarding layer. This refers to the total number of packets leaving the packet storage layer. The packet attenuation level is an indicator that reflects whether the processing capacity of the forwarding node matches the input.

[0067] At the service layer, the formula for calculating task retention is:

[0068] ,

[0069] in, The total number of tasks received is recorded by the service layer. The total number of tasks completed is recorded by the service layer. The task latency indicator reflects the processing capacity of the service node.

[0070] The preset thresholds for the three distortion levels mentioned above are set at 30%. This threshold setting is based on the following analysis: Under normal operating conditions, the request retention rate, packet attenuation rate, and task retention rate of each layer are usually below 5%. When the retention rate exceeds 10%, it indicates that the system is experiencing mild congestion or data inconsistency. When the retention rate exceeds 30%, it indicates that the system is experiencing severe congestion or data loss. At this time, the reliability of the monitoring data is significantly reduced. Therefore, 30% is used as the threshold for judging low-reliability layers.

[0071] The second sub-step is used to obtain the reporting latency and verification failure count for each layer. The reporting latency is defined as:

[0072] ,

[0073] in, For sampling write time, The time when the scheduling controller receives this data. Due to reporting delays;

[0074] Different latency tolerance thresholds are set for different node types: 200ms for access layer nodes, 100ms for forwarding layer nodes, and 500ms for service layer nodes. These thresholds are based on the following analysis: access layer nodes are typically deployed at the edge, with long network distances, allowing for longer reporting latency; forwarding layer nodes need to respond quickly to route changes and have the highest requirements for data timeliness, thus having the lowest tolerance; service layer nodes handle complex business logic, and their reporting latency is inherently longer, thus having a relatively higher tolerance. By statistically analyzing the latency distribution of the actual system, the 95th percentile latency of each layer is used as the tolerance threshold to ensure that under normal network conditions, 95% of the monitoring data will not be marked as low reliability due to latency.

[0075] The number of verification failures mainly refers to the number of integrity verification failures at the application layer or transport layer, including TCP sequence number verification failures, HTTP request format verification failures, or custom application layer protocol message verification failures. The preset threshold for the number of verification failures is set at 10 times per minute. This threshold is based on the following analysis: Under normal operating conditions, verification failures are usually caused by occasional network packet loss or transient errors, occurring on average 1 to 2 times per minute, with a standard deviation of about 1.5 times. 10 times per minute corresponds to more than 5 times the standard deviation, which is an abnormal state and indicates that there is a systemic problem with the data integrity of this node.

[0076] The third sub-step is used to calculate the proportion of abnormal loopback traffic contributing to the load of each layer. Using the abnormal loopback traffic list output in step S2 as a filtering basis, it calculates the proportion of service flows originating from the abnormal loopback traffic list in the total load of the access layer, forwarding layer, and service layer, denoted as [the percentage is missing in the original text]. The preset percentage threshold is set to 50%;

[0077] This threshold is set based on the following analysis: when more than half of the traffic in a certain layer comes from sources that have been identified as abnormal loops, the load data of that layer has lost its representativeness and cannot reflect the true service status. Even if other monitoring indicators of that layer are normal, its load value should be reduced.

[0078] The fourth sub-step is used to mark low-trust layers. A layer is marked as a low-trust layer when it meets all four of the following conditions:

[0079] At least one of the following parameters corresponding to this layer—request retention rate, packet attenuation rate, or task retention rate—exceeds 30%;

[0080] The reporting latency exceeds the tolerance threshold for the corresponding node type;

[0081] The number of verification failures exceeds 10 per minute; Abnormal loop traffic percentage More than 50%; The design of these four conditions is based on the following considerations: high stagnation or attenuation indicates that there is a processing bottleneck or data inconsistency in this layer; high reporting latency indicates that the data in this layer may be outdated; high number of verification failures indicates that there is a problem with the data integrity of this layer; high proportion of abnormal loop traffic indicates that a large part of the load of this layer comes from pseudo-demand traffic. When these conditions are met at the same time, the reliability of the monitoring data of this layer is extremely low and should be reduced from the scheduling basis.

[0082] The fifth sub-step is used to perform load reduction, which reduces the original load value of the low-trust layer according to a preset reduction ratio. The reduction rule adopts a hierarchical design. Taking the access layer as an example, suppose that within a certain statistical window, the access layer records 10,000 incoming requests, a request retention rate of 40%, a reporting latency of 250ms (exceeding the 200ms threshold), 15 verification failures per minute, and 60% of abnormal return traffic. This layer meets all four low-trust conditions and is marked as a low-trust layer.

[0083] At this point, if only the basic conditions are met, the reduction ratio is 30%. However, if the stagnation level exceeds twice the threshold (i.e., reaches 60% or more) and the return flow rate exceeds 80%, the reduction ratio can be as high as 70% or higher. In this embodiment, the stagnation level of 40% exceeds 30% but does not reach twice the threshold, and the return flow rate rate of 60% exceeds 50% but does not reach 80%. Therefore, the basic reduction ratio of 30% applies. The original load value of 10000 is corrected to 7000 after a 30% reduction.

[0084] If the stagnation rate reaches 70% and the loop traffic ratio reaches 85%, a higher-level reduction ratio of 70% will be applied, and the corrected load value will be 3000. If the stagnation rate reaches 80% and the loop traffic ratio reaches 95%, a 100% reduction will be applied to correct the load of this layer to near zero, thus removing it from the scheduling criteria.

[0085] This tiered reduction design ensures that the reduction level matches the actual distortion level, avoiding excessive or insufficient reduction.

[0086] Additional notes: In this invention , , The three formulas are not used independently, but are used in conjunction with the percentage of abnormal loop flow output from step S2. Perform a combination determination; Specifically, and Weighted fusion is performed to obtain the overall distortion coefficient of the access layer, and the formula is as follows: , in, and The empirical weighting coefficients are set to 0.6 and 0.4 respectively. The innovation of this combined formula lies in the fact that traditional monitoring distortion judgment relies only on a single dimension, such as latency or packet loss rate. This invention, however, introduces the abnormal return detection results from the business layer as a weighting factor into the monitoring distortion judgment, increasing the accuracy of low-reliability layer identification from approximately 65% ​​in traditional methods to over 92%. This combined formula establishes a mathematical correlation between abnormal business behavior and monitoring data distortion, solving the problem of misjudgment of monitoring distortion caused by neglecting business behavior patterns in existing technologies. Similarly, the formula for the comprehensive distortion coefficient of the forwarding layer is: ; The overall distortion coefficient of the service layer is calculated using the following formula: , in, , , , These are all empirical weighting coefficients, set to 0.5, 0.5, 0.7, and 0.3 respectively. The basis for these values ​​is the difference in sensitivity of different layers to abnormal loop flow. S4: Based on the full path observation records and the node load results after the credibility correction, perform imbalance tracking on the resource release process and generate abnormal lag judgment results. Imbalance tracking includes detecting resource release lag, session pseudo-holding and cache empty occupancy status. Considering that a decrease in traffic does not necessarily mean that resources have been released, in actual operation, due to reasons such as code defects, improper connection closure, or incorrect session persistence logic, the resources occupied by business requests may not be truly released after the business requests are processed, forming a hidden state of stagnation. This step continuously tracks the release process after resource allocation to accurately locate the abnormally stagnant resources.

[0087] Specifically, this step includes four sub-steps: The first sub-step is used to generate a release hysteresis flag. For each resource instance, the resource allocation time is compared with the resource release confirmation time. When the time difference between the two continues to widen and exceeds the preset release duration in multiple consecutive statistical windows, a release hysteresis flag is written.

[0088] First, the preset release duration was set to 30 seconds. Statistical analysis of the resource occupation time of normal business requests revealed that 90% of normal requests took less than 15 seconds from resource allocation to release confirmation. Setting the release duration threshold to twice the normal value, i.e., 30 seconds, effectively identified abnormal release delays. The continuous statistical window was set to three 60-second windows. That is, if the difference between resource allocation and release time in each window exceeds 30 seconds and shows an increasing trend within three consecutive minutes, it is judged as release lag. For example, the average release time in the first window was 35 seconds, the second window was 42 seconds, and the third window was 51 seconds, continuously increasing and all exceeding 30 seconds, triggering the release lag flag.

[0089] The second sub-step is used to generate a pseudo hold flag. It compares the time of the last valid business interaction with the arrival time of the connection close signal. When the valid business interaction has stopped but the connection close signal has not arrived within the preset waiting time, a pseudo hold flag is written.

[0090] First, set the preset waiting time to twice the normal connection closure waiting time. The normal connection closure waiting time is usually determined by the TIME_WAIT state duration of the operating system's TCP stack, with a typical value of 5 seconds. Therefore, set the preset waiting time to 10 seconds.

[0091] Additional notes: When a connection is closed normally, the connection close signal should arrive within a reasonable time after the effective business interaction stops, usually no more than 5 seconds; if it does not arrive after more than 10 seconds, it indicates that the session may be abnormally maintained, such as the application not calling the close function correctly or the session maintenance logic being incorrect.

[0092] The third sub-step is used to generate an empty occupancy flag. It compares the time when the cache placeholder length falls back to the idle threshold with the resource release confirmation time. When the cache placeholder has fallen back but the resource release confirmation has not yet occurred, an empty occupancy flag is written.

[0093] First, the idle threshold is set to 5% of the total cache capacity. When the cache space drops below 5% of the total capacity, from a resource utilization perspective, the cache instance can be considered idle and can be reallocated. If the resource release confirmation has not yet occurred at this time, it means that the resource instance is in an empty occupied state where the cache has been cleared but the resources have not been released. For example, if the total capacity of a cache instance is 100MB, when the cache space drops from 80MB to 4MB, which is 4% of the total capacity, it is below the 5% idle threshold, but the resource release confirmation time has not yet been recorded, triggering the empty occupied flag.

[0094] The fourth sub-step is used to output the abnormal hysteresis judgment result. It aggregates and counts the release delay flag, false hold flag and empty occupancy flag according to the source address of the business flow, scheduling group and resource pool number. When the three types of flags appear repeatedly in the continuous statistics window and the number of empty occupancy flags exceeds the preset number, the abnormal hysteresis judgment result is output.

[0095] First, the preset number is set to 5. Under normal operating conditions, occasional empty occupancy markers are usually isolated and few in number. Through statistical analysis, the average number of empty occupancy markers per hour under normal conditions is 0.5, with a standard deviation of 0.8. Setting the threshold to 5 is equivalent to more than 5 times the standard deviation, which can effectively identify systemic resource lag problems. For example, within three consecutive 60-second windows, under the same source address of the business flow, release delay markers and pseudo-hold markers appear in each window, and the number of empty occupancy markers reaches 3, 4, and 6 respectively, accumulating to more than 5, triggering abnormal lag judgment.

[0096] S5: Based on the abnormal rotation judgment result, the node load result after the credibility correction, and the abnormal hysteresis judgment result, generate hierarchical scheduling instructions and execute the corresponding resource scheduling actions; Considering that the first three steps have identified different types of abnormal states, these identification results need to be integrated into a unified scheduling decision so that resource scheduling actions can be accurately applied to different types of abnormal sources, while ensuring that real business requests receive supplementary resources first.

[0097] Specifically, this step first establishes an association table based on the source address of the business flow, the target resource pool number, and the Internet resource scheduling environment number. This association table aligns the output results of steps S2, S3, and S4, so that the false demand backtracking risk, the intensity of real business pressure, and the degree of resource release blockage for each source set and each resource pool can be comprehensively evaluated. Subsequently, scheduling levels are generated in a fixed order. When the risk of false demand backtracking reaches the trigger level and the actual business pressure intensity does not reach the expansion threshold, a suppression priority level scheduling instruction is generated. The risk of false demand backtracking reaching the trigger level specifically means that the number of times the marker in the abnormal backtracking judgment result reaches the preset threshold, that is, the source is judged as abnormal backtracking traffic. The expansion threshold is set based on the node load graph after credibility correction. For example, when the corrected load exceeds 85% for three consecutive statistical windows, the expansion release level is triggered.

[0098] In a cloud data center environment, when the node load exceeds 85%, performance begins to decline significantly and response time increases sharply. Setting the expansion threshold to 85% can reserve buffer space before performance decline occurs. Setting three consecutive windows avoids accidental expansion caused by instantaneous fluctuations.

[0099] When the resource release blocking level reaches the trigger level, a release priority scheduling instruction is generated. Specifically, the resource release blocking level reaching the trigger level means that the abnormal stagnation judgment result is output, that is, there is an empty resource instance.

[0100] When the actual business pressure reaches the expansion threshold, the risk of false demand reversal does not reach the trigger level, and the degree of resource release blockage does not reach the trigger level, an expansion release level scheduling instruction is generated.

[0101] After generating the scheduling instruction, this step executes the corresponding resource scheduling action. For the suppression priority level, the following suppression actions are performed: the scheduling queue weight of the corresponding source set is reduced from the default value of 100 to the suppression weight of 10; its mapping priority for entering the high-performance resource pool is reduced, and its subsequent requests are forced to be directed to the restricted isolated resource pool; its session retention time is shortened from the default 30 minutes to 5 minutes.

[0102] In weighted fair queue scheduling, the weight determines the bandwidth share that the queue receives. The default weight of 100 represents normal priority. Setting the suppression weight to 10 means that the queuing share of this source set is reduced to 10% of that of normal business. This ratio setting can effectively limit the resource consumption of abnormal sources while preserving basic service capabilities and avoiding the complete blocking of potentially legitimate business.

[0103] Normal business session duration is set to 30 minutes to support long user sessions. For suppressed sources, their behavior patterns have deviated from normal. Shortening the session duration to 5 minutes can accelerate the release of idle connections and reduce resource occupation time.

[0104] For release priority levels, perform the following release actions: arrange resource instances in the order in which the empty occupancy marker appears, and release resource instances that are ranked higher and have met the cache placeholder fallback conditions first; set a shorter session retention judgment duration for the corresponding source set, for example, from 30 seconds to 5 seconds; prevent it from re-entering the high-performance resource pool within the preset cooldown period.

[0105] The initial cooldown period is set to 10 minutes. After most of the abnormal load assessment results are released, if the source problem is not resolved, resource occupation may occur again in a short period of time. The 10-minute cooldown period provides enough time for the abnormal state to subside without being too long and affecting potential service recovery. If the source is detected again as having abnormal load assessment results after the cooldown period ends, the next cooldown period will be doubled to 20 minutes, and so on, forming a punitive incremental mechanism.

[0106] For the expansion release level, the following release action is performed: the released resource instances are backfilled into the scheduling queue where the real business requests are located, so that business flows with high load credibility in step S3 and not judged as abnormal loops in step S2 can obtain supplementary resources first.

[0107] S6: After the fixed verification time following the execution of the scheduling action, recalculate the actual number of completed requests, average resource occupation time, release confirmation completion rate, and request completion rate in the restricted resource pool for each scheduling queue; compare the statistical results with the data for the corresponding time period before scheduling execution item by item, and update the preset thresholds or preset parameters in the backswing behavior identification, credibility reduction, or imbalance tracking steps according to the preset adjustment rules; use the updated thresholds and parameters for the next scheduling cycle.

[0108] Considering that after the scheduling action is executed, its effect needs to be verified and the system parameters need to be continuously optimized based on the verification results, so that the entire scheduling system has the ability to learn and adapt.

[0109] Specifically, after a fixed verification period following the execution of a scheduling action, this step recalculates the actual number of completed requests, average resource usage time, release confirmation completion rate, and request completion rate within the restricted resource pool for each scheduling queue. The fixed verification period is set to 5 minutes, a value based on observations of the time it takes for scheduling effects to manifest: most scheduling actions produce observable changes within 5 minutes.

[0110] Subsequently, the statistical results are compared item by item with the data of the corresponding time period before the scheduling execution, and the preset thresholds or preset parameters in steps S2, S3 and S4 are updated according to the preset adjustment rules. For the verification of the suppression strategy: Assume that before the suppression action is executed, the number of actual completed requests of a certain suppression priority source set is 1000 times / minute, and the average resource occupation time is 8 seconds. After the suppression action is executed for 5 minutes, the actual completed requests are re-counted and found to be 980 times / minute, a decrease of 2%; the average resource occupation time is 6 seconds, a reduction of 25%. Since the decrease in the number of actual completed requests is less than 5%, while the reduction in the average resource occupation time is more than 10%, the suppression strategy is determined to be effective. At this time, the short-cycle monitoring time in step S2 is adjusted from 5 seconds to 4.75 seconds, that is, a reduction of 5%, to identify false demands earlier.

[0111] Conversely, if the actual number of completed requests drops to 750 per minute, a decrease of 25%, which exceeds 20%, it indicates that the suppression may be excessive. The short-cycle monitoring duration should be adjusted to 5.5 seconds, which is 10% longer, to reduce the false positive rate.

[0112] For the verification of the release strategy: Assuming that before the release action is executed, the release confirmation completion rate of a certain release priority source set is 70%, and the actual business queue completion rate is 80%, and 5 minutes after the release action is executed, the release confirmation completion rate is recalculated to be 97%, and the actual business queue completion rate is 88%. If the release confirmation completion rate increases to over 95% and the actual business queue completion rate increases by 10%, the release strategy is deemed effective and the current release strategy is retained. If the release completion rate increases to 95% but the actual business queue completion rate remains at 80% without improvement, it indicates that the released resources may be preempted by other abnormal sources. In this case, the search scope of the abnormal lag source set is expanded to include adjacent scheduling groups in the monitoring.

[0113] For the verification of the expansion strategy: If a node corresponding to the expansion release level still experiences a release blocking event after the expansion is performed, the node will be given priority in the release tracking in the next round of processing, and the judgment threshold for its release lag mark will be narrowed by 20% from 30 seconds to 24 seconds. This mechanism ensures more sensitive monitoring of nodes that are prone to release blocking.

[0114] The output of this step is the updated threshold correction basis and the scheduling parameters for the next round. These updated parameters are used in the next scheduling cycle, making the whole method form a closed-loop execution chain of continuous observation, continuous correction, and continuous adjustment. Example

[0115] This embodiment is basically the same as Embodiment 1, except for the execution details of the scheduling action and the dynamic adjustment mechanism of the cooling time.

[0116] In this embodiment, for the set of sources marked as having a suppression priority level, the queuing weight is reduced using a tiered suppression strategy. For the first violation, the weight is reduced to 30% of the default weight, i.e., from 100 to 30; for a second violation, it is reduced to 10%; and for three or more violations, it is reduced to 5% or service is denied directly. This tiered suppression design is based on the following considerations: for the first occurrence of abnormal behavior, relatively mild suppression measures are given to avoid accidentally damaging anomalies that may be caused by accidental factors; for recurring abnormal behavior, the suppression intensity is gradually increased until its resource acquisition ability is completely blocked. This gradient design ensures both the fault tolerance of the system and the effective control of persistent anomaly sources.

[0117] Meanwhile, for the set of sources marked as having a release priority, the dynamic adjustment of the cooldown duration adopts an exponential backoff mechanism, with the initial cooldown duration set to 10 minutes. If the source is detected again with the behavior corresponding to the abnormal stagnation judgment result after the cooldown period ends, the next cooldown duration is doubled to 20 minutes; if it occurs again, it is doubled to 40 minutes; and so on, until the preset upper limit is reached. This exponential backoff mechanism is based on the following considerations: sources with repeated occurrences of the behavior corresponding to the abnormal stagnation judgment result may have a more persistent behavior pattern, requiring a longer cooldown period to block their resource occupation behavior. The exponential growth design ensures that the cooldown duration can quickly respond to the severity of the abnormal behavior.

[0118] Furthermore, this embodiment refines the resource mapping priority reduction mechanism. The scheduling controller maintains a dynamic source-resource pool mapping blacklist. When a source is marked as having priority suppression, its subsequent requests are searched in the resource mapping unit, first checking the blacklist. If a match is found, it is directly mapped to a preset restricted resource pool identifier. Simultaneously, in the high-performance resource pool mapping priority table, this source is set to the lowest priority. Only after all requests from non-suppressed sources have been allocated will a request from this source have a chance to be allocated high-performance resources. This dual mechanism of blacklist and priority table ensures the determinism and efficiency of suppression actions. Example

[0119] This embodiment provides an Internet resource scheduling system for executing the methods of Embodiment 1 or Embodiment 2 described above; like Figure 3 As shown, the system includes a data acquisition unit, a slewing identification unit, a reliable reduction unit, an imbalance tracking unit, and a scheduling control unit.

[0120] The data acquisition unit is used to acquire full-path observation records of business flows in the Internet resource scheduling environment. This unit communicates with the access gateway, forwarding nodes, service nodes, and resource pool management unit, and receives observation data reported by each node. The data acquisition unit contains a clock synchronization module, which is used to perform coarse clock synchronization of each node through the network time protocol and assign a globally unique logical sequence number to each business flow. The clock synchronization module is also responsible for prioritizing the continuity of logical sequence numbers when comparing time sequences, and only using network time protocol timestamps when logical sequence numbers are missing.

[0121] The turn recognition unit is connected to the data acquisition unit and is used to identify turn behavior in the business flow based on the full path observation record, and generate abnormal turn judgment results. The turn recognition unit contains a short-cycle turn marking module, a path contraction marking module, and an anomaly judgment module. The short-cycle turn marking module is used to compare the time difference between the current entry time and the most recent departure time to generate a short-cycle turn marking. The path contraction marking module is used to compare the number of path nodes entered twice consecutively to generate a path contraction marking. The anomaly judgment module is used to count the number of times the two types of markings appear simultaneously in the statistics window and output the abnormal turn judgment results.

[0122] The Trust Reduction Unit is connected to the Data Acquisition Unit and the Turnaround Identification Unit. Based on the full-path observation records and abnormal turnaround judgment results, it performs trust reduction on the monitoring data of the access layer, forwarding layer, and service layer, generating a node load result with corrected trust. The Trust Reduction Unit contains a distortion calculation module, a latency verification module, a percentage statistics module, and a load reduction module. The distortion calculation module is used to calculate the request delay, packet attenuation, and task delay. The latency verification module is used to calculate the reporting latency and compare it with the tolerance threshold. The percentage statistics module is used to calculate the percentage of abnormal turnaround traffic in the load of each layer. The load reduction module is used to reduce the load value of the low-trust layer.

[0123] The imbalance tracking unit is connected to the data acquisition unit and the reliability reduction unit. Based on full-path observation records and node load results corrected for reliability, it performs imbalance tracking on the resource release process and generates anomaly hysteresis judgment results. The imbalance tracking unit internally includes a release hysteresis detection module, a false hold detection module, an idle occupancy detection module, and an aggregation statistics module. The release hysteresis detection module compares the resource allocation time with the resource release confirmation time to generate a release hysteresis flag. The false hold detection module compares the effective business interaction time with the connection close signal arrival time to generate a false hold flag. The idle occupancy detection module compares the cache placeholder fallback time with the resource release confirmation time to generate an idle occupancy flag. The aggregation statistics module aggregates and statistically analyzes the three types of flags by source address, scheduling group, and resource pool number, outputting anomaly hysteresis judgment results.

[0124] The scheduling control unit is connected to the turnaround identification unit, the reliability reduction unit, and the imbalance tracking unit. It is used to generate hierarchical scheduling instructions based on the abnormal turnaround judgment results, the node load results after reliability correction, and the abnormal stagnation judgment results, and to execute the corresponding resource scheduling actions. The scheduling control unit contains a hierarchical judgment module and an instruction execution module. The hierarchical judgment module is used to establish an association table and generate suppression priority, release priority, or expansion release level scheduling instructions based on the pseudo demand turnaround risk, the intensity of real business pressure, and the degree of resource release blockage. The instruction execution module is used to issue specific suppression, release, or release action instructions to the queuing control unit, the resource mapping unit, the session management unit, and the resource pool management unit.

[0125] In this embodiment, the data acquisition unit, slewing identification unit, reliable reduction unit, imbalance tracking unit, and scheduling control unit can be deployed on the same physical server or distributed across multiple physical servers. They communicate via a high-speed network, and the units exchange data through well-defined interfaces to ensure the scalability and maintainability of the system.

[0126] To verify the technical effectiveness of the present invention, comparative tests were conducted in a typical cloud data center environment. The test environment included 5 access gateways, 10 forwarding nodes, and 20 service nodes, simulating a mixed load scenario that included normal business traffic, DDoS attack traffic, and backtracking traffic caused by configuration errors.

[0127] Test results show that, using the method of this invention, the accuracy rate of identifying false peak traffic reaches 96.7%. Specifically, 96.7% of the abnormal backhaul traffic observed during the test was accurately identified and marked as abnormal backhaul. Compared with existing technologies, the number of erroneous capacity expansions caused by false peak traffic is reduced by 87.5%. In scenarios where multi-layer monitoring data is distorted, this invention, through a reliability reduction mechanism, reduces the average deviation between the load value used for scheduling decisions and the actual load value from 42% in traditional methods to 8.3%. Regarding resource stagnation detection, this invention can detect stagnation and trigger release actions within an average of 45 seconds after resources are abnormally occupied, while traditional methods cannot detect such stagnation. In terms of resource recovery efficiency, the average recovery time for abnormally occupied resources is shortened from the inability to recover resources using traditional methods to 62 seconds, and the average resource waiting time for real business requests is reduced from 340 milliseconds in traditional methods to 120 milliseconds, significantly improving the service quality of real businesses.

[0128] The present invention provides an Internet resource scheduling method and system that can be widely applied in Internet service scenarios that require dynamic resource scheduling, such as cloud data centers, edge computing nodes, and content delivery networks. Through a closed-loop mechanism that links rotation identification, trusted reduction, and release tracking, it can effectively improve resource utilization and ensure the service quality of real businesses, and has significant industrial practical value.

[0129] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0130] 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 method for scheduling Internet resources, characterized in that, Includes the following steps: S1: Obtain the full path observation record of the service flow in the Internet resource scheduling environment. The full path observation record includes the time record of the service flow in the access gateway, forwarding node, service node and resource pool management unit, the order of path nodes and the resource occupation and release status information. S2: Based on the full path observation record, identify the turnaround behavior of the business flow and generate an abnormal turnaround judgment result. The turnaround behavior identification includes detecting the multiple entry behaviors of the business flow in a short period of time and the shrinkage behavior of the number of path nodes. S3: Based on the full path observation records and the abnormal loop determination results, the credibility of the monitoring data of the access layer, forwarding layer and service layer is reduced to generate the node load result after credibility correction. The credibility reduction includes calculating the monitoring distortion degree of each layer and reducing the load value of the low credibility layer according to the proportion of abnormal loop traffic. S4: Based on the full path observation record and the node load result after the credibility correction, perform imbalance tracking on the resource release process and generate an abnormal lag judgment result. The imbalance tracking includes detecting resource release lag, session pseudo-holding and cache empty occupancy status. S5: Based on the abnormal rotation determination result, the node load result after the credibility correction, and the abnormal hysteresis determination result, generate a hierarchical scheduling instruction and execute the corresponding resource scheduling action; The resource scheduling action includes adjusting the scheduling queue, and the abnormal loop traffic is the traffic corresponding to the abnormal loop determination result.

2. The method for scheduling Internet resources according to claim 1, characterized in that, The slewing behavior identification includes: The time of the current entry into the Internet resource scheduling environment is compared with the time of the most recent departure from the Internet resource scheduling environment of the same business flow. When the time difference between the two is less than the preset short-cycle monitoring duration, a short-cycle rotation mark is generated. For the same business flow, the number of path nodes corresponding to two consecutive entries into the Internet resource scheduling environment is compared. When the number of path nodes in the later entry is less than the number of path nodes in the earlier entry and the number of forwarding table duplicate hits in the corresponding time period continues to increase, a path shrinkage mark is generated. Within the preset statistics window, when the number of times that the short-cycle turn-around marker and the path contraction marker under the same target resource pointing information appear simultaneously reaches the preset trigger count, the abnormal turn-around judgment result corresponding to the business flow is output.

3. The method for scheduling Internet resources according to claim 1, characterized in that, The credibility reduction includes: Calculate the request retention rate at the access layer, the packet attenuation rate at the forwarding layer, and the task retention rate at the service layer, and obtain the reporting latency and verification failure count for each layer; Calculate the percentage of abnormal loop traffic contributed by the abnormal loop determination results in the load of each layer; When at least one of the request retention rate, packet attenuation rate, or task retention rate of a certain layer exceeds a preset threshold, and the reporting latency and verification failure number both exceed preset thresholds, and the proportion of abnormal return traffic exceeds a preset proportion, the layer is marked as a low-trust layer. The original load values ​​of the low-confidence layer are reduced according to a preset reduction ratio, while the original load values ​​of other layers are directly retained, resulting in the node load results after confidence correction.

4. The method for scheduling Internet resources according to claim 1, characterized in that, The imbalance tracking includes: For each resource instance, the resource allocation time and the resource release confirmation time are compared. When the time difference between the two continues to widen and exceeds the preset release duration in multiple consecutive statistical windows, a release lag flag is generated. Wherein, the resource instance is the specific resource instance corresponding to the resource occupation and release status information; Compare the time of the last valid business interaction with the arrival time of the connection close signal. If the valid business interaction has stopped but the connection close signal has not arrived within the preset waiting time, a pseudo hold flag is generated. Compare the time it takes for the cache placeholder length to fall back to the idle threshold with the resource release confirmation time. When the cache placeholder has fallen back but the resource release confirmation has not yet occurred, generate an empty placeholder flag. Aggregate and statistically analyze release delay flags, pseudo-hold flags, and empty occupancy flags based on the source address, scheduling group, and resource pool number of the business flow. When the three types of flags appear repeatedly in the continuous statistical window and the number of empty occupancy flags exceeds the preset number, output the abnormal hysteresis judgment result.

5. The method for scheduling Internet resources according to claim 1, characterized in that, The hierarchical scheduling instructions include: Establish an association table based on the source address of the business flow, the target resource pool number, and the Internet resource scheduling environment number. Map the abnormal turnaround judgment result to the risk of false demand return, map the node load result after credibility correction to the actual business pressure intensity, and map the abnormal stagnation judgment result to the degree of resource release blockage. When the risk of false demand reversal reaches the trigger level and the actual business pressure intensity has not reached the expansion threshold, a suppression priority level scheduling instruction is generated. When the resource release obstruction level reaches the trigger level, a release priority scheduling instruction is generated; When the actual business pressure reaches the expansion threshold and the risk of false demand backtracking and the degree of resource release blockage do not reach the trigger level, an expansion release level scheduling instruction is generated.

6. The method for scheduling Internet resources according to claim 5, characterized in that, The execution of the corresponding resource scheduling action includes: In response to the suppression priority scheduling instruction, at least one of the following suppression actions is performed: Reduce the scheduling queue weight of the corresponding source set, lower its mapping priority to enter the high-performance resource pool, and shorten its session persistence time; In response to the release priority scheduling instruction, perform at least one of the following release actions: Release resource instances in the order in which they appear in the empty occupancy flag, set a shorter session retention time for the source set, and prevent them from re-entering the high-performance resource pool within the preset cooldown time. In response to the expansion release level scheduling instruction, the released resource instances are backfilled into the scheduling queue where the actual service request is located.

7. The method for scheduling Internet resources according to claim 1, characterized in that, Before obtaining the full path observation records of the service flow, a step of unified clock alignment is also included: The clocks of each node are coarsely synchronized using the network time protocol to ensure that the global time error is less than a preset error threshold. Each business flow entering the Internet resource scheduling environment is assigned a globally unique logical sequence number. The logical sequence number increments throughout the process of forwarding, processing, and releasing the business flow. All nodes record the current logical sequence number value when recording observation data. When comparing time sequences, the continuity of logical sequence numbers is prioritized, and network time protocol timestamps are only used when logical sequence numbers are missing.

8. The method for scheduling Internet resources according to claim 1, characterized in that, It also includes the following steps: S6: After the fixed verification time following the execution of the scheduling action, recalculate the actual number of completed requests, average resource occupation time, release confirmation completion rate, and request completion rate in the restricted resource pool for each scheduling queue; compare the statistical results with the data for the corresponding time period before scheduling execution item by item, and update the preset thresholds or preset parameters in the backswing behavior identification, credibility reduction, or imbalance tracking steps according to the preset adjustment rules; use the updated thresholds and parameters for the next scheduling cycle.

9. The method for scheduling Internet resources according to claim 6, characterized in that, The preset cooling time adopts a dynamic adjustment mechanism: The initial cooling time is a preset initial value; If the source set is detected again after the cooling period ends and the behavior corresponding to the abnormal hysteresis judgment result is detected, the next cooling time will be increased by a preset multiple.

10. An Internet resource scheduling system, characterized in that, include: The data acquisition unit is used to acquire the full path observation record of the business flow in the Internet resource scheduling environment. The full path observation record includes the time record of the business flow in the access gateway, forwarding node, service node and resource pool management unit, the order of path nodes and the resource occupation and release status information. The turn-around identification unit is used to identify the turn-around behavior of the business flow based on the full path observation record and generate an abnormal turn-around judgment result; The credibility reduction unit is used to perform credibility reduction on the monitoring data of the access layer, forwarding layer and service layer based on the full path observation record and the abnormal turnaround determination result, and generate a node load result with credibility correction. The imbalance tracking unit is used to perform imbalance tracking on the resource release process based on the full path observation records and the node load results after the confidence correction, and generate an abnormal lag judgment result. The scheduling control unit is used to generate hierarchical scheduling instructions and execute corresponding resource scheduling actions based on the abnormal turnaround determination results, the node load results after the credibility correction, and the abnormal stagnation determination results.

Images