Host load prediction method and device

Abstract

The invention relates to the technical field of cloud computing, and particularly provides a host load prediction method and device, and the method comprises the following steps: S1, collecting historical workload data based on time sequence decomposition and a neural network; s2, performing standardization processing on the historical workload data; s3, decomposing the standardized time sequence to obtain a trend term and a residual term; s4, respectively carrying out neural network modeling on the trend sequence and the residual sequence; and S5, outputting a host load prediction result. Compared with the prior art, the method has the advantages that the load of the host can be accurately predicted in real time, and resource management of the cloud computing system is helped to automatically allocate resources to adapt to the change of the workload of each server.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing, and specifically provides a host load prediction method and device. Background Art

[0002] One of the main characteristics of cloud computing systems is elasticity, that is, the resource management system can automatically allocate resources to adapt to changes in the workload of each server. By reasonably predicting the host load, cloud service providers can prevent potential undersupply when scheduling service resources, avoid performance degradation, increased latency, or system crashes caused by excessive host load, ensure service continuity and stability, reduce the risk of violating the service level agreement (SLA), and improve user experience; on the other hand, they can deal with oversupply in a timely manner and reduce unnecessary hardware investment and maintenance costs.

[0003] As a key technology for cluster resource management, the load prediction method collects data on resource information such as CPU and memory from each server at regular intervals under the premise of normal server operation. By analyzing the historical data collected by the cloud data center, the trend and change rules of the load data are mastered, so as to predict the load value of the next cycle. The load prediction of cloud computing resources is a typical time series prediction problem, and establishing an accurate model is the focus of research work.

[0004] In the research of related fields, the solutions for host load prediction have experienced a transformation from traditional statistical methods to machine learning technology, and then to deep learning. For example, the patent "A host load prediction method based on long short-term memory network" (CN106502799A) adopts LSTM (Long Short Term Memory) network; the patent "A host load prediction method in cloud environment" (CN108196957A) adopts ARMA (Auto-Regressive Moving Average) model; the patent "Cloud computing host load prediction method combining attention mechanism and gated recurrent unit" (CN113076196A) combines attention mechanism and GRU (Gate Recurrent Unit).

[0005] At the same time, as the sequence modeling architecture Transformer shines in various natural language processing tasks, Transformer-based time series solutions are also surging. For example, the paper "Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting, NeurIPS 2019" proposes the LogTrans model, the paper "Informer: Beyond efficient transformer for long sequence time-series forecasting, AAAI 2021" proposes the Informer model, the paper "Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting, NeurIPS2021" proposes the Autoformer model, the paper "Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting, ICLR 2022Oral" proposes the Pyraformer model, and the paper "Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting, ICML 2022" proposes the FEDformer model.

[0006] Although the Transformer-based time series model has made up for the shortcomings of traditional methods in capturing long-distance time dependencies, processing high-dimensional data, and combating noise to a certain extent, it still has the following three problems in the application of load prediction of cloud computing resources:

[0007] (1) The computational complexity of attention is high, and KVcache occupies a large amount of memory, resulting in high training and deployment costs of the model;

[0008] (2) The reasoning process of Transformer is a single-step prediction. This IMS (Iterated Multi-step) method is prone to error accumulation, thus affecting the prediction accuracy of the model;

[0009] (3) Although Transformer uses positional encoding to retain some sorting information, the attention mechanism will inevitably lead to the loss of some temporal information, which will affect the prediction accuracy of the model and further affect the cloud service provider's evaluation of resource scheduling tasks, bringing serious business decision-making risks. Summary of the invention

[0010] The present invention aims at solving the above-mentioned deficiencies of the prior art and provides a host load prediction method with strong practicability.

[0011] A further technical task of the present invention is to provide a host load prediction device that is reasonably designed, safe and applicable.

[0012] The technical solution adopted by the present invention to solve its technical problem is:

[0013] A host load prediction method based on time series decomposition and neural network has the following steps:

[0014] S1. Collect historical workload data;

[0015] S2, standardize historical workload data;

[0016] S3, decomposing the standardized time series to obtain trend terms and residual terms;

[0017] S4, neural network modeling is performed on the trend sequence and the residual sequence respectively;

[0018] S5. Output the host load prediction result.

[0019] Further, in step S1, the resource monitoring system of the cloud computing service cluster center obtains historical workload data with a lookback window L, totaling C dimensions, and the historical workload data constitutes a multivariate time series sample set with a length of L in, Represents the host load of the i-th dimension at time t.

[0020] Furthermore, in step S2, the values ​​of historical load data may vary greatly in different time intervals, and the original data needs to be standardized and preprocessed. The preprocessed data accelerates the convergence of the subsequent deep learning algorithm training process. The original sequence is processed using Z-score standardization, and the processed data meets the mean value of 0 and the standard deviation of 1.

[0021] Furthermore, the Z-score standardization calculation formula is as follows:

[0022]

[0023] Where, t=1,2,…,L, i=1,2,,…C, is the i-th dimension time series The sample mean of is the i-th dimension time series The sample standard deviation of .

[0024] Furthermore, in step S3, the trend term of the time series is calculated using the m-order moving average method, where m is a hyperparameter value that needs to be set manually, and the i-th dimension time series The trend items are:

[0025]

[0026] Where m = 2k + 1, that is, the estimated value of the trend term at time point t It is obtained by averaging the k distances at time t. The average eliminates some randomness in the data, thus obtaining a smoother trend estimate.

[0027] Further calculate the residual term of the time series:

[0028]

[0029] Further, in step S4, the trend sequence and the remaining sequence Conduct neural network modeling, and is the input of the neural network, and the output is obtained through weight allocation, linear combination and nonlinear activation:

[0030]

[0031] in, are the weight and bias parameters of the trend neural network, are the weight and bias parameters of the remaining neural network, both obtained through training, and σ(·) is a nonlinear activation function.

[0032] Further, in step S5, the output O of the trend neural network T and the output of the remaining neural network O R Add together to get the host load value for the next T time steps:

[0033]

[0034] A host load prediction device comprises: at least one memory and at least one processor;

[0035] The at least one memory is used to store a machine-readable program;

[0036] The at least one processor is used to call the machine-readable program to execute a host load prediction method.

[0037] Compared with the prior art, the host load prediction method and device of the present invention have the following outstanding beneficial effects:

[0038] The present invention accelerates the convergence of deep learning algorithm training through standardized preprocessing modules. On the one hand, it solves the shortcomings of traditional methods in capturing long-term dependencies, processing high-dimensional data and combating noise. On the other hand, it avoids the high cost of training and deployment of Transformer-based time series models, error accumulation caused by single-step prediction, and partial time information loss. It optimizes the prediction model and improves the prediction effect, reducing the business decision-making risks caused by insufficient model prediction accuracy.

[0039] The method of the present invention can predict the host load in real time and accurately, and help the resource management of the cloud computing system to automatically allocate resources to adapt to the changes in the workload of each server. On the one hand, it can prevent potential undersupply and avoid problems such as performance degradation, increased latency or system crash caused by excessive host load, ensure service continuity and stability, reduce the risk of violating service level agreements, and improve user experience; on the other hand, it can deal with oversupply in a timely manner and reduce unnecessary hardware investment and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Attached Figure 1 is a flow chart of a host load prediction method;

[0042] Attached Figure 2 It is a schematic diagram of time series decomposition in a host load prediction method;

[0043] Attached Figure 3 It is a schematic diagram of a single-layer neural network model in a host load prediction method. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] A best embodiment is given below:

[0046] like Figure 1 As shown, a host load prediction method in this embodiment has the following steps:

[0047] S1. Collect historical workload data;

[0048] In this implementation, the historical workload data with a lookback window L is obtained through the resource monitoring system of the cloud computing service cluster center, including CPU load sequence, memory load sequence and disk I / O load sequence, totaling C dimensions.

[0049] The historical data constitutes a multivariate time series sample set of length L in represents the host load of the i-th dimension at time t. The goal of this method is to predict the host load value in the future T time steps based on historical workload data.

[0050] S2, standardize historical workload data;

[0051] The values ​​of historical load data vary greatly in different time intervals, so the original data needs to be standardized and preprocessed. The preprocessed data can accelerate the convergence of the subsequent deep learning algorithm training process. In this step, the original sequence is processed using Z-score standardization, and the processed data meets the average value of 0 and the standard deviation of 1. The Z-score standardization calculation formula is as follows:

[0052]

[0053] Where, t=1,2,…,L, i=1,2,,…C, is the i-th dimension time series The sample mean of is the i-th dimension time series The sample standard deviation of .

[0054] S3, decomposing the standardized time series to obtain trend terms and residual terms;

[0055] The trend component of a time series represents the persistent, long-term change in the mean of the series and represents the importance of the largest time scale.

[0056] In this implementation, the m-th order moving average method is used to calculate the trend term of the time series, where m is a hyperparameter value that needs to be set manually. The trend items are:

[0057]

[0058] Where m = 2k + 1. That is, the estimated value of the trend term at time point t is It is obtained by averaging the k distances at time t. The average value eliminates some randomness in the data, so that we can get a smoother estimate of the trend term. Based on this, the residual term of the time series can be further calculated:

[0059]

[0060] Attached Figure 2 This is a schematic diagram of the decomposition of a standardized time series. However, the present invention is not limited to the decomposition method disclosed above. Those skilled in the art may also use methods such as a linear trend model with change points (Linear Trend with Changepoints) and a nonlinear saturating growth model (Nonlinear Saturating Growth) to decompose trend terms and residual terms.

[0061] S4, neural network modeling is performed on the trend sequence and the residual sequence respectively;

[0062] For trend series and the remaining sequence Conduct neural network modeling. Figure 3 Take the single-layer neural network model shown in the figure as an example. and is the input of the neural network, and the output is obtained through weight allocation, linear combination and nonlinear activation:

[0063]

[0064] in, are the weight and bias parameters of the trend neural network, are the weight and bias parameters of the remaining neural network, both obtained through training, and σ(·) is a nonlinear activation function.

[0065] S5, host load prediction result output;

[0066] Output of the trend neural network O Tand the output of the remaining neural network O R Add together to get the host load value for the next T time steps:

[0067]

[0068] The present invention directly outputs the host load value at the future T time steps, belongs to the DMS (direct multi-step) method, and will not cause error accumulation due to multiple iterations.

[0069] Based on the above method, a host load prediction device in this embodiment includes: at least one memory and at least one processor;

[0070] The at least one memory is used to store a machine-readable program;

[0071] The at least one processor is used to call the machine-readable program to execute a host load prediction method.

[0072] The above-mentioned specific implementations are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementations. Any technical solutions that conform to the above-mentioned specific implementations of the present invention and any appropriate changes or substitutions made by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0073] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A host load prediction method, characterized in that: Based on time series decomposition and neural network, the steps are as follows: S1. Collect historical workload data; S2, standardize historical workload data; S3, decomposing the standardized time series to obtain trend terms and residual terms; S4, neural network modeling is performed on the trend sequence and the residual sequence respectively; S5. Output the host load prediction result.

2. A host load prediction method according to claim 1, characterized in that: In step S1, the resource monitoring system of the cloud computing service cluster center obtains historical workload data with a lookback window L, totaling C dimensions. The historical workload data constitutes a multivariate time series sample set with a length of L. in, Represents the host load of the i-th dimension at time t.

3. A host load prediction method according to claim 2, characterized in that: In step S2, the values ​​of historical load data may vary greatly in different time intervals, and the original data needs to be standardized and preprocessed. The preprocessed data accelerates the convergence of the subsequent deep learning algorithm training process. The original sequence is processed using Z-score standardization, and the processed data meets the mean value of 0 and the standard deviation of 1.

4. A host load prediction method according to claim 3, characterized in that: The Z-score standardization calculation formula is as follows: Where, t=1,2,…,L, i=1,2,,…C, is the i-th dimension time series The sample mean of is the i-th dimension time series The sample standard deviation of .

5. A host load prediction method according to claim 4, characterized in that: In step S3, the trend term of the time series is calculated using the m-order moving average method, where m is a hyperparameter value that needs to be set manually, and the i-th dimension time series The trend items are: Where m = 2k + 1, that is, the estimated value of the trend term at time point t It is obtained by averaging the k distances at time t. The average eliminates some randomness in the data, thus obtaining a smoother trend estimate. Further calculate the residual term of the time series:

6. A host load prediction method according to claim 5, characterized in that: In step S4, the trend sequence and the remaining sequence Conduct neural network modeling, and is the input of the neural network, and the output is obtained through weight allocation, linear combination and nonlinear activation: in, are the weight and bias parameters of the trend neural network, are the weight and bias parameters of the remaining neural network, both obtained through training, and σ(·) is a nonlinear activation function.

7. A host load prediction method according to claim 6, characterized in that: In step S5, the output O of the trend neural network is T and the output of the remaining neural network O R Add together to get the host load value for the next T time steps:

8. A host load prediction device, characterized in that: include: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 7.

Images