CN112667394B - Computer resource utilization rate optimization method - Google Patents

Computer resource utilization rate optimization method Download PDF

Info

Publication number
CN112667394B
CN112667394B CN202011539370.4A CN202011539370A CN112667394B CN 112667394 B CN112667394 B CN 112667394B CN 202011539370 A CN202011539370 A CN 202011539370A CN 112667394 B CN112667394 B CN 112667394B
Authority
CN
China
Prior art keywords
utilization rate
resource utilization
prediction
network
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
Application number
CN202011539370.4A
Other languages
Chinese (zh)
Other versions
CN112667394A (en
Inventor
蒋锴
赵宇
张政伟
戴大伟
徐瑞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
CETC 28 Research Institute
Original Assignee
CETC 28 Research Institute
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by CETC 28 Research Institute filed Critical CETC 28 Research Institute
Priority to CN202011539370.4A priority Critical patent/CN112667394B/en
Publication of CN112667394A publication Critical patent/CN112667394A/en
Application granted granted Critical
Publication of CN112667394B publication Critical patent/CN112667394B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

The invention provides a method for optimizing the utilization rate of computer resources, which comprises the following steps: step 1, data preprocessing is carried out; step 2, coding the relation between the task characteristic sample and the resource utilization rate by using an LSTM neural network; step 3, learning the resource utilization rate sequence of the real cluster operation data to generate a resource utilization rate prediction function; step 4, inputting a task running state in the test data, predicting by using a resource utilization rate prediction function, comparing the predicted resource utilization rate with the resource utilization rate in the test data, and counting and calculating overhead and prediction errors, wherein the prediction errors comprise mean square errors, average absolute errors and standard deviations of the average absolute errors; and after the prediction is finished, optimizing the utilization rate of the computer resources. The method can be applied to optimization of the utilization rate of the cloud resources and the like. The resource utilization rate optimization method has the advantages of high prediction accuracy, good optimization stability and the like.

Description

Computer resource utilization rate optimization method
Technical Field
The invention relates to a method for optimizing the utilization rate of computer resources.
Background
The problem of resource scheduling is an important research problem in the field of cloud computing, and the difficulty of the problem lies in how to accurately predict cluster resources in real time.
In recent years, many resource scheduling algorithms have been developed, but these algorithms have some disadvantages, some algorithms are too rough for predicting used resources, and some algorithms use a rule-based resource scheduling algorithm, so that the resource usage in a real cluster cannot be predicted. Therefore, the resource utilization rate prediction method based on the inverse reinforcement learning and the LSTM neural network is provided, the change characteristics of the machine resource utilization rate are predicted by using real cluster data according to the task running condition, and the resource utilization rate is convenient to optimize.
Disclosure of Invention
The purpose of the invention is as follows: the technical problem to be solved by the invention is to provide a method for optimizing the utilization rate of computer resources aiming at the defects of the prior art, which comprises the following steps:
step 1, extracting real data of cluster operation in a period of time, and performing data preprocessing;
step 2, encoding the relation between the task characteristic sample and the resource utilization rate by using an LSTM neural network;
step 3, learning the resource utilization rate sequence of the real cluster operation data by using a reverse reinforcement learning algorithm to generate a resource utilization rate prediction function, wherein the reverse reinforcement learning algorithm comprises a strategy network and a judgment network, the strategy network is used for predicting the resource utilization rate, and the judgment network is used for evaluating the quality of the strategy network prediction;
and 4, inputting the task running state in the test data, predicting by using a resource utilization rate prediction function, comparing the predicted resource utilization rate with the resource utilization rate in the test data, and counting the calculation cost and the prediction error, wherein the prediction error comprises a mean square error, an average absolute error and a standard deviation of the average absolute error. After the algorithm training is finished, scheduling the tasks to a computer with lower predicted resource utilization rate by using a peak clipping and valley filling method, thereby optimizing the computer resource utilization rate.
The step 1 comprises the following steps:
step 1-1, original data is extracted in blocks: dividing cluster running data larger than one hour into small segments of running data in more than two continuous time periods according to the timestamp sequence of the real cluster running data and the set time length; the set time length is determined according to the demand of cluster resource utilization rate prediction;
step 1-2, packing small sections of running data in more than two continuous time periods into a whole section of running track, specifically, packing task characteristic samples and resource utilization rate sequences in the running data into state variables and action variables respectively;
step 1-3, setting the running track as a prediction environment capable of interacting with an algorithm, wherein the prediction environment refers to an environment interacting with the algorithm, the algorithm receives a given state of the prediction environment, acts on the current state, and then obtains feedback of the prediction environment.
In step 1-2, the state variables
Figure GDA0003743777360000021
Is defined as:
Figure GDA0003743777360000022
wherein i is the machine number, t is the current time, m is the length of the historical data contained in the state variable,
Figure GDA0003743777360000023
the task running state on the ith machine from t-m to t-1,
Figure GDA0003743777360000024
the real resource utilization rate of the ith machine from t-m to t-1;
action variable
Figure GDA0003743777360000025
Is defined as:
Figure GDA0003743777360000026
wherein
Figure GDA0003743777360000027
The predicted resource utilization for machine i at time t.
The step 2 comprises the following steps:
step 2-1, inputting task characteristic samples into LSTM neural networkNetwork, the task feature sample
Figure GDA0003743777360000028
Is defined as:
Figure GDA0003743777360000029
wherein
Figure GDA00037437773600000210
For the running task set at time t:
Figure GDA00037437773600000211
wherein task 1 ,task 2 Representing two different tasks;
the training goal of the LSTM neural network is a sequence of resource utilization over the past m moments
Figure GDA00037437773600000212
Step 2-2, using a back propagation method to
Figure GDA00037437773600000213
And (3) training the LSTM neural network for the fitting target, iterating for more than two times until the relative error between the LSTM neural network training target and the true value is less than ten percent, and taking out the last hidden layer output of the LSTM neural network as the code of the task characteristic sample.
The step 3 comprises the following steps:
step 3-1, initializing parameters of a strategy network and a judgment network, wherein the strategy network refers to a model for predicting resource utilization rate, and the judgment network refers to a model for judging the fitting effect of the strategy network; preparing real cluster data as a learning target of a strategy network;
step 3-2, sampling the predicted resource utilization rate from the policy network, sampling the real resource utilization rate from the real cluster data, and updating the parameters of the discrimination network by using a gradient descent method according to the error between the predicted resource utilization rate and the real resource utilization rate to serve as a cost function of the policy network;
and 3-3, training the strategy network by using a reinforcement learning method until the strategy network converges, thereby generating a resource utilization rate prediction function (reference: Schulman J, Wolski F, Dhariwal P, et al. Proximal Policy Optimization Algorithms [ J ]. 2017.).
The error calculation method in the step 4 comprises the following steps:
the mean square error MSE calculation method comprises the following steps:
Figure GDA0003743777360000031
the calculation method of the average absolute error MAE comprises the following steps:
Figure GDA0003743777360000032
where n is the total number of samples predicted, U s (i) For predicted resource utilization, U R (i) Is the true resource utilization.
The method of calculating the mean absolute error refers to the method of calculating the total standard deviation. The calculation method of the calculation overhead is to count the time required by the prediction algorithm to finish multiple predictions and calculate the average value.
Compared with the prior art, the invention has the remarkable advantages that: compared with other prediction algorithms, the cloud resource utilization rate prediction algorithm based on the inverse reinforcement learning can be combined with real cluster data prediction to better accord with a resource utilization rate sequence in a production environment, so that the use condition of the computer resource utilization rate is helped to be optimized through real-time accurate prediction, and the resource scheduling strategy is convenient to adjust in real time.
Drawings
The foregoing and/or other advantages of the invention will become further apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings.
FIG. 1 is a schematic workflow of the process of the present invention;
FIG. 2 is a plot of the mean square error predicted by the method of the present invention on the Alibab and Google datasets;
FIG. 3 is a graph of the mean absolute error predicted by the method of the present invention on the Alibab and Google datasets;
FIG. 4 is a standard deviation of the mean absolute error predicted by the method of the present invention on the Alibab and Google datasets;
FIG. 5 is the computational overhead predicted by the method of the present invention on the Alibab and Google datasets.
Detailed Description
The invention discloses a computer resource utilization rate optimization method, which is applied to the problem of resource utilization rate prediction of a real cloud environment.
As shown in FIG. 1, the present invention provides a method for optimizing the utilization rate of computer resources, which comprises the following steps:
step 1, extracting real data of cluster operation within a period of time, and performing data preprocessing. The data preprocessing comprises three parts of data blocking, operation track packaging, prediction environment setting and the like. The data blocking part divides longer time cluster running data into small sections of running data in a plurality of continuous time periods according to the time stamp sequence of the cluster running data and with a set time length. The set time length is determined according to the demand of cluster resource utilization rate prediction. And the operation track packaging part is used for packaging the task operation information and the resource utilization rate information into variables such as states, actions and the like. The prediction environment setting refers to setting a state action sequence as an interactively predictable environment. The prediction environment refers to an environment interacting with an algorithm, and the algorithm receives a given state of the prediction environment, acts on the current state and then obtains feedback of the prediction environment.
And 2, coding the relation between the task running state and the resource utilization rate by using an LSTM neural network. The task running state comprises task request resources, starting time and ending time. And the resource utilization rate sequence is used as the output of the LSTM neural network, and the neural network is optimized by using a back propagation method. The process is iteratively trained for a plurality of times until the fitting of the neural network to the resource utilization rate is completely converged. Then, the last layer of the LSTM neural network hidden layer is taken out to be used as the code of the current task running state.
And 3, learning the real resource utilization rate sequence by using an inverse reinforcement learning-based algorithm. The inverse reinforcement learning algorithm firstly learns a resource utilization rate cost function in a real environment through an inverse reinforcement learning process, and then performs reinforcement learning from the cost function to generate a resource utilization rate prediction function. The resource utilization rate prediction function is mapped to the cluster resource utilization rate from the cluster state, so that the resource utilization rate of a real cluster can be predicted by an algorithm, the inverse reinforcement learning comprises a strategy network and a judgment network, the strategy network is used for predicting the resource utilization rate, and the judgment network is used for evaluating the strategy network prediction.
And 4, inputting the task running state in the test data, predicting by using a resource utilization rate prediction function generated by a strategy network, comparing the resource utilization rate output by a prediction result with the resource utilization rate in the test data, and counting the calculation cost and the prediction error, wherein the prediction error comprises a mean square error, an average absolute error and a standard deviation of the average absolute error. After the algorithm training is finished, optimizing the resource utilization rate by using a peak clipping and valley filling method, namely scheduling the task to a computer with lower predicted resource utilization rate.
The step 1 comprises the following steps:
step 1-1, performing block extraction on original data: dividing cluster operation data larger than one hour into small sections of operation data in more than two continuous time periods according to the timestamp sequence of the real cluster operation data and the set time length; the set time length is determined according to the demand of cluster resource utilization rate prediction;
step 1-2, packing small sections of running data in more than two continuous time periods into a whole section of running track, specifically, packing task characteristic samples and resource utilization rate sequences in the running data into state variables and action variables respectively;
step 1-3, setting the running track as a prediction environment capable of interacting with an algorithm, wherein the prediction environment refers to an environment interacting with the algorithm, the algorithm receives a given state of the prediction environment, acts on the current state, and then obtains feedback of the prediction environment.
The step 2 comprises the following steps:
step 2-1, inputting task characteristic samples into an LSTM neural network, wherein the task characteristic samples
Figure GDA0003743777360000051
Is defined as:
Figure GDA0003743777360000052
wherein
Figure GDA0003743777360000053
For the running task set at time t
Figure GDA0003743777360000054
The goal of the LSTM neural network is a sequence of resource utilizations over the past m moments
Figure GDA0003743777360000055
Step 2-2, using a back propagation method to
Figure GDA0003743777360000056
And training the LSTM neural network for the fitting target, iterating for multiple times until the relative error between the LSTM neural network training target and the true value is less than ten percent, and taking out the last hidden layer output of the LSTM neural network as the encoding of the task feature sample.
The step 3 comprises the following steps:
step 3-1, initializing a strategy network and judging network parameters, wherein the strategy network refers to a model for predicting the resource utilization rate, and the judging network refers to a model for judging the fitting effect of the strategy network. Real cluster data is prepared as a learning target for the policy network.
And 3-2, sampling the predicted resource utilization rate from the strategy network, and sampling the real resource utilization rate from the real cluster data. And updating the parameters of the discrimination network according to the difference between the predicted resource utilization rate and the real resource utilization rate, and taking the parameters as the loss function of the strategy network.
The error calculation method in the step 4 comprises the following steps:
the method for calculating the mean square error comprises the following steps:
Figure GDA0003743777360000061
the average absolute error is calculated by the following method:
Figure GDA0003743777360000062
where n is the total number of samples predicted, U S (i) For predicted resource utilization, U R (i) Is the true resource utilization.
The method of calculating the mean absolute error refers to the method of calculating the total standard deviation. The calculation method for calculating the overhead is the time required by the statistical prediction algorithm to complete one prediction. As shown in fig. 3, 4 and 5.
Example (b):
as shown in fig. 2. The experiment was performed in a local cluster with five physical servers, each server was equipped with two 4-core intel (r) xeon (r) E5-2650 processors, four GeForce RTX 2080Ti GPUs, 16GB memory, 1TB disk space, and running software environments Pycharm (2018.1.4), Python 3.6, Keras 2.3.1, and tensrflow 2.2.0.
In order to evaluate the resource utilization predictor based on the inverse reinforcement learning and LSTM neural networks, three best algorithms, called sparse autoencoder and recurrent neural networks based resource prediction algorithm (L-PAW), recurrent neural networks based resource prediction algorithm (RNN), and bayesian information criterion based resource prediction algorithm (BIC-E), were used for comparison, and a brief description of each algorithm is given below.
The L-PAW algorithm uses a sparse self-encoder to encode the resource utilization rate and then inputs the resource utilization rate into a recurrent neural network for learning
The RNN algorithm inputs the resource utilization rate into a recurrent neural network for learning
BIC-E algorithm selects optimal statistical learning method based on Bayesian information criterion to predict resource utilization rate
The data set used in the experiment was a cluster operating data set for Alibaba in 2018 and a cluster operating data set for Google in 2011. The data set comprises resource allocation information, task running states and resource use conditions of each machine, and the interval of sampling points is 10 seconds. The experiment mainly uses the task use condition and resource allocation information of each machine at each time point to make a data set, and uses the invention to predict the resource utilization rate.
Step 1, extracting real data of each cluster operation in a period of time, and performing data preprocessing. The data preprocessing comprises three parts of data blocking, operation track packaging, prediction environment setting and the like. The data blocking part divides longer time cluster running data into small sections of running data in a plurality of continuous time periods according to the time stamp sequence of the cluster running data and with a set time length. The set time length is determined according to the demand of cluster resource utilization rate prediction. And the operation track packaging part is used for packaging the task operation information and the resource utilization rate information into variables such as states, actions and the like. The prediction environment setting refers to setting the state action sequence as an interactive prediction environment. The prediction environment refers to an environment interacting with an algorithm, and the algorithm receives a given state of the prediction environment, acts on the current state and then obtains feedback of the prediction environment.
And 2, coding the relation between the task running state and the resource utilization rate by using an LSTM neural network. The task running state comprises a task request resource, a starting time and an ending time. And the resource utilization rate sequence is used as the output of the LSTM neural network, and the neural network is optimized by using a back propagation method. The process is iteratively trained multiple times until the fit of the neural network to the resource utilization is fully converged. Then, the last layer of the LSTM neural network hidden layer is taken out to be used as the code of the current task running state.
And 3, learning the real resource utilization rate sequence by using an inverse reinforcement learning algorithm. The inverse reinforcement learning algorithm firstly learns a resource utilization rate cost function in a real environment through an inverse reinforcement learning process, and then performs reinforcement learning from the cost function to generate a resource utilization rate prediction function. The resource utilization rate prediction function is mapped to the cluster resource utilization rate from the cluster state, so that an algorithm can predict the resource utilization rate of a real cluster, the inverse reinforcement learning comprises a strategy network and a judgment network, the strategy network is used for predicting the resource utilization rate and inputting the task running state, and the strategy network outputs the predicted resource utilization rate through calculation. The judgment network is used for evaluating the quality of the strategy network prediction, namely the input task running state and the predicted resource utilization rate, the judgment network outputs the judgment on the quality of the predicted value through calculation, the better the predicted value is, the larger the value output by the judgment network is, the worse the predicted value is, and the smaller the value output by the judgment network is.
And 4, inputting a task running state in the test data, predicting by using a resource utilization rate prediction function generated by a strategy network, comparing the resource utilization rate output by a prediction result with the resource utilization rate in the test data, and counting the calculation cost and the prediction error, wherein the prediction error comprises a mean square error, an average absolute error and a standard deviation of the average absolute error. After the algorithm training is finished, optimizing the resource utilization rate by using a peak clipping and valley filling method, namely scheduling the task to a computer with lower predicted resource utilization rate.
The step 1 comprises the following steps:
step 1-1, carrying out block extraction on the original data of the Ali Bargo Google: dividing cluster operation data larger than one hour into small sections of operation data in more than two continuous time periods according to the timestamp sequence of the real cluster operation data and the set time length; the set time length is determined according to the demand predicted by the utilization rate of the cluster resources;
step 1-2, packing small sections of running data in more than two continuous time periods into a whole section of running track, specifically, packing task characteristic samples and resource utilization rate sequences in the running data into state variables and action variables respectively;
step 1-3, setting the running track as a prediction environment capable of interacting with an algorithm, wherein the prediction environment refers to an environment interacting with the algorithm, the algorithm receives a given state of the prediction environment, acts on the current state, and then obtains feedback of the prediction environment.
The step 2 comprises the following steps:
step 2-1, inputting task characteristic samples into an LSTM neural network, wherein the task characteristic samples
Figure GDA0003743777360000081
Is defined as:
Figure GDA0003743777360000082
wherein
Figure GDA0003743777360000083
For the running task set at time t
Figure GDA0003743777360000084
The goal of the LSTM neural network is a sequence of resource utilizations over the past m moments
Figure GDA0003743777360000085
Step 2-2, Using the reverseA propagation method of
Figure GDA0003743777360000086
And training the LSTM neural network for the fitting target, iterating for multiple times until the relative error between the LSTM neural network training target and the true value is less than ten percent, and taking out the last hidden layer output of the LSTM neural network as the encoding of the task feature sample.
The step 3 comprises the following steps:
step 3-1, initializing a strategy network and judging network parameters, wherein the strategy network refers to a model for predicting the resource utilization rate, and the judging network refers to a model for judging the fitting effect of the strategy network. Real cluster data is prepared as a learning target for the policy network.
And 3-2, sampling the utilization rate of the predicted resources from the policy network, and sampling the utilization rate of the real resources from the real cluster data. And updating the parameters of the discrimination network according to the difference between the predicted resource utilization rate and the real resource utilization rate, and taking the parameters as a loss function of the strategy network.
The error calculation method in the step 4 comprises the following steps:
the method for calculating the mean square error comprises the following steps:
Figure GDA0003743777360000091
the average absolute error is calculated by the following method:
Figure GDA0003743777360000092
where n is the total number of samples predicted, U S (i) For predicted resource utilization, U R (i) Is the true resource utilization.
The method of calculating the mean absolute error refers to the method of calculating the total standard deviation. The calculation method for calculating the overhead is the time required by the statistical prediction algorithm to complete one prediction.
The experimental result shows that when the prediction length is shorter, the method is not much different from the prior art, but along with the increase of the prediction length, the prediction accuracy and stability of the method are obviously superior to those of the traditional resource utilization rate prediction method.
Compared with the prior art, the invention has the remarkable advantages that: compared with the resource utilization rate method, the resource utilization rate prediction method based on the inverse reinforcement learning and the LSTM neural network can improve the prediction accuracy and stability and can predict the resource utilization rate under the condition of not providing input.
In a specific implementation, the present invention further provides a computer storage medium, wherein the computer storage medium may store a program, and the program may include some or all of the steps of the embodiments of the method of the present invention when executed. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a Random Access Memory (RAM).
The present invention provides a method for optimizing computer resource utilization, and a plurality of methods and approaches for implementing the technical solution, and the above description is only a preferred embodiment of the present invention, it should be noted that, for those skilled in the art, a plurality of modifications and embellishments can be made without departing from the principle of the present invention, and these modifications and embellishments should also be regarded as the protection scope of the present invention. All the components not specified in the present embodiment can be realized by the prior art.

Claims (4)

1. A method for optimizing the utilization rate of computer resources is characterized by comprising the following steps:
step 1, extracting real data of cluster operation within a period of time, and performing data preprocessing;
step 2, encoding the relation between the task characteristic sample and the resource utilization rate by using an LSTM neural network;
step 3, learning the resource utilization rate sequence of the real cluster operation data by using a reverse reinforcement learning algorithm to generate a resource utilization rate prediction function, wherein the reverse reinforcement learning algorithm comprises a strategy network and a judgment network, the strategy network is used for predicting the resource utilization rate, and the judgment network is used for evaluating the quality of the strategy network prediction;
step 4, inputting a task running state in the test data, predicting by using a resource utilization rate prediction function, comparing the predicted resource utilization rate with the resource utilization rate in the test data, and counting and calculating overhead and prediction errors, wherein the prediction errors comprise mean square errors, average absolute errors and standard deviations of the average absolute errors; after prediction is finished, scheduling the tasks to a computer with the predicted resource utilization rate lower than fifty percent by using a peak clipping and valley filling method, so as to optimize the computer resource utilization rate;
the step 1 comprises the following steps:
step 1-1, performing block extraction on original data: dividing cluster running data larger than one hour into small segments of running data in more than two continuous time periods according to the timestamp sequence of the real cluster running data and the set time length; the set time length is determined according to the demand of cluster resource utilization rate prediction;
step 1-2, packing small sections of running data in more than two continuous time periods into a whole section of running track, specifically, packing task characteristic samples and resource utilization rate sequences in the running data into state variables and action variables respectively;
step 1-3, setting the running track as a prediction environment capable of interacting with an algorithm, wherein the prediction environment refers to an environment interacting with the algorithm, the algorithm receives a given state of the prediction environment, acts on the current state, and then obtains feedback of the prediction environment;
in step 1-2, the state variables
Figure FDA0003743777350000011
Is defined as:
Figure FDA0003743777350000012
wherein i is the machine numberT is the current time, m is the length of the history data contained in the state variable,
Figure FDA0003743777350000013
the task running state on the ith machine from t-m to t-1,
Figure FDA0003743777350000014
the real resource utilization rate of the ith machine from t-m to t-1;
action variable
Figure FDA0003743777350000015
Is defined as:
Figure FDA0003743777350000016
wherein
Figure FDA0003743777350000021
The predicted resource utilization for machine i at time t.
2. The method of claim 1, wherein the step 2 comprises:
step 2-1, inputting task characteristic samples into an LSTM neural network, wherein the task characteristic samples
Figure FDA0003743777350000022
Is defined as follows:
Figure FDA0003743777350000023
wherein
Figure FDA0003743777350000024
For the running task set at time t:
Figure FDA0003743777350000025
wherein task 1 ,task 2 Representing two different tasks;
the training goal of the LSTM neural network is a sequence of resource utilization over the past m moments
Figure FDA0003743777350000026
Step 2-2, using a back propagation method to
Figure FDA0003743777350000027
And (3) training the LSTM neural network for the fitting target, iterating for more than two times until the relative error between the LSTM neural network training target and the true value is less than ten percent, and taking out the last hidden layer output of the LSTM neural network as the code of the task feature sample.
3. The method of claim 2, wherein step 3 comprises:
step 3-1, initializing parameters of a strategy network and a judgment network, wherein the strategy network refers to a model for predicting resource utilization rate, and the judgment network refers to a model for judging the fitting effect of the strategy network; preparing real cluster data as a learning target of a strategy network;
step 3-2, sampling the predicted resource utilization rate from the policy network, sampling the real resource utilization rate from the real cluster data, and updating the parameters of the discrimination network by using a gradient descent method according to the error between the predicted resource utilization rate and the real resource utilization rate to serve as a cost function of the policy network;
and 3-3, training the strategy network by using a reinforcement learning method until the strategy network is converged, thereby generating a resource utilization rate prediction function.
4. The method of claim 3, wherein the error calculation method of step 4 is:
the mean square error MSE calculation method comprises the following steps:
Figure FDA0003743777350000028
the method for calculating the average absolute error MAE comprises the following steps:
Figure FDA0003743777350000031
where n is the total number of samples predicted, U s (i) For predicted resource utilization, U R (i) Is the true resource utilization.
CN202011539370.4A 2020-12-23 2020-12-23 Computer resource utilization rate optimization method Active CN112667394B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011539370.4A CN112667394B (en) 2020-12-23 2020-12-23 Computer resource utilization rate optimization method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011539370.4A CN112667394B (en) 2020-12-23 2020-12-23 Computer resource utilization rate optimization method

Publications (2)

Publication Number Publication Date
CN112667394A CN112667394A (en) 2021-04-16
CN112667394B true CN112667394B (en) 2022-09-30

Family

ID=75408845

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011539370.4A Active CN112667394B (en) 2020-12-23 2020-12-23 Computer resource utilization rate optimization method

Country Status (1)

Country Link
CN (1) CN112667394B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113296951A (en) * 2021-05-31 2021-08-24 阿里巴巴新加坡控股有限公司 Resource allocation scheme determination method and equipment

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018211140A1 (en) * 2017-05-19 2018-11-22 Deepmind Technologies Limited Data efficient imitation of diverse behaviors
CN108986470A (en) * 2018-08-20 2018-12-11 华南理工大学 The Travel Time Estimation Method of particle swarm algorithm optimization LSTM neural network
CN110689171A (en) * 2019-09-05 2020-01-14 哈尔滨工程大学 Turbine health state prediction method based on E-LSTM

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11989647B2 (en) * 2019-02-08 2024-05-21 Adobe Inc. Self-learning scheduler for application orchestration on shared compute cluster

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018211140A1 (en) * 2017-05-19 2018-11-22 Deepmind Technologies Limited Data efficient imitation of diverse behaviors
CN108986470A (en) * 2018-08-20 2018-12-11 华南理工大学 The Travel Time Estimation Method of particle swarm algorithm optimization LSTM neural network
CN110689171A (en) * 2019-09-05 2020-01-14 哈尔滨工程大学 Turbine health state prediction method based on E-LSTM

Also Published As

Publication number Publication date
CN112667394A (en) 2021-04-16

Similar Documents

Publication Publication Date Title
CN110969290B (en) Runoff probability prediction method and system based on deep learning
CN110751318A (en) IPSO-LSTM-based ultra-short-term power load prediction method
CN111950810B (en) Multi-variable time sequence prediction method and equipment based on self-evolution pre-training
CN109787821B (en) Intelligent prediction method for large-scale mobile client traffic consumption
CN112765894B (en) K-LSTM-based aluminum electrolysis cell state prediction method
CN114493014A (en) Multivariate time series prediction method, multivariate time series prediction system, computer product and storage medium
CN116307215A (en) Load prediction method, device, equipment and storage medium of power system
CN113220450A (en) Load prediction method, resource scheduling method and device for cloud-side multi-data center
CN112667394B (en) Computer resource utilization rate optimization method
CN115983497A (en) Time sequence data prediction method and device, computer equipment and storage medium
CN108073442B (en) Simulation request execution time prediction method based on depth fuzzy stack self-coding
CN114925808B (en) Anomaly detection method based on incomplete time sequence in cloud network end resource
CN115759455A (en) Load probability density prediction method based on time sequence Gaussian mixture density network
CN116483624A (en) Monitoring data recovery model training method, recovery method, device and storage medium
CN115936184A (en) Load prediction matching method suitable for multi-user types
CN115081609A (en) Acceleration method in intelligent decision, terminal equipment and storage medium
CN115619028A (en) Clustering algorithm fusion-based power load accurate prediction method
CN115907000A (en) Small sample learning method for optimal power flow prediction of power system
CN113657501A (en) Model adaptive training method, apparatus, device, medium, and program product
CN111652701A (en) Personal credit evaluation method and system based on fusion neural network
CN117667606B (en) High-performance computing cluster energy consumption prediction method and system based on user behaviors
CN110648183A (en) Grey correlation and QGNN-based resident consumption price index prediction method
CN113313313B (en) City perception-oriented mobile node task planning method
KR102466482B1 (en) System and method for accelerating deep neural network training using adaptive batch selection
CN118193507A (en) Traffic flow data interpolation method and system based on tensor completion and graph annotation meaning network

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