A host scheduling node method based on a computer remote control system

Abstract

This invention relates to the field of computer remote control and cluster scheduling technology, specifically a host scheduling node method based on a computer remote control system. The method includes: collecting operating parameters of each node in a remote host cluster and predicting fault risk using Markov chains; calculating node load rate and health; determining node status; constructing an ARIMA-LSTM hybrid model to predict future load rate; receiving a set of tasks to be scheduled, calculating task priority coefficients, and generating a scheduling queue; constructing a task-node bipartite graph, using graph neural network embedding vectors, and combining resource constraints with cosine similarity and the Hungarian algorithm to solve for the global optimal match; detecting faulty nodes, migrating incomplete tasks, isolating faulty nodes, and performing recovery detection. This invention achieves accurate perception and prediction of node status and load trends, and improves the reliability, resource utilization, and execution efficiency of high-priority tasks in cluster scheduling through intelligent global matching and proactive fault tolerance.

Description

Technical Field

[0001] This invention relates to the field of computer remote control and cluster scheduling technology, specifically a host scheduling node method based on a computer remote control system. Background Technology

[0002] With the rapid development of information technology, computer remote control systems are widely used in industrial automation, cloud computing, the Internet of Things, and edge computing. These systems consist of a control center (remote control terminal) and a host cluster composed of multiple computers. The control center is responsible for issuing various computing, control, or data processing tasks to the cluster, and the nodes in the cluster are responsible for executing these tasks. Efficient and reliable task scheduling is the core of the stable operation of such systems. Its goal is to rationally distribute tasks across the nodes in the cluster while meeting specific task requirements (such as real-time performance and resource requirements), thereby achieving high throughput, low latency, high reliability, and good cluster load balancing.

[0003] Traditional task scheduling methods primarily rely on static rules or simple dynamic strategies. Examples include round-robin scheduling, greedy algorithms based on resource utilization (such as assigning tasks to the node with the least available CPU or memory), or scheduling based on fixed-priority queues. These methods are suitable for scenarios with small system sizes, simple task types, and stable node states. However, as system sizes increase, task types become more complex, and requirements for system reliability and real-time performance rise, traditional scheduling methods reveal numerous limitations:

[0004] Traditional methods often rely solely on single, real-time metrics such as CPU and memory, neglecting critical resources like bandwidth and disk I / O. Furthermore, they fail to incorporate historical failure patterns for reliability prediction, resulting in an inability to accurately identify sub-optimal or potential fault states of nodes. This can lead to the assignment of critical tasks to unreliable nodes, causing task failures or performance degradation. Scheduling decisions are mostly based on current instantaneous load, failing to predict future load trends. When faced with fluctuating task flows, this can easily lead to resource contention and localized overload, hindering true forward-looking load balancing and causing significant fluctuations in system resource utilization. Task prioritization is simplistic, often based on fixed levels or submission times, failing to dynamically integrate multi-dimensional attributes such as task real-time performance, data volume, and business importance for refined differentiation. Under resource constraints, it's difficult to ensure that truly critical and urgent tasks receive priority access to resources, impacting the continuity of core business operations.

[0005] The matching of tasks and nodes often adopts local optimization strategies such as greedy algorithms, which makes it difficult to model the complex resource competition and hidden relationships between tasks and nodes. It is easy to get trapped in local optima and lacks in-depth exploration of the complex relationships between tasks and nodes. It is difficult to achieve the dual goals of low-latency execution of high-priority tasks and cluster load balancing, resulting in resource waste and system performance degradation.

[0006] Existing methods have relatively simple mechanisms for detecting and handling node failures, and lack forward-looking prediction of failure risks, resulting in delayed failure handling and easily causing problems such as task interruption and data loss, which affects the overall reliability of the system.

[0007] Therefore, existing methods have significant shortcomings in terms of comprehensive node health assessment, load trend prediction, dynamic task priority adaptation, global scheduling optimization, and proactive fault tolerance. Summary of the Invention

[0008] The purpose of this invention is to provide a host scheduling node method based on a computer remote control system to solve the problems mentioned in the background.

[0009] The objective of this invention can be achieved through the following technical solution: a host scheduling node method based on a computer remote control system, comprising the following steps:

[0010] Step 1, Node Status Awareness: Collect the operating parameters of each node in the remote host cluster at the current moment, calculate the evaluation value of the operating parameters of the node at the current moment, and predict the fault risk coefficient of each node at the current moment through Markov chain based on the historical fault record data of each node, and construct a segmented node health model to determine the node status.

[0011] Step 2, Node Load Trend Prediction: By constructing an ARIMA-LSTM hybrid prediction model, the overall node load rate at future moments is predicted based on historical node load rate data;

[0012] Step 3: Task Priority Sorting: Receive the set of tasks to be scheduled sent by the remote control terminal, calculate the priority coefficient of each task, and generate a task scheduling queue;

[0013] Step 4, Task-Node Scheduling Decision: Construct a task-node bipartite graph. Calculate the initial matching degree of edges based on node health, predicted load rate, and task priority. Inject feature vectors of tasks, nodes, and edges, and generate embedding vectors through GNN iterative training. Combine resource constraints with cosine similarity calculation and Hungarian algorithm to solve for the global optimal matching. Finally, generate and issue standardized scheduling instructions to achieve the optimal matching of tasks and nodes, balancing low-latency execution of high-priority tasks with cluster load balancing.

[0014] Step 5, Proactive Fault Tolerance: Detect faulty nodes, migrate incomplete tasks, isolate faulty nodes, and perform recovery checks.

[0015] The beneficial effects of this invention are:

[0016] This invention collects the operating parameters of each node in a remote host cluster, predicts node failure risks based on Markov chains, and constructs a segmented node health model to accurately classify node states. This makes node state analysis more reasonable and reliable, effectively avoiding the misjudgment problem caused by relying on a single indicator in traditional methods. This allows for reasonable task allocation and prevents performance degradation or system crashes caused by misjudgment of node states.

[0017] This invention constructs an ARIMA-LSTM hybrid prediction model, utilizing the ARIMA model's ability to capture linear trends and the LSTM model's ability to handle nonlinear fluctuations, to predict the overall load rate of nodes at future times, accurately reflecting the changing trend of node load and providing a more reliable basis for task allocation.

[0018] This invention receives a set of tasks to be scheduled from a remote control terminal, calculates the priority coefficient of each task based on the task's attribute data, and generates a task scheduling queue. This ensures that high-priority tasks are processed first, improving the execution efficiency and response speed of critical tasks. When faced with multiple concurrent tasks, resources can be intelligently allocated according to the task priority coefficients to ensure that urgent tasks are processed in a timely manner, while avoiding low-priority tasks from consuming too many resources.

[0019] This invention constructs a task-node bipartite graph by selecting available nodes, calculates the initial matching degree of edges based on node health, predicted load rate, and task priority, generates embedding vectors through iterative training using a GNN, and solves for the globally optimal match using cosine similarity calculation and the Hungarian algorithm, combined with resource constraints. A globally optimized scheduling strategy ensures optimal matching between tasks and nodes, balancing low-latency execution of high-priority tasks with cluster load balancing. When allocating tasks, the most suitable node is selected based on multiple factors such as health, predicted load rate, and task priority, thereby improving overall resource utilization and operational reliability.

[0020] This invention also has proactive fault tolerance capabilities, which can detect faulty nodes in real time and migrate unfinished tasks to healthy nodes, while isolating faulty nodes and performing recovery detection, effectively avoiding task interruption and data loss caused by node failure, and improving fault tolerance and stability. Attached Figure Description

[0021] The invention will now be further described with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a flowchart of the ARIMA-LSTM model construction process of the present invention. Detailed Implementation

[0024] 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.

[0025] Please see Figures 1-2 As shown, this invention is a host scheduling node method based on a computer remote control system, comprising the following steps:

[0026] Step 1: Node Status Awareness: Collect the operating parameters of each node in the remote host cluster at the current moment, calculate the evaluation value of the node's operating parameters at the current moment, and predict the failure risk coefficient of each node at the current moment through Markov chain based on the historical fault record data of each node. Construct a segmented node health model to determine the node status; the specific execution process is as follows:

[0027] Furthermore, the operating parameters of each node in the remote host cluster are collected at the current moment, and the evaluation value of the operating parameters of the nodes at the current moment is calculated. The specific execution process is as follows:

[0028] Get the set of nodes N in the remote host cluster. (i≥2, where i is the number of nodes in the remote host cluster), obtain the processor utilization, memory utilization, bandwidth utilization, and disk I / O utilization data of each node in the remote host cluster at the current moment. It should be noted that processor utilization reflects the load pressure of computationally intensive tasks and is the core indicator of the load rate of computationally intensive nodes; memory utilization reflects the load pressure of memory-intensive tasks, and insufficient memory will lead to frequent page swapping, which will significantly affect node performance; bandwidth utilization reflects the load pressure of transmission-intensive tasks, and bandwidth bottlenecks will lead to increased data transmission latency; disk I / O utilization reflects the load pressure of I / O-intensive tasks, and disk I / O bottlenecks will lead to increased latency in file read / write and database operations.

[0029] The standard deviation and mean of the processor utilization data of each node at the current moment are calculated to generate the standard deviation and mean of the processor utilization of each node at the current moment, denoted as . ;

[0030] Using the coefficient of variation algorithm The processor utilization evaluation value of each node at the current moment is calculated. ;

[0031] Similarly, the memory utilization assessment value, bandwidth utilization assessment value, and disk I / O utilization assessment value of each node at the current moment are calculated and denoted as follows: It should be noted that the coefficient of variation algorithm is used to quantify the fluctuation of the load of a single resource. The larger the value, the worse the resource stability, providing a basic characteristic for the subsequent overall load rate.

[0032] Furthermore, based on the historical fault record data of each node, the fault risk coefficient of each node at the current moment is predicted using a Markov chain, as follows:

[0033] Extract nodes from the node monitoring database Historical fault record dataset ,in These are the start and end timestamps for the d-th fault, respectively. d1 represents the fault type; d1 represents the total number of faults.

[0034] Set time slice length Divide the historical timeline into continuous time slices ; It represents the set of positive integers.

[0035] For each time slice k, define a binary state variable. , , This indicates that a fault occurred within time slice k (node ​​fault state). This indicates that no faults occurred within time slice k (the node is in a normal state), and generates a sequence of the node's historical states. Where T-1 is the index of the last complete time slice before the current time t;

[0036] Traverse all consecutive state pairs in the history sequence Initialize the counter ,like ,but ,like ,but ,like ,but ,like ,but Thus, the frequency of state transitions is obtained statistically.

[0037] Based on the state transition frequency results, the first-order Markov transition probability is calculated using the following formula: ,in, This represents the total number of times the previous time step was normal (0). The state transition probability matrix is ​​constructed by taking the total number of faults (1) in the previous time step. It should be noted that this matrix satisfies the row sum normalization condition: ;

[0038] Obtain node information through the real-time node monitoring module. At any moment The actual state of (i.e., time slice T-1) If the monitoring data shows that there are no fault events within that time frame, then If monitoring data shows that a fault event occurred within that time slice, then ,

[0039] according to Query transition probability matrix Output the fault risk coefficient of each node at the current moment, using the following formula: It should be noted that the failure probability of each node at the current moment, calculated using the Markov model, reflects the cumulative effect of historical failures and the current failure risk of the node. The larger the value, the higher the probability of a failure occurring at the current moment.

[0040] For example, the state sequence of a node over the past 10 time steps is as follows: (0 indicates normal, 1 indicates fault), the transition from the previous moment to the current moment is statistically analyzed. See the table below:

[0041]

[0042] The total number of times the previous time step was normal (0) was 5, including time steps 1, 2, 4, 6, and 7;

[0043] The total number of times the previous time was fault (1) was 4, including times 3, 5, 8, and 9;

[0044] The transition probabilities for each case are shown in the table below:

[0045]

[0046] Obtain the transition probability matrix ;

[0047] If the previous moment (Normal), then the probability of failure at the current moment is: If the previous moment (Fault), then the probability of a fault at the current moment is This leads to the fault risk coefficient at the current moment.

[0048] It should be noted that by learning the evolution pattern of node states through historical state sequences, real-time and quantifiable prediction of fault risks can be achieved, providing key input parameters for calculating node health.

[0049] Furthermore, the specific process of constructing a segmented node health model to determine the node status is as follows:

[0050] Based on the processor utilization, memory utilization, bandwidth utilization, disk I / O utilization, and fault risk coefficient of each node in the remote host cluster at the current moment, the overall load rate of each node at the current moment is calculated using a load rate determination model. ;

[0051] The load rate determination model is as follows: ;in, This represents the Sigmoid function, where e represents the natural constant. It is an intermediate variable used to calculate the overall load factor, consisting of a weighted sum of multiple different types of load, reflecting the node's... At any moment The comprehensive state value after taking into account various load factors. These are preset adaptive weight factors, and satisfy... Its specific size can be dynamically adjusted according to the business scenario. For example, in a computationally intensive task scenario, it can be increased. Emphasize the weight of processor load; for memory-intensive task scenarios, increase the weight. Emphasize the weight of memory load; for transmission-intensive task scenarios, increase the weight. Emphasize the weight of bandwidth load; for I / O intensive task scenarios, increase the weight of bandwidth load. Emphasize the weight of disk I / O load; in scenarios with frequent failures, increase the weight. This emphasizes the weighting of failure risk. By adapting load rates to task types and node states, the rigidity of static weights is avoided.

[0052] Based on the overall load rate of each node at the current moment, a segmented node health model is constructed to achieve a fine-grained division of node states. The segmented node health model is as follows: ;

[0053] when When this happens, the node is determined to be in a normal state and can undertake various priority tasks.

[0054] when If this happens, the node is determined to be in a sub-healthy state and will only accept low-priority tasks.

[0055] when If the node is in a critical fault state, it will stop accepting new tasks and start task migration preparation.

[0056] when If the node is in a faulty state, the fault tolerance process is immediately triggered, and the task is migrated and the node is isolated.

[0057] Step 2: Node Load Trend Prediction: By constructing an ARIMA-LSTM hybrid prediction model, the overall node load rate at future times is predicted based on historical node load rate data; the specific execution process is as follows:

[0058] 201. Obtain Time Series Sample Data: Obtain the comprehensive load rate sequence of nodes within a preset historical time period. The preset historical time period is, for example, the past hour. Where T is the total sample length. This represents the overall load factor at time t.

[0059] 202. Data Stationarity Verification and Difference Processing: The original composite load factor sequence is subjected to stationarity verification. The ADF test is used to determine whether the original load factor data is a non-stationary sequence. If the data is non-stationary, difference operations are performed until the data becomes stationary, and the optimal difference order d is recorded. * The d-th order difference formula is: ,in, For a d-order difference operator, For lag operators, For example, when d=1, .

[0060] 203. ARIMA Model Order Determination and Parameter Estimation: By using d-order differencing to bring the overall load rate to a stable state, plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the differencing series. Based on the truncation characteristics, preliminarily determine the autoregression order p and the moving average order q; for example, if PACF is truncated after order p, then p is preliminarily selected; if ACF is truncated after order q, then q is preliminarily selected.

[0061] The Akaike Information Criterion (ACI) values ​​for different (p, q) combinations are calculated using the following formula: Where N is the effective sample size. To determine the model residual variance, the order that minimizes AIC is selected as the optimal order, and ARIMA (p) is constructed. * d * q * )Model;

[0062] ARIMA (p) was evaluated using the maximum likelihood estimation (MLE) method. * d * q * The parameters of the model are estimated, including the autoregressive coefficients and moving average coefficients.

[0063] 204. Linear part prediction of ARIMA model: Input the differencing stationary sequence into the ARIMA (p) model with determined order and completed parameter estimation. * d * q* The model is used for model fitting and prediction to obtain a sequence of fitted values ​​from historical data. and future moments linear trend forecast value , This is a preset time window.

[0064] 205. Residual Sequence Calculation: After processing the load factor sequence data, the ARIMA model generates a series of residual values. These residual values ​​represent data variations that the ARIMA model failed to capture, including some non-linear trends. These residual values ​​are calculated using the formula... The fitted residual sequence of the ARIMA model over the historical period was calculated. The fitted residual sequence is used to characterize nonlinear dynamic features not captured by the linear model.

[0065] 206. Fit the residual sequence As input to the LSTM model, it outputs future time moments. Nonlinear residual prediction values ;

[0066] It should be noted that the LSTM model includes a forget gate, input gate, output gate, and memory unit, and processes sequence data through the following steps:

[0067] The forget gate controls the degree of information retention in memory cells. It outputs a value between 0 and 1 through the sigmoid activation function, representing the degree of information retention in each memory cell. The calculation formula is as follows: ,in, It is the sigmoid activation function. Used to limit the output of gate control to between 0 and 1; The input vector at the current time step is the fitted residual sequence. This is the hidden state from the previous time step. This refers to the cell state (or memory cell) at the previous time step. The output of the forget gate is used to control the degree to which information is retained in the cell state. and Here are the weight matrix and bias terms for the forget gate;

[0068] The input gate controls the degree to which new information is added to the memory cell. The sigmoid activation function outputs a value between 0 and 1, representing the importance of the new information in each memory cell. The calculation formula is as follows: ,in, The output of the input gate, and The weight matrix and bias term of the input gate are used to control the degree of new information input.

[0069] The memory unit updates its contents based on the outputs of the forget gate and the input gate, as well as the current input. The calculation formula is as follows: ,in, This represents the cell state at the current time step. This represents the cell state at the previous time step. To update the weight matrix and bias terms of the memory cells;

[0070] The output gate determines its output at the current moment based on the current input and the contents of the memory unit. The calculation formula is as follows: ,in, The output of the output gate is used to control the output at the current time step. These are the weight matrix and bias terms for the output gate;

[0071] The final output is calculated as follows: ,in, is the output of the LSTM model, and tanh is the hyperbolic tangent activation function, used to restrict the output of the memory unit to between -1 and 1.

[0072] 207. Predict future moments using the ARIMA model. linear trend forecast value LSTM model for predicting future moments Nonlinear residual prediction values By performing a weighted combination, we can obtain future moments. Load rate forecast The formula is: ,in, The weights of the ARIMA model prediction results. The weights for the prediction results of the LSTM model.

[0073] It should be noted that the weights of the ARIMA model predictions and the LSTM model predictions are determined using the inverse of the relative error method. The weight of each individual model is determined based on the magnitude of the error; the larger the relative error, the lower the prediction accuracy of the model, and the smaller the weight value. Conversely, the smaller the relative error, the larger the weight value corresponding to the individual model. The calculation formula is as follows: , The relative error of the ARIMA prediction model. This represents the relative error of the LSTM prediction model.

[0074] In the ARIMA-LSTM hybrid forecasting model, the ARIMA (Autoregressive Moving Average) model is used to capture the linear trend of the composite load rate sequence, while the LSTM (Long Short-Term Memory) model is used to capture the nonlinear fluctuation characteristics of the composite load rate sequence.

[0075] Step 3: Task Priority Sorting: Receive the set of tasks to be scheduled from the remote control terminal, calculate the priority coefficient of each task, and generate a task scheduling queue; the specific execution process is as follows:

[0076] Receive a set of tasks to be scheduled M sent by a remote control terminal, wherein the set of tasks to be scheduled M includes multiple tasks to be scheduled. Let j = 1, 2, ..., n, where n is the total number of task numbers to be scheduled. Obtain the attribute data of each task in the task set M. The attribute data includes real-time requirements, data size, and business importance. It should be noted that real-time requirements reflect the task's sensitivity to time, such as emergency control commands or normal query responses; data size reflects the scale of data to be processed or transmitted, such as KB-level small messages, MB-level files, or GB-level large data transmissions; and business importance reflects the task's criticality to the business process, such as equipment start / stop control, order processing, and log backup.

[0077] The real-time requirements, data volume, and business importance of each task to be scheduled are matched with preset real-time coefficients, data volume coefficients, and business importance coefficients corresponding to each real-time requirement, data volume, and business importance, respectively, to obtain the real-time coefficient, data volume coefficient, and business importance coefficient of each task to be scheduled. It should be noted that the higher the real-time requirement of the task to be scheduled, the larger its real-time coefficient value. The real-time coefficient value ranges from [0, 1]. For urgent tasks such as emergency control commands, the real-time coefficient value is 0.9-1.0. For ordinary tasks such as user query requests, the real-time coefficient value is usually 0.3-0.5. For low real-time tasks such as scheduled backups, the real-time coefficient value is usually 0.1-0.2. The larger the data volume of the task to be scheduled, the larger its data volume coefficient value, ranging from [0, 1]. For ultra-large data tasks such as GB-level video transmission, the data volume coefficient is 0.8-1.0; for medium data tasks such as MB-level file upload, the data volume coefficient is 0.3-0.5; and for small data tasks such as KB-level instructions, the data volume coefficient is 0.1-0.2. The more core and important the business of the task to be scheduled, the larger its business importance coefficient, ranging from 0 to 1. For core businesses such as device start / stop control and payment transactions, the business importance coefficient is usually 0.8-1.0, and business failure will lead to serious consequences. For ordinary businesses such as user information query, the business importance coefficient is usually 0.4-0.6, which is important but not core. For non-core business tasks such as log upload and monitoring data collection, the business importance coefficient is usually 0.1-0.3, and can tolerate delay scheduling.

[0078] Computation model based on priority coupling The priority coefficient of each task to be scheduled is calculated. Where tanh is the hyperbolic tangent function. The priority potential value of a task is an intermediate calculation variable, composed of a linear weighted sum of three factors: the real-time nature of the task to be scheduled, the amount of data, and the importance of the business. Tasks to be scheduled The real-time performance coefficient, The data volume coefficient for the task to be scheduled. The business importance coefficient of the task to be scheduled. These are preset weighting factors, satisfying It can be dynamically adjusted according to actual business needs. For example, It should be noted that the linear weighted coupling formula... The priority potential value of a task is calculated, and then the linear weighting result is nonlinearly mapped using the hyperbolic tangent function to amplify the distinguishability of high-priority tasks, avoid the priority ambiguity caused by linear weighting, and make the priority coefficient of high-priority tasks significantly prominent. It is easier to distinguish core tasks from ordinary tasks during scheduling. The larger the real-time coefficient, data volume coefficient, and business importance coefficient of the task to be scheduled, the larger the priority coefficient of the task to be scheduled.

[0079] All tasks in the set of tasks to be scheduled are sorted from highest to lowest priority to form a task scheduling queue. This ensures that high-priority tasks (such as emergency control instructions and core business tasks) are placed at the front of the queue, thus prioritizing node allocation.

[0080] Step 4, Task-Node Scheduling Decision: A task-node bipartite graph is constructed by selecting available nodes. Preliminary edge matching degrees are calculated based on node health, predicted load rate, and task priority. Feature vectors of tasks, nodes, and edges are injected, and embedded vectors are generated through iterative training using a GNN. Based on resource constraints, the global optimal match is solved using cosine similarity calculation and the Hungarian algorithm. Finally, standardized scheduling instructions are generated and issued to achieve optimal matching between tasks and nodes, balancing low-latency execution of high-priority tasks with cluster load balancing. The specific execution process is as follows:

[0081] 401. Construct a task-node bipartite graph: Extract all nodes in the current normal state and sub-healthy state, mark them as available nodes, and construct a set of available nodes;

[0082] Define the set of tasks to be scheduled as the set of the left vertices of a bipartite graph. Where j1 is the number of tasks to be scheduled; the set of available nodes is defined as the set of right vertices. , where r is the number of available nodes;

[0083] For each task With available nodes Establish edges This forms a complete bipartite graph structure, representing the task. It is possible at the node Execute above;

[0084] 402. Preliminary matching degree calculation: For each edge Based on the current health of available nodes Predicted load rate of available nodes at future times and task priority coefficient Through formula Calculate the initial matching degree And the initial matching degree is used as the initial weight of the edge, where, These are respectively represented as preset scaling factors, for example, .

[0085] 403. Graph Neural Network Embedding Learning: For each task vertex in the left vertex set TA of a bipartite graph Constructing eigenvectors Among them, CPU requirement and memory requirement are the actual resource requirements of the task, and deadline is the preset completion time of the task.

[0086] For each vertex in the right vertex set NA of the bipartite graph Construct feature vectors Among them, remaining CPU and remaining memory are the node's current available resource parameters, and historical response time is the node's average response time for processing similar tasks in the past;

[0087] For each edge in the bipartite graph Construct feature vectors ,in, The initial matching degree of the edges obtained in step 402;

[0088] GNN is iteratively trained on the bipartite graph. In each iteration, for each task vertex... Aggregate the feature information of all connected nodes, based on the edges Initial weights (preliminary matching degree) For connected vertices eigenvectors Perform weighted summation to generate a task embedding vector that incorporates the interaction information of neighboring nodes. For each node vertex According to all its connected edges The initial weights for connected task vertices eigenvectors Perform weighted summation to generate node embedding vectors that incorporate neighbor task interaction information. ;

[0089] A matching loss function is constructed to measure the difference between the actual and ideal matching between the task and the node. This loss function is minimized using the backpropagation algorithm to drive the task embedding vector. With node embedding vector The embedding vector is iterated and updated until the loss function converges. At this point, the embedding vector can accurately capture the hidden association features between the task and the node.

[0090] 404. Global Optimal Matching and Scheduling: Embedding Vector for Each Task With each node embedding vector Using the cosine similarity algorithm Calculate the final matching score between the two. ,in, This represents the dot product of the task embedding vector and the node embedding vector. Represents task embedding vector The model, Represents node embedding vector The model, The value range is [0, 1]. The larger the value, the higher the matching degree between the task and the node.

[0091] 405. Constrained Bipartite Graph Matching: Set resource capacity constraints, i.e., the tasks to be assigned. CPU requirements ≤ corresponding node Remaining CPU, tasks Memory requirements ≤ corresponding node The remaining memory is used only for task-node combinations that satisfy this constraint to participate in the optimal matching solution;

[0092] Based on the final match score Using edge weights, the Hungarian algorithm is used to solve the global optimal matching of a bipartite graph with resource constraints, generating a task-node allocation scheme that maximizes the total matching score of tasks and nodes while satisfying resource constraints.

[0093] 406. Issuance of scheduling instructions: The task-node allocation scheme is converted into standardized scheduling instructions, which are then issued to the corresponding nodes through a distributed scheduling system. The scheduling instructions include core parameters such as task identifier, target node identifier, node execution port, number of allocated CPU cores, and size of allocated memory.

[0094] After scheduling is completed, the process returns to step one to update the real-time load rate of each node in the remote host cluster for the next round of node status awareness. At the same time, the task allocation results are fed back to the remote control terminal for users to monitor and maintain in real time.

[0095] Step 5, Proactive Fault Tolerance: Detect faulty nodes, migrate incomplete tasks, isolate faulty nodes, and perform recovery checks; the specific execution process is as follows:

[0096] When a node is in a critical or fault state, a fault warning signal is triggered. To avoid misjudgment caused by abnormal parameter acquisition or network fluctuations, three high-frequency heartbeat packets are sent to the target node consecutively. If no valid response is received for three consecutive times, the node is determined to be a real fault. If a valid response is received, it is determined to be a temporary parameter fluctuation, and only a warning is triggered without performing subsequent fault-tolerant operations.

[0097] Once a node is confirmed to be a real failure, a task pause command is immediately sent to that node to suspend the execution of all unfinished tasks on that node. At the same time, the task snapshot generation program is triggered to collect and store the complete status information of each task to be migrated, including task execution progress, context data, temporary cache, business importance tags and real-time requirements, to form a task breakpoint snapshot.

[0098] The breakpoint resume procedure transmits the breakpoint snapshots of each task to be migrated to the corresponding target migration node one by one, and monitors the transmission progress and integrity in real time during the transmission process. After receiving the breakpoint snapshots, the target migration node automatically loads the task context and execution progress, resumes task execution, and ensures that the task is not interrupted and the data is not lost. It should be noted that the target migration node is determined by the method in step four.

[0099] Remove and isolate critical fault nodes and faulty nodes from the remote host cluster node set N, mark them as unavailable nodes, and prohibit new tasks from being assigned to these nodes to avoid affecting operational stability;

[0100] The isolated nodes are monitored for status. When a node health level of ≥0.3 is detected, the node is determined to have recovered and is added back to the remote host cluster node set N, marked as recovered, and gradually takes over low-priority tasks. Normal scheduling is resumed after confirming that there are no abnormalities.

[0101] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A host scheduling node method based on a computer remote control system, characterized in that, Includes the following steps: Step 1: Collect the operating parameters of each node in the remote host cluster at the current moment, calculate the evaluation value of the operating parameters of the node at the current moment, and predict the fault risk coefficient of each node at the current moment through Markov chain based on the historical fault record data of each node, and construct a segmented node health model to determine the node status. Step 2: By constructing an ARIMA-LSTM hybrid prediction model, predict the node's overall load rate at future times based on historical node overall load rate data; Step 3: Receive the set of tasks to be scheduled sent by the remote control terminal, calculate the priority coefficient of each task and generate a task scheduling queue; Step 4: Construct a task-node bipartite graph. Calculate the initial matching degree of edges based on node health, predicted load rate, and task priority. After injecting the feature vectors of tasks, nodes, and edges, generate embedding vectors through GNN iterative training. Combine resource constraints with cosine similarity calculation and Hungarian algorithm to solve for the global optimal matching. Finally, generate and issue scheduling instructions. Step 5: Detect faulty nodes, migrate incomplete tasks, isolate faulty nodes, and perform recovery tests.

2. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, The evaluation values ​​of the computing node's operating parameters at the current moment are as follows: Obtain the set of nodes N in the remote host cluster, as well as the processor utilization, memory utilization, bandwidth utilization, and disk I / O utilization data of each node in the remote host cluster at the current moment; The standard deviation and mean of the processor utilization data of each node at the current moment are calculated to generate the standard deviation and mean of the processor utilization of each node at the current moment; the processor utilization evaluation value, memory utilization evaluation value, bandwidth utilization evaluation value and disk I / O utilization evaluation value of each node at the current moment are calculated by the coefficient of variation algorithm.

3. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, Based on the historical failure records of each node, the failure risk coefficient of each node at the current moment is predicted using a Markov chain, as follows: Extract historical fault record datasets of nodes from the node monitoring database; set the time slice length and divide the historical timeline into continuous time slices; For each time slice, define a binary state variable and generate a sequence of historical states for the node; Traverse all consecutive state pairs in the historical sequence, initialize the counter, and then obtain the state transition frequency results. Calculate the first-order Markov transition probability based on the state transition frequency results; The actual status of a node at any given time is obtained through the real-time node monitoring module: if the monitoring data shows no fault events within that time slice, the actual status value is zero; if the monitoring data shows fault events within that time slice, the actual status value is one. Based on the actual state, query the transition probability matrix and output the failure risk coefficient of each node at the current moment.

4. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, Based on the processor utilization, memory utilization, bandwidth utilization, disk I / O utilization, and fault risk coefficient of each node in the remote host cluster at the current moment, the comprehensive load rate of each node at the current moment is calculated through the load rate determination model. Based on the overall load rate of each node at the current moment, a segmented node health model is constructed to obtain the node health score. ;when When, the node is determined to be in a normal state; when When, the node is determined to be in a sub-healthy state; when When the node is in a critical fault state, it is determined to be in a critical fault state; when If the node is in a faulty state, then the node is determined to be in a faulty state.

5. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, The process of predicting the overall node load rate at future moments based on historical node load rate data is as follows:

201. Obtain the comprehensive load rate sequence of the node within the preset historical time period; 202. Perform a stationarity check on the original comprehensive load rate sequence to determine whether the original load rate data is a non-stationary sequence. If the data is non-stationary, perform a difference operation until the data becomes stationary, and record the optimal difference order.

203. By using d-order differencing to bring the overall load rate to a stable state, plot the autocorrelation function and partial autocorrelation function of the differencing sequence, and preliminarily determine the autoregression order p and the moving average order q based on the truncation characteristics. Calculate the Akaike information content criterion values ​​for different (p, q) combinations, select the one that minimizes the Akaike information content criterion value as the optimal order, and construct the ARIMA(p) function. * d * q * )Model; ARIMA(p) was analyzed using the maximum likelihood estimation method. * d * q * The parameters of the model are estimated, including autoregressive coefficients and moving average coefficients; 204. Input the differencing stationary sequence into an ARIMA algorithm that has been determined in order and whose parameters have been estimated. * d * q * The model is used to fit and predict historical data, resulting in a sequence of fitted values ​​and a linear trend prediction for future time periods.

205. Calculate the fitting residual sequence of the ARIMA model over the historical period; 206. Use the fitted residual sequence as input to the LSTM model and output the nonlinear residual prediction values ​​for future time steps.

207. The load rate prediction for future time is obtained by weighting and combining the linear trend prediction value of the ARIMA model and the nonlinear residual prediction value of the LSTM model.

6. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, The priority coefficient of each task is calculated and a task scheduling queue is generated. The specific execution process is as follows: Receive a set of tasks to be scheduled sent by a remote control terminal, the set of tasks to be scheduled includes multiple tasks to be scheduled, and obtain the attribute data of each task to be scheduled in the set of tasks to be scheduled; The real-time requirements, data volume, and business importance of each task to be scheduled are matched with the preset real-time coefficients, data volume coefficients, and business importance coefficients corresponding to each real-time requirement, data volume, and business importance, respectively, to obtain the real-time coefficient, data volume coefficient, and business importance coefficient of each task to be scheduled. The priority coefficient of each task to be scheduled is calculated by the priority coupling calculation model. All tasks in the set of tasks to be scheduled are sorted from largest to smallest according to their priority coefficients to form a task scheduling queue.

7. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, The specific execution process for step four is as follows:

401. Extract all nodes in normal and sub-healthy states at the current moment and mark them as available nodes to construct a set of available nodes; define the set of tasks to be scheduled as the left vertex set of a bipartite graph; define the set of available nodes as the right vertex set; establish an edge between each task and an available node to form a complete bipartite graph structure; 402. For each edge, based on the current health of available nodes, the predicted load rate of available nodes in the future, and the task priority coefficient, the preliminary matching degree is calculated by formula, and the preliminary matching degree is used as the initial weight of the edge.

403. Construct a feature vector for each task vertex in the left vertex set of the bipartite graph; construct a feature vector for each node vertex in the right vertex set of the bipartite graph; then construct a feature vector for each edge in the bipartite graph. The bipartite graph is trained iteratively using GNN. In each iteration, for each task vertex, the feature information of all connected nodes is aggregated. The feature vectors of connected nodes are weighted and summed according to the initial weight of the edges to generate a task embedding vector that integrates the interaction information of neighboring nodes. For each node vertex, the feature vectors of the connected task vertices are weighted and summed according to the initial weights of all its connected edges to generate a node embedding vector that incorporates the interaction information of neighboring tasks. Construct a matching loss function, minimize the matching loss function through the backpropagation algorithm, and drive the task embedding vector and node embedding vector to update iteratively until the matching loss function converges.

404. For each task embedding vector and each node embedding vector, calculate the final matching score between them using the cosine similarity algorithm; 405. Set resource capacity constraints, and only task-node combinations that meet the constraints participate in the optimal matching solution; use the final matching score as the edge weight, and use the Hungarian algorithm to solve the global optimal matching of the bipartite graph with resource constraints to generate a task-node allocation scheme; 406. Convert the task-node allocation scheme into scheduling instructions and send the scheduling instructions to the corresponding nodes. After the scheduling is completed, re-enter step one to update the real-time load rate of each node in the remote host cluster for the next round of node status awareness, and at the same time feed back the task allocation results to the remote control terminal.

8. The host scheduling node method based on a computer remote control system according to claim 1, characterized in that, The specific execution process for step five is as follows: When a node is in a critical or faulty state, a fault warning signal is triggered, and three high-frequency heartbeat packets are sent to the target node consecutively. If no valid response is received after three consecutive attempts, the node is determined to be a real fault. If a valid response is received, it is determined to be a temporary parameter fluctuation, and only a warning is triggered without performing subsequent fault-tolerant operations. Once a node is confirmed to be a real failure, a task pause command is immediately sent to that node to pause the execution of all unfinished tasks on that node. At the same time, the task snapshot generation program is triggered to collect and store the complete status information of each task to be migrated, forming a task breakpoint snapshot. The breakpoint resume program transmits the breakpoint snapshots of each task to be migrated to the corresponding target migration node one by one. After receiving the breakpoint snapshots, the target migration node automatically loads the task context and execution progress and resumes task execution. Remove critical fault nodes and faulty nodes from the remote host cluster node set N, mark them as unavailable nodes, and prohibit new tasks from being assigned to these nodes; The isolated nodes are monitored for status. When a node health level of ≥0.3 is detected, the node is determined to have recovered and is added back to the remote host cluster node set, marked as recovered, and gradually takes over low-priority tasks. Normal scheduling is resumed after confirming that there are no abnormalities.

Images