CN113220450B - Load prediction method, resource scheduling method and device for cloud-side multi-data center - Google Patents
Load prediction method, resource scheduling method and device for cloud-side multi-data center Download PDFInfo
- Publication number
- CN113220450B CN113220450B CN202110473131.1A CN202110473131A CN113220450B CN 113220450 B CN113220450 B CN 113220450B CN 202110473131 A CN202110473131 A CN 202110473131A CN 113220450 B CN113220450 B CN 113220450B
- Authority
- CN
- China
- Prior art keywords
- load
- input
- load prediction
- representing
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 78
- 239000013598 vector Substances 0.000 claims abstract description 91
- 230000008569 process Effects 0.000 claims abstract description 33
- 238000003062 neural network model Methods 0.000 claims abstract description 25
- 230000006870 function Effects 0.000 claims description 38
- 238000004364 calculation method Methods 0.000 claims description 28
- 230000004913 activation Effects 0.000 claims description 19
- 230000008859 change Effects 0.000 claims description 12
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 238000013507 mapping Methods 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 11
- 230000007246 mechanism Effects 0.000 claims description 10
- 230000003044 adaptive effect Effects 0.000 claims description 5
- 238000012549 training Methods 0.000 claims description 5
- 230000005540 biological transmission Effects 0.000 claims description 3
- 238000009795 derivation Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000013215 result calculation Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 230000005284 excitation Effects 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 10
- 238000004590 computer program Methods 0.000 description 9
- 238000004458 analytical method Methods 0.000 description 6
- 238000009826 distribution Methods 0.000 description 6
- 238000013461 design Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000007781 pre-processing Methods 0.000 description 3
- 238000013468 resource allocation Methods 0.000 description 3
- 238000003860 storage Methods 0.000 description 3
- UFHFLCQGNIYNRP-UHFFFAOYSA-N Hydrogen Chemical compound [H][H] UFHFLCQGNIYNRP-UHFFFAOYSA-N 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 229910052739 hydrogen Inorganic materials 0.000 description 2
- 239000001257 hydrogen Substances 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000004134 energy conservation Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
- G06F9/505—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals considering the load
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5077—Logical partitioning of resources; Management or configuration of virtualized resources
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a load prediction method facing cloud multi-data center, comprising the following processes: acquiring a log record file for recording the resource use condition of the virtual machine at each time point, and extracting required characteristic quantity data and historical load data from the log record file; converting the characteristic quantity data and the historical load data into corresponding input characteristic sequences and historical load vectors; calculating to obtain a nonlinear component of load prediction by utilizing a pre-constructed neural network model; calculating to obtain a linear component of load prediction by utilizing a pre-constructed autoregressive model; and integrating the nonlinear component and the linear component of the load prediction to obtain a final load prediction result. The method comprehensively considers the linear trend and the nonlinear characteristic of the load sequence changing along with time under the environment of multiple data centers, combines the neural network model with the statistical learning method of the autoregressive model, and can effectively improve the future load prediction precision.
Description
Technical Field
The invention particularly relates to a load prediction method facing cloud multi-data centers, and further relates to a resource scheduling method facing cloud multi-data centers, and belongs to the technical field of cloud computing and data mining.
Background
The continuous increase of the scale of the cloud data center is promoted by the continuous increase of the computing demand of the cloud computing technology, and the business market scale of the domestic data center is expected to increase to 3200 hundred million yuan by the end of 2022 years, and the scale structure of the domestic data center is also changed from a single data center to a cloud multi-data center. In order to realize the development of green energy conservation, the data center needs to have decision-making capability of dynamically adjusting the internal resource allocation, and integrate the computing resources to realize service flexibility. Effective load prediction is a precondition for realizing flexible resource allocation, and can provide decision reference for a high-quality extension scheme. Therefore, based on the current cloud multi-data center environment, establishing a proper load prediction model by combining load characteristics has important significance.
Workload is a time-series problem that is closely related to context, however most studies of load prediction have stagnated within a single data center. Compared with the traditional single data center, the multi-data center resource scheduling expansion capability is stronger, the service deployment is more flexible, and the load change is more obviously influenced by the user behavior drive. The computing resource requirements needed by the cloud-oriented data centers are dynamically changed, and the load fluctuation trends of different data centers are greatly different, so that the cloud-oriented data center-oriented load prediction needs to be combined with the specific situation of each data center for modeling analysis, and cannot be based on the original algorithm oriented to a single data center.
In view of the above, it is necessary to provide a cloud-oriented load prediction method and a resource distributor system for cloud-oriented multi-data centers to solve the above problems.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, and provides a cloud-side multi-data-center-oriented load prediction method, which combines a neural network model and an autoregressive model to predict the non-linear component and the linear component of future load change on a server, so that the accuracy of a load prediction result is improved.
In order to solve the technical problems, the technical scheme of the invention is as follows.
In a first aspect, the present invention provides a load prediction method for cloud-side multiple data centers, including the following processes:
acquiring a log record file for recording the resource use condition of the virtual machine at each time point, and extracting required characteristic quantity data and historical load data from the log record file; converting the characteristic quantity data and the historical load data into corresponding input characteristic sequences and historical load vectors;
calculating to obtain a nonlinear component of load prediction by utilizing a pre-constructed neural network model based on the obtained input characteristic sequence and the historical load vector;
calculating to obtain a linear component of load prediction by utilizing a pre-constructed autoregressive model based on the obtained historical load vector;
and integrating the nonlinear component and the linear component of the load prediction to obtain a final load prediction result.
Optionally, the characteristic quantity includes: sampling and recording time, virtual machine kernel number configuration, CPU capacity, virtual machine memory configuration capacity, virtual machine memory effective usage, disk reading throughput, disk writing throughput, network receiving throughput and network transmission throughput; the load refers to the effective use amount of the CPU.
Optionally, the converting the feature quantity data into a corresponding input feature sequence includes:
and segmenting the characteristic quantity data by using a sliding window to form a time sequence with fixed time step length as an input characteristic sequence.
Optionally, the neural network model includes: an encoder, a decoder and a multi-layer perceptron network; the input of the encoder is an acquired input feature sequence, the input of the decoder is an input feature sequence which is output by an upper encoder and extracted in a self-adaptive mode, and the input of the multilayer perceptron network is a text vector output by the decoder.
Optionally, the obtaining, by using a pre-constructed neural network model, a nonlinear component of load prediction based on the obtained input feature sequence and the historical load vector includes:
the encoder module comprises an input attention layer, a softmax layer and an LSTM neural network layer, wherein the output of each layer is the input of the next layer; hidden layer state output by previous LSTM unit is required in cyclic update of encoder module data
And cell state s t-1 As input parameters, use is made of:
h t ＝f 1 (h t-1 ,X t )
representing one such update calculation process; wherein f is 1 The LSTM unit in question is represented by,
real number vector space, X, representing m dimensions t Representing the input signature sequence at time t;
in the LSTM unit, at each time step there is a formula:
wherein, f t ,i t ,o t Respectively showing a forgetting gate, an input gate and an output gate;
are respectively applied to r t And h t-1 The weight matrix of (a) is determined,
represents 4m × d r The real number vector space of the dimension is,
real vector space representing 4m x m dimensions, m being the hidden layer dimension, d r Is an input of r t The vector dimension of (a); r is a radical of hydrogen t Is the input at time t; h is a total of t-1 Is a hidden state vector output at the time of t-1;
the candidate unit state vector is output by the neural network model at the current moment; sigmoid and tanh represent different activation functions, respectively;
then, the weight corresponding to each input feature sequence can be obtained through an input attention mechanism
Wherein,
is an intermediate variable, has no specific actual meaning,
and
is to note the parameters that need to be learned in the force model,
a real number vector space representing the dimension T,
a real vector space representing the dimension T x 2m,
a real vector space representing a dimension of T x T; tanh represents a hyperbolic tangent activation function, exp represents an exponential function;
the calculation of (1) is processed by a softmax layer;
through the attention weight, the input feature sequence extracted in a self-adaptive mode can be obtained:
the hidden layer state of the LSTM unit may then be updated as:
use of
Representing the adaptive input signature sequence at time t, thereby enabling the encoder to selectivelyThe method focuses on more important relevant input features, treats all input feature sequences equally, and realizes the mutual dependency relationship among the feature sequences;
in a decoder module, a time attention mechanism is used for adaptively selecting relevant hidden layer states in a decoding layer of a model because of the problem that the characterization capability of an input sequence is reduced when the input sequence is too long in a traditional coder-decoder model, and the model effect is rapidly deteriorated.
Similar to the approach in the encoder block, the attention mechanism in the decoder block also requires hidden layer states from the previous LSTM unit output
And cell state
As input parameters, where z represents the dimension of the hidden layer in the encoder,
real vector space representing z dimension, importance weight thereof
The formula derivation process and the input feature quantity sequence attention degree
The calculation process of (a) is the same,
representing the k-th encoder concealment state h k The magnitude of importance to the final prediction. Then the decoder sums all the states of the hidden layers of the encoder according to the weights to obtain a text vector:
incorporating historical load data L T-P ,L T-P+1 ,...,L T-1 And the obtained text vectors are subjected to vector splicing and linear change to obtain the decoding layer input of self-adaptive extraction:
wherein,
for the concatenation of the load at time t-1 and the computed text vector, m is the hidden layer dimension in the decoder described earlier,
representing a real vector space in the m +1 dimension.
And
is the parameter to be learned during the linear transformation. Then using the calculation
Updating decoder hidden layer state:
wherein, f 2 Is a non-linear activation function of LSTM unit, which is specifically updated by calculation method and f 1 And (5) the consistency is achieved.
The output layer is composed of a multilayer perceptron, and the final hidden layer state output by the coding and decoding model is namely { d T-P ,d T-P+1 ,...,d T-1 And (5) as an input, outputting the final model prediction result through a three-layer perceptron network, wherein the first two layers of the multi-layer perceptron use a PReLU as an activation function:
f 3 ＝max(μd t ,d t )
wherein f is 3 Representing the PReLU activation function, μ is a parameter that is updated only during the training process. The PReLU activation function avoids the problem that part of the parameters cannot be updated. And the activation function of the last layer of sensor is a Sigmoid function so as to ensure that the prediction result can be limited within a reasonable range.
Finally, T time steps are carried out to obtain the nonlinear part of the final load prediction result
Wherein,
representing the nonlinear part of the load prediction after T time steps, we denote the time step size by T, non denotes the sign of the above nonlinearity, F non Non-linear calculation procedure for representing the above-mentioned overall load prediction, { L T-P ,...,L T-1 Denotes the load size of P-1 time steps before T time step, { X T-P ,...,X T And represents the input characteristic sequence P time steps before the T time step.
Is a concatenated vector of hidden layer states and text vectors,
real vector space, parameter W, representing z + m dimensions y And b w Mapping of the stitching vector to the size of the decoding layer hidden layer state is realized, wherein W y Representing the corresponding non-linear mapping weights of the stitching vector during the calculation process of the decoder, b w Representing bias values, the calculation being effected by a computer program, the same way as
And u w Representing non-linear mapping weights and bias values in the output layer, respectively, using F non A neural network prediction function representing the above process,
representing the non-linear components of the calculated result, i.e. the final predicted result.
Optionally, a specific calculation formula of the autoregressive model is as follows:
wherein, { L T-P ,L T-P+1 ,...,L T-1 Is historical data, ε T For random disturbance variables, λ t For the weight size corresponding to each moment, the two variables can be initialized and automatically updated in the design of the autoregressive model, and F is used linear An autoregressive prediction function representing the above process,
representing the linear component of the calculated result, i.e. the final predicted result.
In a second aspect, the present invention further provides a cloud-oriented multi-data center load prediction apparatus, including:
the data processing module is used for acquiring a log record file for recording the resource use condition of the virtual machine at each time point and extracting required characteristic quantity data and historical load data from the log record file; converting the characteristic quantity data and the historical load data into corresponding input characteristic sequences and historical load vectors;
the nonlinear component prediction module is used for calculating to obtain a nonlinear component of load prediction by utilizing a pre-constructed neural network model based on the obtained input characteristic sequence and the historical load vector;
the linear component prediction module is used for calculating to obtain a linear component of load prediction by utilizing a pre-constructed autoregressive model based on the obtained historical load vector;
and the prediction result calculation module integrates the nonlinear component and the linear component of the load prediction to obtain a final load prediction result.
In a third aspect, the present invention further provides a cloud-oriented multi-data center resource scheduling method, including the following processes:
calculating to obtain a load prediction result of the virtual machines on each server in the cluster under the cloud multi-data center environment based on the method;
and generating a corresponding resource scheduling strategy based on the load prediction result of the virtual machine on each server.
In a fourth aspect, the present invention further provides a resource allocator, including:
the load prediction module is used for calculating and obtaining a load prediction result of the virtual machine on each server in the cluster under the cloud multi-data center environment based on the method;
and the resource scheduling module is used for generating a corresponding resource scheduling strategy based on the load prediction result of the virtual machine on each server.
Compared with the prior art, the invention has the following beneficial effects: the invention comprehensively considers the linear trend and the nonlinear characteristic of the load sequence changing along with time under the environment of multiple data centers, combines the neural network method based on the LSTM with the statistical learning method of the autoregressive model, and can effectively improve the prediction precision of future load measurement.
Drawings
FIG. 1 is a schematic diagram of the system of the present invention;
FIG. 2 is a flow chart of the method of the present invention;
FIG. 3 is a schematic diagram of an encoder decoder module in a neural network model;
FIG. 4 is a schematic diagram of a multi-layer perceptron output network in a neural network model;
FIG. 5 is a block diagram of a detailed schematic of the system of the present invention;
fig. 6 shows the actual load change trend of a server in a data center over a period of time.
Detailed Description
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
Example 1
The invention designs a load prediction method and a resource scheduling method for cloud-side multi-data centers, the system takes log record files of all virtual machines acquired by a cloud-side data center as input, linear and nonlinear changes of loads are predicted by preprocessing data, and finally the system can generate a corresponding resource allocation scheduling strategy according to load prediction results.
The invention discloses a load prediction method facing cloud multi-data center, as shown in fig. 2, comprising the following steps:
data collection and processing are all oriented to all virtual machines, a large number of virtual machine log record files are collected in the process, and the processing and analysis methods are the same, so that analysis in the embodiment takes one virtual machine as an example to process the log record files and predict load.
In order to simplify the complicated data information in the mass log recording files, non-time sequence information which has no influence on load prediction is removed. The invention firstly appoints the type of the characteristic quantity needed by the model in the preprocessing stage and standardizes the time sequence data format. Thereby facilitating learning richer structural information and contextual associations in a variety of data that contribute to load.
The specific process of preprocessing the log record file comprises the following steps:
1) Firstly, data cleaning is carried out on the collected log record files, required characteristic quantity data is extracted, the rest of record data irrelevant to load prediction is abandoned, and only the resource use condition of the virtual machine at each time point is reserved. If the local data abnormality occurs, averaging adjacent data before and after a field to replace abnormal data, and when the sequence loss of a certain section reaches two or more, considering the possible complex nonlinear relation among the characteristics of the data center, in order to avoid the influence of a simple data processing mode on the training precision, discarding the abnormal data with data loss.
2) The finally acquired data should contain, but not be limited to, the following characteristic quantities: sampling and recording time, virtual machine kernel number configuration, CPU capacity, CPU effective usage, virtual machine memory configuration capacity, virtual machine memory effective usage, disk reading throughput, disk writing throughput, network receiving throughput and network transmission throughput. The effective CPU usage is used as the target of load prediction, and the rest characteristic quantities are used as the input data of the model.
Wherein the recording time 1970.01.01 of the characteristic quantity is the default starting timing time of the computer, the unit MHZ represents one of the units of megahertz wave frequency, and the unit KB/S represents the number of kilobytes that can be processed per second.
3) And then, the collected characteristic quantity data is segmented by utilizing a sliding window to form a time sequence curve with fixed time step length, and the time sequence curve is subsequently used as an input characteristic sequence of the model.
By using
Representing an input characteristic sequence obtained after finishing the log file of a certain virtual machine, wherein n represents the characteristic of load prediction inputDimension, P represents the time step of the input feature sequence, T represents the predicted target time,
representing a real space of dimension n x P. Is provided with
Each X k Each representing a sequence of features of time step P as one of the input features, and introducing
An input signature sequence representing a load prediction of n input signatures at a certain time t.
In addition, in the case of the present invention, using L = { L T-P ,L T-P+1 ,...,L T-1 Represents the historical load data vector, using
The effective usage amount of the CPU of the cloud data center at time T, that is, the current instant, is also our prediction target.
In summary, after the log record file is preprocessed, the required input feature sequence X = { X } is obtained 1 ,X 2 ,...,X n } P And a historical load data vector L = { L = { (L) T-P ,L T-P+1 ,...,L T-1 }。
And 2, respectively transmitting the input feature sequence and the historical load data vector obtained after the processing in the step 1 into a pre-designed neural network model and an autoregressive model, and outputting a load prediction result at the next moment.
The neural network model comprises an encoder, a decoder and a multilayer perceptron network. The input of the encoder is an acquired input feature sequence, the input of the decoder is an input feature sequence which is output by an upper-layer encoder and extracted in a self-adaptive mode, and the input of the multilayer perceptron network is a text vector output by the decoder. And extracting the interdependence relation among the characteristic sequences by using a neural network module, analyzing the nonlinear variation trend among the characteristic quantities, and finally outputting the nonlinear component part of the load prediction result. An attention mechanism is embedded in the encoder and the decoder for exploring the weight magnitude of the influence of the characteristic data on the load and analyzing the influence of the previous load sequence and each characteristic quantity on the load.
The structure and the specific processing procedure of the neural network are shown in fig. 3 and fig. 4, and first all the variables and symbols appearing in fig. 3 and fig. 4 are explained: { X 1 ,X 2 ,...,X n } P Representing the input signature sequence collected, it may also be referred to as a time sequence because it is data collected at individual time nodes. The time series can be split so that e.g. { x } can be obtained 1 T-P ,x 1 T-P+1 ,...,x 1 T To { x } n T-P ,x n T-P+1 ,...,x n T N feature vectors composed of single feature quantities are shown, wherein n represents a feature dimension of the load prediction input, P represents a time step of the input feature sequence, and T represents a predicted target time.
h t Hidden layer tensor data corresponding to a certain time t is generated in an encoder representing a neural network model. The softmax function may also be called a normalized exponential function, in order to make the sum of all hidden layer weights generated by the encoder 1.
Representing the value of the k-th input feature sequence at time t,
and representing the weight corresponding to the value of the kth input feature sequence obtained in the encoder at the time t.
Representing the input vector from the kth adaptive extraction at time t. LSTM stands for long-short term memory neural network, which is an important component of model parameter updates in codec models. d is a radical of t RepresentsAnd hidden layer tensor data corresponding to the t moment generated in the decoder.
Representing the weight corresponding to the value of the i-th encoder hidden layer state obtained in the decoder at time t.
Is a text vector resulting from the decoder summing all the encoder hidden states according to weights,
is a summation symbol representing
To
All values of (a) are added. L is t Representing the corresponding load at time t i.e. the active use of the CPU,
representing the nonlinear component of the neural network model at time T that yields the load prediction results.
In the encoder module, an input attention mechanism is introduced, so that the input feature sequence is weighted in an adaptive mode. The encoder modules include the input attention layer, softmax layer, and LSTM neural network layer as shown in fig. 3, and the output of each layer is the input of the next layer. Hidden layer state output by previous LSTM unit is required in cyclic update of encoder module data
And cell status s t-1 As input parameters, use is made of:
h t ＝f 1 (h t-1 ,X t )
indicating one such update calculation process. Wherein f is 1 Represents the LSTM sheetThe number of the meta-data is one,
real number vector space, X, representing m dimensions t Representing the input signature sequence at time t.
In the LSTM unit, at each time step there is a formula:
wherein f is t ,i t ,o t Respectively showing a forgetting gate, an input gate and an output gate;
are respectively applied to r t And h t-1 The weight matrix of (a) is determined,
represents 4m × d r A real number vector space of a dimension that,
real vector space representing 4m x m dimensions, m being the hidden layer dimension, d r Is an input of r t The vector dimension of (a); r is a radical of hydrogen t Is the input at time t; h is a total of t-1 Is a hidden state vector output at the time of t-1;
the candidate unit state vector is output by the neural network model at the current moment; sigmoid and tanh represent different activation functions, respectively.
Then, the weight corresponding to each input feature sequence can be obtained through an input attention mechanism
Wherein,
is an intermediate variable, has no specific actual meaning,
and
is to note the parameters that need to be learned in the force model,
a real number vector space representing the dimension T,
a real vector space representing the dimension T x 2m,
a real vector space representing the dimension T x T. tanh represents a hyperbolic tangent activation function, and exp represents an exponential function.
The calculation of (b) is processed by the softmax layer. Through the attention weight, the input feature sequence extracted in a self-adaptive mode can be obtained:
the hidden layer state of the LSTM unit may then be updated as:
use of
And the adaptive input characteristic sequences at the time t are shown, so that the encoder can selectively focus on more important related input characteristics, not treat all input characteristic sequences equally, and the mutual dependency relationship among the characteristic sequences is explored.
In the decoder module, because the characterization capability of the traditional coder decoder model is reduced when the input sequence is too long, the model effect is rapidly deteriorated, and therefore a time attention mechanism is used for adaptively selecting relevant hidden layer states in the decoding layer of the model.
Similar to the approach in the encoder block, the attention mechanism in the decoder block also requires hidden layer states from the previous LSTM unit output
And cell state
As input parameters, where z represents the dimension of the hidden layer in the encoder,
real number vector space representing z dimension, importance weight thereof
The formula derivation process and the input feature quantity sequence attention degree
The calculation process of (a) is the same,
representing the k-th encoder concealment state h k The magnitude of importance to the final prediction. Then the decoder sums all the states of the hidden layers of the encoder according to the weights to obtain a text vector:
incorporating historical load data L T-P ,L T-P+1 ,...,L T-1 And the obtained text vectors are subjected to vector splicing and linear change to obtain the decoding layer input of self-adaptive extraction:
wherein,
for the concatenation of the load at time t-1 and the computed text vector, m is the hidden layer dimension in the decoder described earlier,
representing a real vector space in the m +1 dimension.
And
are the parameters to be learned during the linear transformation. Then using the calculation
Updating decoder hidden layer state:
wherein, f 2 Is a non-linear activation function of LSTM unit, which is specifically updated by calculation method and f 1 And (5) the consistency is achieved.
The output layer is composed of a multilayer perceptron, and the final hidden layer state output by the coding and decoding model is namely { d T-P ,d T-P+1 ,...,d T-1 Taking the PReLU as an activation function in the first two layers of the multi-layer perceptron:
f 3 ＝max(μd t ,d t )
wherein f is 3 μ is a parameter that is updated only during the training process, representing the PReLU activation function. The PReLU activation function avoids the problem that part of the parameters cannot be updated. And the activation function of the last layer of sensor is a Sigmoid function so as to ensure that the prediction result can be limited within a reasonable range.
Finally, T time steps are carried out to obtain the nonlinear part of the final load prediction result
Wherein,
representing the nonlinear part of the load prediction after T time steps, we denote the time step size by T, non denotes the sign of the above nonlinearity, F non Non-linear calculation procedure for representing the above-mentioned overall load prediction, { L T-P ,...,L T-1 Denotes the load size of P-1 time steps before T time step, { X T-P ,...,X T And represents the input characteristic sequence P time steps before the T time step.
Is a concatenated vector of hidden layer states and text vectors,
real vector space, parameter W, representing z + m dimensions y And b w Mapping the splicing vector into the size of the hidden layer state of the decoding layer is realized, wherein W y Representing the corresponding non-linear mapping weights of the stitching vector during the calculation process of the decoder, b w Representing bias values, the calculation being effected by a computer program, the same way as
And u w Representing non-linear mapping weights and bias values in the output layer, respectively, using F non A neural network prediction function representing the above process,
representing the non-linear components of the calculated result, i.e. the final predicted result.
In the prediction of the linear change of the load, the load data of a plurality of past time steps are used as the input of an autoregressive model according to the collected historical load data, and the predicted load value of the next time step is predicted. Therefore, the long-term linear change trend of the load change can be explored, and the problem that the input and output scale of the neural network model is not sensitive is avoided.
The specific calculation formula of the autoregressive model is as follows:
wherein, { L } T-P ,L T-P+1 ,...,L T-1 Is historical data, ε T For randomly disturbing variables, λ t For the weight size corresponding to each moment, the two variables can be initialized and automatically updated in the design of the autoregressive model, and F is used linear An autoregressive prediction function representing the above process,
representing the linear component of the calculated result, i.e. the final predicted result.
And 3, returning the obtained load prediction result to the data center resource distributor, and sending the generated resource distribution strategy to the cloud data center by the resource distributor to perform resource distribution work of service.
As shown in fig. 1 and fig. 5, applying the load prediction result to the resource allocator specifically includes:
in the above analysis modeling process, as shown in the real load variation trend chart shown in fig. 6, the variation process of the load tends to be a linear variation trend and a nonlinear variation trend coexisting in the whole. Therefore, in the load prediction analysis, the comprehensive analysis of the load nonlinear component and the load linear component is helpful for improving the accuracy of the prediction result. Therefore, the proposed method combines the output results of the nonlinear predictive component and the linear predictive component of the load, i.e. the neural network model and the autoregressive model
And
as a final prediction result.
The final prediction result can be expressed as:
wherein,
representing the predicted components of the autoregressive model obtained at time T,
representing the predicted component of the neural network model at time T. [ d T ；C T ]∈R p+m Is a concatenation vector of hidden layer states and text vectors, parameter W y And b w Mapping the stitching vector to the size of the decoding layer hidden layer state is achieved,
and b v Respectively representing non-linear mapping weights and bias values in the output layer, { L T-P ,L T-P+1 ,...,L T-1 Denotes historical load data, { X } T-P ,X T-P+1 ,...,X T Represents the input sequence vector.
And then, the resource distributor generates a corresponding resource scheduling strategy according to the prediction result, and flexibly adjusts the occupied resources of the virtual machines of the servers.
The system needs to predict the virtual machine load on each server at time T
And feeding back the load prediction result to the resource distributor, generating a corresponding resource distribution strategy by the resource distributor according to the load prediction result, and dynamically adjusting the resource distribution condition of each virtual machine by the cloud data center according to the strategy.
Step 4, predicting the load results at a plurality of past moments
The current actual load data collected by the cloud data center are fed back to the prediction model, so that the model can continuously obtain new experimental data to perform model training, the performance of model prediction is further improved, and errors are reduced.
The system can continuously feed back the load prediction result and the actual load error to the model, so that the prediction deviation of the load model caused by lack of data is reduced.
The resource distributor generates a corresponding resource distribution result after future load prediction is carried out on a server cluster in a cloud multi-data center environment based on a load prediction method. The specific implementation details of the resource allocator can be realized by computer coding, and the scheduling method is selected as follows: according to the load prediction result, more computing resources such as the number of CPU cores, internal memory and the like are distributed for the server which possibly bears more computing tasks in a future period of time, so that the current server can continuously support the service operation of the server. When the computing resources are on line in the distribution of a single server, the resource distributor needs to consider redistributing the upcoming computing tasks, and avoids the task blocking and even the system crash caused by the single server node bearing excessive task load. If the load among the servers is balanced, a counter can be set, and the calculation tasks are evenly distributed. Under the environment of multiple data centers, when the computing resource margin is limited, the computing tasks are considered to be returned to the cloud end for redistribution, so that the computing delay caused by the accumulation of the tasks is avoided.
In summary, the present invention predicts the future load change condition of each virtual machine on the server according to the collected log record file, and dynamically adjusts and configures the data center resources by using the resource allocator. The long-term change trend of load sequence change and the interdependency of load characteristics and other characteristic sequences under the data center environment are comprehensively considered, and a fusion model based on a neural network and an autoregressive method is designed to predict the load of multiple data centers. The robustness of the prediction result is improved, and the problem of insensitive scale of the neural network model is avoided. And the flexibility of the model is ensured, so that the method can adapt to the load prediction of the data center with different variation trends.
Example 2
Based on the same inventive concept as embodiment 1, an embodiment of the present invention provides a cloud-side-oriented load prediction apparatus for multiple data centers, including:
the data processing module is used for acquiring a log record file for recording the resource use condition of the virtual machine at each time point and extracting required characteristic quantity data and historical load data from the log record file; converting the characteristic quantity data and the historical load data into corresponding input characteristic sequences and historical load vectors;
the nonlinear component prediction module is used for calculating to obtain a nonlinear component of load prediction by utilizing a pre-constructed neural network model based on the obtained input characteristic sequence and the historical load vector;
the linear component prediction module is used for calculating to obtain a linear component of load prediction by utilizing a pre-constructed autoregressive model based on the obtained historical load vector;
and the prediction result calculation module integrates the nonlinear component and the linear component of the load prediction to obtain a final load prediction result.
The specific implementation scheme of each module of the device is shown in the implementation process of each step of the method in the embodiment 1.
Example 3
Based on the same inventive concept as embodiment 1, a resource allocator of an embodiment of the present invention includes:
the load prediction module is used for calculating and obtaining a load prediction result of the virtual machines on each server in the cluster under the cloud multi-data center environment based on the method;
and the resource scheduling module is used for generating a corresponding resource scheduling strategy based on the load prediction result of the virtual machine on each server.
The specific implementation scheme of each module of the device is shown in the implementation process of each step of the method in the embodiment 1.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.
Claims (7)
1. A load prediction method facing cloud multi-data center is characterized by comprising the following processes:
acquiring a log record file for recording the resource use condition of each time point virtual machine, and extracting required characteristic quantity data and historical load data from the log record file; converting the characteristic quantity data and the historical load data into corresponding input characteristic sequences and historical load vectors;
calculating to obtain a nonlinear component of load prediction by utilizing a pre-constructed neural network model based on the obtained input characteristic sequence and the historical load vector;
calculating to obtain a linear component of load prediction by utilizing a pre-constructed autoregressive model based on the obtained historical load vector;
integrating the nonlinear component and the linear component of the load prediction to obtain a final load prediction result;
the neural network model includes: an encoder, a decoder and a multi-layer perceptron network; the input of the encoder is an acquired input feature sequence, the input of the decoder is an input feature sequence which is output by an upper-layer encoder and is extracted in a self-adaptive mode, and the input of the multilayer perceptron network is a text vector output by the decoder;
the method for obtaining the nonlinear component of the load prediction by utilizing the pre-constructed neural network model calculation based on the obtained input feature sequence and the historical load vector comprises the following steps:
the encoder module comprises an input attention layer, a softmax layer and an LSTM neural network layer, wherein the output of each layer is the input of the next layer; hidden layer state output by previous LSTM unit in data cycle update of encoder module
And cell state s t-1 As input parameters, use is made of:
h t ＝f 1 (h t-1 ,X t )
representing one such update calculation process; wherein f is 1 It is indicated that the LSTM unit in question,
real number vector space, X, representing m dimensions t Representing the input signature sequence at time t;
in the LSTM unit, at each time step, there is the following calculation:
wherein f is t ，i t ，o t Respectively showing a forgetting gate, an input gate and an output gate;
are respectively applied to r t And h t-1 The weight matrix of (a) is determined,
represents 4m × d r The real number vector space of the dimension is,
real vector space representing 4m x m dimensions, m being the hidden layer dimension, d r Is an input of r t The vector dimension of (a); r is t Is the input at time t; h is t-1 Is a hidden state vector output at time t-1;
the candidate unit state vector is output by the neural network model at the current moment; sigmoid and tanh respectively represent different activation functions;
then obtaining the weight corresponding to each input feature sequence through an input attention mechanism
Wherein,
is an intermediate variable, has no specific actual meaning,
and
is to be noted the parameters to be learned in the force model,
a real vector space representing the dimension T,
a real vector space representing the dimension T x 2m,
a real vector space representing a dimension of T x T; tanh represents a hyperbolic tangent activation function, exp represents an exponential function;
the calculation of (1) is processed by a softmax layer;
obtaining an input feature sequence extracted in a self-adaptive mode according to the attention degree weight:
use of
The adaptive input feature sequence at time t is represented, and the hidden layer state of the LSTM unit is updated as follows:
in the decoder module, similar to the encoder module, the attention mechanism in the decoder module is to hide the layer state output by the previous LSTM unit
And cell status
As input parameters, where z represents the dimension of the hidden layer in the encoder,
real vector space representing z dimension, importance weight thereof
The formula derivation process and the input feature quantity sequence attention degree
The calculation process of (a) is the same,
representing the kth encoder hidden state h k The magnitude of importance to the final prediction; then the decoder sums all the hidden layer states of the encoder according to the weights to obtain a text vector:
incorporating historical load data L T-P ,L T-P+1 ,...,L T-1 And (4) obtaining the self-adaptively extracted decoding layer input by vector splicing and linear change of the obtained text vector:
wherein,
is the concatenation of the load at time t-1 and the computed text vector, m is the hidden layer dimension in the decoder,
a real number vector space representing a dimension of m + 1;
and
is a parameter to be learned in the linear transformation process; then using the calculation
Updating decoder hidden layer state:
wherein f is 2 Is a non-linear activation function of LSTM unit, which is specifically updated by calculation method and f 1 The consistency is achieved;
the output layer is the final hidden layer state output by the coding and decoding model, namely { d T-P ,d T-P+1 ,...,d T-1 Taking the PReLU as an activation function in the first two layers of the multi-layer perceptron:
f 3 ＝max(μd t ,d t )
wherein f is 3 Represents the PReLU activation function, μ is a parameter that is updated only during the training process; excitation of last layer perceptronThe live function is a sigmoid function;
obtaining the non-linear part of the final load prediction result after T time steps
Wherein,
representing the non-linear part of the load prediction after T time steps, T representing the size of the time step, non representing the identity of the non-linearity, F non Non-linear calculation procedure for representing the entire load prediction, { L T-P ,...,L T-1 Denotes the load size of P-1 time steps before T time step, { X T-P ,...,X T Representing an input characteristic sequence P time steps before the T time step;
is a concatenated vector of hidden layer states and text vectors,
real vector space, parameter W, representing the z dimension y And b w Mapping of the stitching vector to the size of the decoding layer hidden layer state is realized, wherein W y Representing the corresponding non-linear mapping weights of the concatenated vector during the decoder computation, b w Indicating the offset value, in the same way
And u w Representing non-linear mapping weights and bias values in the output layer, respectively, using F non A neural network prediction function representing the above process.
2. The cloud-oriented multi-data center load prediction method as claimed in claim 1, wherein the feature quantity includes: sampling and recording time, virtual machine kernel number configuration, CPU capacity, virtual machine memory configuration capacity, virtual machine memory effective usage, disk reading throughput, disk writing throughput, network receiving throughput and network transmission throughput; the load refers to the effective usage of the CPU.
3. The cloud-oriented multi-data center load prediction method of claim 1, wherein the converting the feature quantity data into the corresponding input feature sequence comprises:
and segmenting the characteristic quantity data by using a sliding window to form a time sequence with fixed time step length as an input characteristic sequence.
4. The cloud-oriented multi-data center load prediction method according to claim 1, wherein a specific calculation formula of the autoregressive model is as follows:
wherein, { L } T-P ,L T-P+1 ,...,L T-1 Is historical data, ε T For randomly disturbing variables, λ t For the weight magnitude, F, corresponding to each time instant linear Represents an auto-regressive prediction function and,
representing the linear component of the calculated result, i.e. the final predicted result.
5. An apparatus of the cloud-oriented multi-data center load prediction method according to any one of claims 1 to 4, comprising:
the data processing module is used for acquiring a log record file for recording the resource use condition of the virtual machine at each time point and extracting required characteristic quantity data and historical load data from the log record file; converting the characteristic quantity data and the historical load data into corresponding input characteristic sequences and historical load vectors;
the nonlinear component prediction module is used for calculating to obtain a nonlinear component of load prediction by utilizing a pre-constructed neural network model based on the obtained input characteristic sequence and the historical load vector;
the linear component prediction module is used for calculating to obtain a linear component of load prediction by utilizing a pre-constructed autoregressive model based on the obtained historical load vector;
and the prediction result calculation module integrates the nonlinear component and the linear component of the load prediction to obtain a final load prediction result.
6. A cloud-side multi-data-center-oriented resource scheduling method is characterized by comprising the following processes:
calculating and obtaining a load prediction result of a virtual machine on each server in a cluster in a cloud multi-data center environment based on the method of any one of claims 1-4;
and generating a corresponding resource scheduling strategy based on the load prediction result of the virtual machine on each server.
7. A resource allocator, comprising:
the load prediction module is used for calculating and obtaining a load prediction result of the virtual machine on each server in the cluster under the cloud multi-data center environment based on the method of any one of claims 1 to 4;
and the resource scheduling module is used for generating a corresponding resource scheduling strategy based on the load prediction result of the virtual machine on each server.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110473131.1A CN113220450B (en) | 2021-04-29 | 2021-04-29 | Load prediction method, resource scheduling method and device for cloud-side multi-data center |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110473131.1A CN113220450B (en) | 2021-04-29 | 2021-04-29 | Load prediction method, resource scheduling method and device for cloud-side multi-data center |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113220450A CN113220450A (en) | 2021-08-06 |
| CN113220450B true CN113220450B (en) | 2022-10-21 |
Family
ID=77089970
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110473131.1A Active CN113220450B (en) | 2021-04-29 | 2021-04-29 | Load prediction method, resource scheduling method and device for cloud-side multi-data center |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113220450B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114064203B (en) * | 2021-10-28 | 2024-07-23 | 西安理工大学 | Cloud virtual machine load prediction method based on multi-scale analysis and depth network model |
| CN114124554B (en) * | 2021-11-29 | 2022-08-30 | 燕山大学 | Virtual network service chain throughput prediction method |
| CN114338694B (en) * | 2022-03-04 | 2022-05-31 | 广州鹏捷科技股份有限公司 | One-stop cloud data center server scheduling method and system |
| CN115460061B (en) * | 2022-08-03 | 2024-04-30 | 中国科学院信息工程研究所 | Health evaluation method and device based on intelligent operation and maintenance scene |
| CN115509752A (en) * | 2022-09-29 | 2022-12-23 | 福州大学 | Edge prediction method based on deep autoregressive recurrent neural network |
| CN118394592A (en) * | 2024-04-16 | 2024-07-26 | 广州视声智能股份有限公司 | Paas platform based on cloud computing |
| CN118227742A (en) * | 2024-05-24 | 2024-06-21 | 浙江口碑网络技术有限公司 | Data trend analysis method, device, equipment, storage medium and program product |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106502799A (en) * | 2016-12-30 | 2017-03-15 | 南京大学 | A kind of host load prediction method based on long memory network in short-term |
| CN108170529A (en) * | 2017-12-26 | 2018-06-15 | 北京工业大学 | A kind of cloud data center load predicting method based on shot and long term memory network |
| CN108196957A (en) * | 2017-12-28 | 2018-06-22 | 福州大学 | A kind of host load prediction method under cloud environment |
| CN111638958A (en) * | 2020-06-02 | 2020-09-08 | 中国联合网络通信集团有限公司 | Cloud host load processing method and device, control equipment and storage medium |
-
2021
- 2021-04-29 CN CN202110473131.1A patent/CN113220450B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106502799A (en) * | 2016-12-30 | 2017-03-15 | 南京大学 | A kind of host load prediction method based on long memory network in short-term |
| CN108170529A (en) * | 2017-12-26 | 2018-06-15 | 北京工业大学 | A kind of cloud data center load predicting method based on shot and long term memory network |
| CN108196957A (en) * | 2017-12-28 | 2018-06-22 | 福州大学 | A kind of host load prediction method under cloud environment |
| CN111638958A (en) * | 2020-06-02 | 2020-09-08 | 中国联合网络通信集团有限公司 | Cloud host load processing method and device, control equipment and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 基于深度学习的主机负载在线预测模型研究;钱声攀 等;《计算机工程》;20200909;第84-89页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113220450A (en) | 2021-08-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113220450B (en) | Load prediction method, resource scheduling method and device for cloud-side multi-data center | |
| Zhu et al. | A novel approach to workload prediction using attention-based LSTM encoder-decoder network in cloud environment | |
| Mishra et al. | A survey on deep neural network compression: Challenges, overview, and solutions | |
| EP3467723B1 (en) | Machine learning based network model construction method and apparatus | |
| US20200265301A1 (en) | Incremental training of machine learning tools | |
| Yang et al. | Multi-step-ahead host load prediction using autoencoder and echo state networks in cloud computing | |
| CN113570064A (en) | Method and system for performing predictions using a composite machine learning model | |
| CN112800053B (en) | Data model generation method, data model calling device, data model equipment and storage medium | |
| Dogani et al. | Multivariate workload and resource prediction in cloud computing using CNN and GRU by attention mechanism | |
| CN111178986B (en) | User-commodity preference prediction method and system | |
| CN113128478A (en) | Model training method, pedestrian analysis method, device, equipment and storage medium | |
| CN114066073A (en) | Power grid load prediction method | |
| CN115699041A (en) | Extensible transfer learning using expert models | |
| Lee et al. | Quantune: Post-training quantization of convolutional neural networks using extreme gradient boosting for fast deployment | |
| Zhong et al. | Cooperative coevolutionary surrogate ensemble-assisted differential evolution with efficient dual differential grouping for large-scale expensive optimization problems | |
| Qin et al. | A hybrid deep learning model for short‐term load forecasting of distribution networks integrating the channel attention mechanism | |
| US11823066B2 (en) | Enterprise market volatility predictions through synthetic DNA and mutant nucleotides | |
| CN115409217A (en) | Multitask predictive maintenance method based on multi-expert hybrid network | |
| CN112667394B (en) | Computer resource utilization rate optimization method | |
| CN112036641B (en) | Artificial intelligence-based retention prediction method, apparatus, computer device and medium | |
| CN116149764A (en) | Cloud desktop distribution method, device, equipment and computer storage medium | |
| Metelkin et al. | Predicting Load in Datacenters Using NARX Model | |
| Klonecki et al. | Cost-constrained Group Feature Selection Using Information Theory | |
| Zhao et al. | Neural-Network-Based Feature Learning: Recurrent Neural Network | |
| US11823064B2 (en) | Enterprise market volatility prediction through synthetic DNA and mutant nucleotides |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |