Cluster load prediction method and system based on clustering and attention mechanism
A technology of load prediction and attention, which is applied in the direction of program control design, computer parts, character and pattern recognition, etc., can solve the problems of high consumption of computing resources and unsatisfactory prediction accuracy, so as to avoid high consumption of computing resources and improve Resource utilization and the effect of improving accuracy
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0052] This embodiment proposes a cluster load prediction method based on clustering and attention mechanism, see Figure 1-Figure 5 , including the following steps:
[0053] S1: Obtain the discrete time series of the computer cluster, and divide the discrete time series into a training set and a test set; in this embodiment, the method of interval sampling is used to obtain the discrete time series X of the computer cluster whose length is t 1 ={x 11 ,x 12 ,...,x 1t},X 2 ={x 21 ,x 22 ,...,x 2t},...,X n ={x n1 ,x n2 ,...,x nt}, where n is the number of discrete time series.
[0054] S2: Divide the discrete time series of the computer cluster into k categories through a clustering algorithm;
[0055] S3: Based on the attention mechanism, set k pairs of encoders and decoders of this class, as well as a global encoder; input all discrete time series of k types into the global encoder for training, and the encoding of each class of this class The encoder uses the disc...
Embodiment 2
[0078] In this embodiment, a cluster load prediction system based on clustering and attention mechanism is proposed, which is applied to a cluster load prediction method based on clustering and attention mechanism proposed in the above-mentioned embodiment 1, see Figure 6 , which includes:
[0079] The sequence acquisition module is used to obtain the discrete time series of the computer cluster, and divide the discrete time series into a training set and a test set;
[0080] A clustering module is used to divide the discrete time series of the computer cluster in the sequence acquisition module into k classes;
[0081] The training module is used to set k pairs of encoders and decoders of this class and a global encoder based on the attention mechanism; input all discrete time sequences of the k classes into the global encoder for training, and each class The encoder of this class uses the discrete time series in this class for training; after the training of the global enc...
Embodiment 3
[0087] This embodiment adopts a cluster CPU utilization data set.
[0088] The data set samples a total of 888 CPU load time series with a time span of one week by sampling every 30 seconds. Through the clustering algorithm, these 888 time series are divided into 10 classes, and the number of time series owned by each class is not equal. The time series is divided into training set, verification set and test set according to 5:1:1.
[0089] In this embodiment, it is assumed that the training step size t, the decoder heuristic sequence step size t′, and the prediction step size l are used to generate time series data by means of a sliding window. For a time series X={x 0 ,x 1 ,...,x t+l}, the input to the encoder is X E ={x 0 ,x 1 ,...,x t}, the decoder input is X D ={x t-t′+1 ,x t-t′+2 ,...,x t ,0,0,...0}. The number of 0s is the prediction step size l. Model output Z 预测 ={x' t+1 ,x′ t+2 ,…,x′ t+l}, target value Z 目标 ={x t+1 ,x t+2 ,...,x t+l}. Where t' ...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 
PatSnap Eureka turns technology decisions into work you can execute. Powered by our Innovation Knowledge Graph, it runs expert workflows across engineering, life sciences, materials and intellectual property. Get your review-ready output in minutes.
