A method and system for efficient operation and maintenance monitoring of computing power servers

Abstract

The application relates to the technical field of computing power server operation and maintenance, in particular to a method and system for efficient operation and maintenance monitoring of a computing power server, which comprises the following steps: for each computing power node in a computing power server cluster, calculating the demand intensity of the node based on the observation values of each dimension resource of the node within a set time window; sorting the demand intensity of the node in the same domain, determining the breakpoint prefix according to the difference between adjacent positions, extracting the maximum connected component, and obtaining a hot node set; obtaining a candidate target node set, performing rank operation on the idle degree of the candidate target node and the topological distance between the candidate target node and the hot node, obtaining an adaptive sequence position, and selecting a migration or scheduling target according to the adaptive sequence position. The application aims to avoid single node overload and ensure the smooth distribution of computing power resources.

Description

Technical Field

[0001] This application relates to the field of computing server operation and maintenance technology, specifically to a method and system for efficient operation and maintenance monitoring of computing servers. Background Technology

[0002] A computing server is a hardware system designed for high-density computing scenarios. It includes a large number of CPUs, GPUs, FPGAs, AI-ASICs, or memory pools. Through NVLink, PCIe switching, and RDMA networks, it quantifies the computing power of multiple cores, graphics memory bandwidth, and memory pool capacity, exposing a unified computing API. Running multiple workloads on a single device results in extremely high instantaneous power consumption. Any single point of overheating can trigger rack-wide power throttling. Therefore, it is essential to use a monitoring system to eliminate hotspots, reduce power consumption, optimize computing resources, and improve server operation and maintenance efficiency while ensuring business continuity.

[0003] Traditional operation and maintenance monitoring typically uses static thresholds, setting fixed upper limits for each dimension, assigning weights to different dimensions based on experience, and then weighting and summing the results to obtain the node load. When the load exceeds the threshold, warnings, migration, or frequency reduction are triggered. Because it does not consider relative anomalies across dimensions, it cannot identify relative hotspots after an overall load increase or extreme anomalies in a specific dimension. It also does not take into account the topological distance between nodes, which may cause hotspot migration to move tasks to devices with a long network hop count or those that are already implicitly overloaded, resulting in the recurrence of hotspots and failure to reduce power consumption in a timely manner. Summary of the Invention

[0004] In view of the above, it is necessary to provide a method and system for efficient operation and maintenance monitoring of computing servers to solve the above problems.

[0005] The first aspect of this application provides a method for efficient operation and maintenance monitoring of a computing server, the method comprising:

[0006] Collect data on various resource dimensions of each computing node at each moment, analyze the distribution characteristics of data fluctuations in various resource dimensions, and determine the demand intensity of each computing node;

[0007] The demand intensity of all computing power nodes in the same domain is sorted, and the breakpoint prefix is ​​obtained based on the difference distribution between adjacent demand intensities. The maximum connected component is extracted from the subgraph induced by the computing power node corresponding to the breakpoint prefix to obtain the hot spot node set.

[0008] The scheduling of tasks to be migrated for hot nodes is specifically as follows: Analyze the characteristics of each computing node (excluding the set of hot nodes) when accepting node tasks to obtain a set of candidate target nodes; analyze the positional distribution differences between each hot node and each candidate target node in the set of hot nodes, and determine the adaptation order between each hot node and each candidate target node based on the demand intensity; obtain the target node for each hot node; schedule the tasks to be migrated for the hot nodes to the target nodes; repeat the scheduling process for the tasks to be migrated for the hot nodes until the hot nodes are no longer hot.

[0009] Preferably, determining the demand intensity of each computing node specifically involves:

[0010] For each computing node, the average data of various resource dimensions at all times within each time window is used as the observation value of each resource dimension in each time window;

[0011] The formula for calculating the demand intensity is: ,in, This indicates the demand intensity of computing power node v; Represents the maximum value function; This represents the observed value of resource dimension j for computing node v within the current time window; This represents the median of the observed values ​​of resource dimension j for computing node v within the current and previous preset number of time windows; This represents the interquartile range of the observed values ​​of resource dimension j within the current and previous preset number of time windows for computing node v; This represents a preset constant to prevent the denominator from being zero.

[0012] Preferably, the process of obtaining the breakpoint prefix is ​​as follows:

[0013] Sort the demand intensity of all computing power nodes in the same domain in descending order to obtain the demand intensity sequence. Calculate the difference between each element in the demand intensity sequence and the next element. Use the set of the index value corresponding to the largest difference and the index values ​​of the previous differences as the breakpoint prefix.

[0014] Preferably, the step of obtaining the hotspot node set is as follows:

[0015] Using the computing power nodes in the breakpoint prefix as vertices, retain the edges in the topology graph within the same domain that have both ends in the breakpoint prefix to obtain a subgraph; run the search algorithm to find all connected components; select the connected component with the most vertices as the hotspot node set.

[0016] Preferably, if there are several connected components with the most vertices, the connected component with the smallest rank sum in the demand intensity sequence of computing power nodes is selected as the hot spot node set.

[0017] Preferably, obtaining the candidate target node set specifically involves:

[0018] The difference between the natural number 1 and the demand intensity of each computing node is used as the idle degree of each computing node;

[0019] For computing power nodes other than the hotspot node set, exclude computing power nodes that are reserved, under maintenance, or faulty. Then exclude computing power nodes whose remaining resources are still insufficient to execute the current hotspot node task after being estimated according to the minimum container specifications. Finally, sort the remaining computing power nodes in the same domain from high to low according to their idle degree, and select the top N as candidate target nodes, where N is a preset value.

[0020] Preferably, the specific formula for determining the adaptation order between each hotspot node and each candidate target node is as follows: ;in, This indicates the matching order between the hot spot node v and the candidate target node t; Indicates the demand intensity of candidate target node t; This represents the topological distance between the hot spot node v and the candidate target node t, which is determined by the positions of the hot spot node and the candidate target node, as well as the data transmission link. This represents the idle degree of candidate target node t; Represents the rank statistic function; This represents a preset constant to prevent the denominator from being zero.

[0021] Preferably, determining the target node specifically involves selecting the candidate target node with the smallest adaptation order with each hotspot node as the target node for each hotspot node.

[0022] Preferably, the condition for scheduling the migration task of the hot spot node to the target node is: the adaptation order of the hot spot node and its target node is less than the first preset percentile of all adaptation orders in the candidate target node set.

[0023] Secondly, embodiments of this application also provide a system for efficient operation and maintenance monitoring of computing power servers, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0024] This application has at least the following beneficial effects:

[0025] This application collects computing power node resource dimension data, continuously monitors the resource usage of each node, and promptly identifies nodes with resource shortages. This helps to quickly identify and locate problems, providing a basis for subsequent task scheduling and ensuring flexible adjustments based on actual load conditions to avoid resource waste or overload. By analyzing data fluctuations, the demand intensity of different computing power nodes is extracted. Demand intensity, as a quantitative indicator, helps to accurately determine the load demand of nodes, ensuring more scientific and reasonable task scheduling and resource allocation. The demand intensity of all computing power nodes within the same domain is sorted, and the differences in demand intensity between adjacent nodes are analyzed to identify breakpoint prefixes of demand changes. By analyzing the differences in demand intensity, regions with drastic changes in resource demand can be quickly identified, helping to discover potential hotspot nodes. The maximum connected component is extracted from the subgraph induced by the computing power node corresponding to the breakpoint prefix, thereby obtaining the set of hotspot nodes. By obtaining the maximum connected component... It can efficiently locate hotspots requiring focused attention, which may be the areas with the greatest performance bottlenecks or resource pressure. For each hotspot node, based on its resource characteristics and demand intensity, it obtains a set of candidate target nodes and analyzes their topological distance, adaptation order, etc., to determine the most suitable target node. It can select the target node with the least resources and closest to the hotspot node, effectively migrating tasks to nodes with lighter loads, ensuring load balancing across the entire system and preventing some nodes from overloading. It migrates tasks from hotspot nodes to target nodes, repeating this process until the load on the hotspot node is alleviated and the hotspot is resolved. Through task migration, the load of the hotspot node is balanced to other nodes, avoiding localized system overload, improving the overall system stability and responsiveness, helping to avoid overloading of individual nodes, ensuring the smooth distribution of computing resources, thereby improving the overall system performance, and preventing hotspot nodes from existing for extended periods, ensuring the efficient utilization of computing resources. Attached Figure Description

[0026] Figure 1 A flowchart illustrating the steps of a method for efficient operation and maintenance monitoring of a computing server, provided in one embodiment of this application;

[0027] Figure 2 A flowchart illustrating the acquisition of a target node is provided for one embodiment of this application. Detailed Implementation

[0028] In the description of the embodiments in this application, the words "exemplary," "or," and "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary," "or," and "for example" is intended to present the relevant concepts in a specific manner.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application's specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0030] It should also be noted that the terms "first" and "second" in this application and its accompanying drawings are used to distinguish similar objects, rather than to describe a specific order or sequence. The methods disclosed in the embodiments of this application or the methods shown in the flowcharts include one or more steps for implementing the method. Without departing from the scope of protection of this application, the execution order of multiple steps can be interchanged, and some steps can also be deleted.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0032] The following description, in conjunction with the accompanying drawings, details the specific solution of the method and system for efficient operation and maintenance monitoring of computing servers provided in this application.

[0033] Please see Figure 1 The diagram illustrates a flowchart of a method for efficient operation and maintenance monitoring of a computing server according to an embodiment of this application. The method includes the following steps:

[0034] Step 1: Collect data on various resource dimensions of each computing node at each moment, analyze the distribution characteristics of data fluctuations in various resource dimensions, and determine the demand intensity of each computing node.

[0035] In heterogeneous computing servers, various dimensions, such as CPU, GPU, video memory, RAM, bandwidth, power consumption, and temperature, have different dimensions and physical meanings. Therefore, traditional fixed threshold or weighted average methods are easily masked by extreme anomalies in certain dimensions. To more accurately measure demand intensity, each dimension should be standardized against its historical baseline. This can avoid misjudgments and lags caused by differences in dimensions.

[0036] A computing node refers to the smallest physical or logical unit in a cluster that can be independently scheduled and carry computing tasks. It can be a complete machine, a CPU socket, a GPU card, an FPGA board, or an abstract service of a container. The node exposes a unified computing power interface to the outside world.

[0037] In this application, the time window is set to 30 seconds, the sliding step is 1 second, and the average value is taken as the current observation value. This ensures rapid capture of sudden hotspots while avoiding false triggers caused by jumps. A monitoring process runs on each computing node, collecting data on various resource dimensions every second. These resource dimensions include: CPU utilization, GPU utilization, VRAM usage, main memory usage, network inbound and outbound bandwidth usage, node real-time power consumption, core temperature, NVMe disk I / O utilization, container-level GPU power consumption, and L2 cache hit rate.

[0038] The formula for calculating the demand intensity of a node is: ,in, This represents the demand intensity of computing power nodes v; the larger the value, the more strained the nodes are. Represents the maximum value function; This represents the observed value of resource dimension j for computing node v within the current time window; This represents the median of the observed values ​​of resource dimension j for computing node v within the current and previous preset number of time windows, serving as the normal baseline for that dimension; It represents the interquartile range of the observed values ​​of resource dimension j within the current and previous preset number of time windows, measuring the normal fluctuation range; The preset constant is used to prevent the denominator from being zero, and its value is 1e-6. In this embodiment, the preset quantity is 10. The implementer can adjust it according to the actual situation. This application does not impose any restrictions on this.

[0039] First, by measuring the deviation of the current value from the normal level by the difference between the observed value and the median, this method is unaffected by the original units. The interquartile range represents the middle 50% of historical data, reflecting the magnitude of normal fluctuations. By converting the deviation into the equivalent of several normal fluctuations, cross-dimensional comparisons can be achieved. Furthermore, because the median and interquartile range are insensitive to abrupt changes, this method avoids the problem of the mean method being affected by extreme values, ensuring that deviations in any dimension can be detected and identified promptly.

[0040] Step 2: Sort the demand intensity of all computing power nodes in the same domain, and obtain the breakpoint prefix based on the difference distribution between adjacent demand intensities; extract the maximum connected component in the subgraph induced by the computing power node corresponding to the breakpoint prefix to obtain the hot spot node set.

[0041] In heterogeneous high-power computing server clusters, changes in node demand intensity often manifest as a plateau at the beginning and a sudden drop at the end, causing a break in the demand intensity curve. To address this issue, truly high-demand nodes can be adaptively identified by maximizing the gap, avoiding mis-expansion or omission due to empirical selection under tidal load conditions. Furthermore, by performing topology analysis on the breakpoint prefixes to identify the largest connected clusters, isolated noise nodes can be effectively eliminated, preserving concentrated areas of high-load nodes. This allows subsequent task migrations to focus more on nodes within the same domain, thereby minimizing the overhead of cross-region migrations.

[0042] Within the same domain, the comparison range in a single sorting is typically the entire cluster, a single rack (within the same PDU power supply unit), all servers connected to the same switch, or a custom fault domain. In this embodiment, the entire cluster is treated as the same domain. First, the demand intensity of all computing nodes within the same domain is sorted in descending order to form a demand intensity sequence.

[0043] The set of hotspot nodes can be represented as ,in, , Let i and y represent the i-th and (i+1)-th values ​​of the demand intensity sequence, respectively. Indicating the demand intensity sequence The index value at the maximum value will be used to determine the maximum gap. The corresponding points are used as breakpoints to divide the sequence into a high-value plateau at the beginning and a low-value area at the end, thereby determining the hotspot boundary. This can not only avoid missing true hotspots when the load is uplinked, but also prevent the erroneous migration of healthy nodes when downlinked, thus achieving adaptive hotspot identification. This indicates a breakpoint prefix.

[0044] Indicates the index value of the computing node; Represents the rank statistic function; Let represent the rank of computing node 'a' in the demand intensity sequence. The breakpoint prefix is ​​obtained by cutting the descending demand intensity sequence from the first node to the node with the largest gap. This provides input for topology connectivity filtering, ensuring that the final hotspots simultaneously possess both high demand intensity and clustered distribution, facilitating batch and low-cost migration. It should be noted that the rank in this application does not refer to the rank of a matrix, but rather to the rank statistic; the rank is the index itself when a set of real numbers is sorted by size.

[0045] This represents the maximum connected component. In the topology graph, the subgraph with the most nodes and interconnected nodes is obtained from the breakpoint prefix, thus eliminating isolated noise points and retaining high-load nodes in contiguous areas. This concentrates subsequent migration actions within physically adjacent connected domains, reducing cross-regional overhead and the risk of secondary hotspots. Specifically: using the computing power nodes in the breakpoint prefix as vertices, edges with both ends in the breakpoint prefix are retained in the original cluster (same domain) topology graph to obtain a subgraph; Breadth-First Search (BFS) is run to find all connected components. In other embodiments, Depth-First Search (DFS) can also be used to find all connected components; the connected component with the most vertices is selected as the hotspot node set; if there are multiple connected components with the most vertices, the connected component with the smallest rank sum among the computing power nodes in the demand intensity sequence is selected as the hotspot node set.

[0046] Step 3: Schedule the tasks to be migrated for hot nodes. Specifically, analyze the characteristics of each computing node (excluding the set of hot nodes) when accepting node tasks to obtain a set of candidate target nodes; analyze the positional distribution differences between each hot node and each candidate target node in the set of hot nodes, and determine the adaptation order between each hot node and each candidate target node based on the demand intensity; obtain the target node for each hot node; schedule the tasks to be migrated for the hot nodes to the target nodes; repeat the scheduling process for the tasks to be migrated for the hot nodes until the hot nodes are no longer hot.

[0047] Traditional weighted scoring directly uses idleness and topology distance as weights. This can lead to issues such as targeting distant, low-load nodes due to differences in measurement units, or repeatedly selecting the wrong target node when business needs change due to fixed weights. This application converts idleness and topology distance into ranking numbers within the same domain, then sums them using their rank. This allows nearby idle nodes to become target nodes under any load condition. The decision result can be interpreted as the node's idleness ranking and distance ranking, thus migrating hot tasks to the most suitable location with the lowest migration cost, avoiding secondary hotspots and unnecessary power consumption.

[0048] For each hot node in the hot node set, firstly, obtain its candidate target node set, which is the set of computing power nodes that can accept hot node tasks and will not create new hot nodes. First, exclude nodes in the hot node set, then exclude nodes that are reserved, under maintenance, or faulty. Next, exclude nodes whose remaining resources are estimated according to the minimum container specifications and still cannot execute the current hot node task. Finally, sort the remaining healthy nodes in the same domain from high to low idle degree and select the top N as the candidate target node set. In this embodiment, N is 15. Implementers can choose a suitable value from 10 to 20 according to the actual situation. This application does not impose any restrictions on this.

[0049] This application uses rank instead of specific numerical values ​​or weighted calculations because in heterogeneous high-power computing server clusters, the dimensions and ranges of idleness and topology distance differ significantly. Furthermore, with fluctuating business activity, idle computing nodes with greater topology distances may be misjudged as optimal, while those with closer topology distances but not idle may be incorrectly identified as optimal. By converting these two metrics into ranking numbers within the same candidate set, the difference in dimensions can be eliminated, the influence of extreme values ​​on the results can be avoided, and errors caused by fluctuations in business activity levels can be prevented, thus achieving more interpretable and adaptive hotspot node migration selection.

[0050] The result of the adaptation order between each hotspot node and its candidate target nodes can be represented as: ,in, This indicates the matching order of hot spot node v and its candidate target node t. The smaller the value, the more suitable the candidate target node t is for receiving task migration or scheduling from hot spot node v, as it is both idle and adjacent to other candidate target nodes in the same domain.

[0051] Let represent the demand intensity of candidate target node t, then This represents the idleness of the candidate target node t. The larger the idleness value, the more idle it is, meaning it has more remaining resources and is more suitable as a target node for task migration. This represents the topological distance between hotspot node v and candidate target node t, expressed as network hop count, switch level, rack number difference, or RDMA round-trip time. A larger value indicates a greater topological distance and a higher cost for task migration between the two nodes. The larger the size, the better it is for migration.

[0052] In this embodiment, the topology distance is determined by the locations of the hotspot node and the candidate target node, as well as the data transmission link. Specifically: when the hotspot node v and the candidate target node t are on the same motherboard, they go through an NVLink / PCIe switch without passing through the network, and the value is 0; when the hotspot node v and the candidate target node t are on different motherboards in the same rack, they go through a Top-of-Rack switch, going through a Level 1 network, and the value is 1; when the hotspot node v and the candidate target node t are in different racks, they go through a Level 1 Spine switch, and the value is 2, and through a Level 2 Leaf-Spine switch, the value is 3; when the hotspot node v and the candidate target node t are in different data centers, they need to go through a core switch or even a wide area link, in which case the bandwidth is low and the latency is high, and the value is 10. During system initialization, the topology discovery service is used to write the computing server topology into the memory table, and it is updated when nodes go online or offline or when the link changes. It should be noted that 1, 2, 3, and 10 are all preset values, and implementers can adjust them according to their relative magnitudes based on actual conditions.

[0053] This represents the idle degree of candidate target node t; This represents the proximity between the hotspot node v and the candidate target node t; This means sorting all nodes in the candidate target node set in descending order according to their idle degree, and obtaining the idle degree rank of candidate target node t; This represents the rank of the proximity of hotspot node v and candidate target node t, sorted in descending order of all proximity in the same domain.

[0054] For a hotspot node v in the same domain and its set of candidate target nodes, the selection is based on the adaptation order of all nodes with the candidate target nodes. The minimum value is the optimal target. For example, if the idle degree rank from v to t is 1 and the proximity degree rank is 1, then the adaptation order of this migration is 2, which is the minimum value.

[0055] If the adaptation order of hotspot node v and target node is less than the top 20% percentile of all adaptation orders in the candidate target node set, then the migration task of hotspot node v is scheduled to the target node. Otherwise, it indicates that the domain is overloaded or the distance is too far, and migration is not performed. In this case, the hotspot node is subjected to frequency reduction, rate limiting, or horizontal scaling to avoid creating secondary hotspots. The flowchart for obtaining the target node is as follows. Figure 2 As shown.

[0056] During migration, for containerized tasks, Kubernetes / Slurm is used to migrate Pods / Jobs online from hot nodes v to candidate target nodes t. For GPU processes, NVML+CUDA MPS hot migration or checkpoint-restore is used, and for virtual machines, live-migration is employed. Simultaneously, tasks migrated from hot nodes are released to reduce resource observations across various resource dimensions, thereby decreasing demand intensity.

[0057] After migration or frequency reduction is completed, the monitoring process will re-collect and calculate the demand intensity of hotspot node v in the next time window. If the hotspot is not resolved, the above steps will continue.

[0058] Based on the same inventive concept as the above method, this application embodiment also provides a system for efficient operation and maintenance monitoring of computing power servers, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described methods for efficient operation and maintenance monitoring of computing power servers.

[0059] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0060] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from its essential characteristics. Therefore, the embodiments described above should be considered exemplary and non-limiting in all respects; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some technical features, without causing the essence of the corresponding technical solutions to deviate from the scope of the technical solutions in the embodiments of this application, should all be included within the protection scope of this application.

Claims

1. A method for efficient operation and maintenance monitoring of a computing power server, characterized in that, The method comprises the following steps: Collecting data of various resource dimensions of each computing node at each time point, analyzing the distribution characteristics of the data fluctuation of various resource dimensions, and determining the demand intensity of each computing node; Sorting the demand intensities of all computing nodes in the same domain, obtaining a breakpoint prefix based on the difference distribution between adjacent demand intensities, extracting the maximum connected component in the subgraph induced by the computing nodes corresponding to the breakpoint prefix, and obtaining a hot node set; Scheduling the tasks to be migrated of the hot nodes, specifically: analyzing the characteristics of each computing node accepting node tasks except the hot node set, obtaining a candidate target node set; analyzing the position distribution difference between each hot node in the hot node set and each candidate target node, combining the demand intensity to determine the adaptation order between each hot node and each candidate target node, obtaining the target node of each hot node, and scheduling the tasks to be migrated of the hot nodes to the target node, repeating the task scheduling process of the hot nodes until the hot nodes are removed. The demand intensity of each computing node is determined, specifically: For each computing node, the mean value of the data of various resource dimensions at all time points in each time window is taken as the observation value of various resource dimensions in each time window; The demand intensity calculation formula is: wherein, represents the demand intensity of the computing power node v; represents the maximum function; represents the observation value of the resource dimension j of the computing power node v in the current time window; represents the median of the observation values of the resource dimension j of the computing power node v in the current and the preset number of previous time windows; represents the interquartile range of the observation values of the resource dimension j of the computing power node v in the current and the preset number of previous time windows; represents a preset constant for preventing the denominator from being zero; The adaptation order between each hot node and each candidate target node is determined, and the specific formula is: ; wherein, represents the adaptation rank of the hotspot node v and the candidate target node t; represents the demand intensity of the candidate target node t; represents the topological distance between the hotspot node v and the candidate target node t, which is determined by the positions of the hotspot node and the candidate target node and the data transmission link; represents the rank statistic function; represents a preset constant for preventing the denominator from being zero.

2. The method for efficient operation and maintenance monitoring of a computing power server according to claim 1, wherein, The process of obtaining the breakpoint prefix is: Arranging the demand intensities of all computing nodes in the same domain in descending order to obtain a demand intensity sequence, calculating the difference value between each element and the next element in the demand intensity sequence, and taking the set composed of the index value corresponding to the maximum difference value and the index value of the previous difference value as the breakpoint prefix.

3. The method for efficient operation and maintenance monitoring of a computing power server according to claim 1, wherein, The hot node set is obtained by: Taking the computing nodes in the breakpoint prefix as vertices, retaining edges in the same domain topology graph with both ends in the breakpoint prefix to obtain a subgraph, running a search algorithm to find all connected blocks, and selecting the connected block with the most vertices as the hot node set.

4. The method for efficient operation and maintenance monitoring of a computing power server according to claim 3, wherein, If there are several connected blocks with the most vertices, the connected block with the smallest rank sum of the computing nodes in the demand intensity sequence is selected as the hot node set.

5. The method for efficient operation and maintenance monitoring of a computing power server according to claim 1, wherein, The candidate target node set is obtained, specifically: The difference between the natural number 1 and the demand intensity of each computing node is taken as the idle degree of each computing node; For computing nodes other than the hot node set, the computing nodes that cannot execute the current hot node task after estimating the remaining resources according to the lowest container specification are excluded, and finally the remaining computing nodes in the same domain are sorted by idle degree from high to low, and the top N are taken as candidate target nodes, where N is a preset value.

6. The method for efficient operation and maintenance monitoring of a computing power server according to claim 1, wherein, The target node of each hot node is obtained by: taking the candidate target node with the smallest adaptation order of each hot node as the target node of each hot node.

7. The method for efficient operation and maintenance monitoring of a computing power server according to claim 1, wherein, The condition for scheduling the tasks to be migrated of the hot nodes to the target node is that the adaptation order of the hot node and its target node is smaller than the front preset quantile of all adaptation orders in the candidate target node set.

8. A system for efficient operation and maintenance monitoring of a computing power server, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor implements the steps of the method according to any one of claims 1-7 when executing the computer program.

Images