CN110231976B - Load prediction-based edge computing platform container deployment method and system - Google Patents
Load prediction-based edge computing platform container deployment method and system Download PDFInfo
- Publication number
- CN110231976B CN110231976B CN201910420328.1A CN201910420328A CN110231976B CN 110231976 B CN110231976 B CN 110231976B CN 201910420328 A CN201910420328 A CN 201910420328A CN 110231976 B CN110231976 B CN 110231976B
- Authority
- CN
- China
- Prior art keywords
- node
- load
- computing
- task
- original
- 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
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
- G06F9/44—Arrangements for executing specific programs
- G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
- G06F9/45533—Hypervisors; Virtual machine monitors
-
- 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
- G06N3/045—Combinations of 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
- G06N3/08—Learning methods
-
- 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
- G06F9/44—Arrangements for executing specific programs
- G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
- G06F9/45533—Hypervisors; Virtual machine monitors
- G06F9/45558—Hypervisor-specific management and integration aspects
- G06F2009/4557—Distribution of virtual machine instances; Migration and load balancing
Abstract
The invention discloses a load prediction-based edge computing platform container deployment method and a load prediction-based edge computing platform container deployment system, which comprise a plurality of computing nodes and a central node, wherein an original load monitoring system is loaded on each computing node, the original load monitoring systems are respectively connected with a node load prediction system and uploaded to a central server through the node load prediction system, a node load prediction system and a computing task management system are loaded on the central node, an LSTM model corresponding to each computing node is arranged on the node load prediction system, and the node load prediction system receives original load information of the nodes and sends a prediction result to the computing task management system; the calculation task management system is responsible for deploying the containers, feeds back the node numbers and the task time to the node load prediction system according to the received information, and issues the containers to the available calculation nodes. The invention reasonably deploys the container to each computing node, thereby reducing the cost of computing tasks.
Description
Technical Field
The invention belongs to the technical field of edge computing platforms, and particularly relates to a load prediction-based edge computing platform container deployment method and system.
Background
In recent years, with the development of mobile internet, the data volume of internet is increased explosively, and more internet businesses are based on the analysis of big data. These have resulted in a rapid increase in demand for computing resources. The computing power of a single machine has not been able to meet the demand. Cloud computing has emerged. Cloud computing is a product of development and fusion of traditional computers and network technologies such as distributed computing, parallel computing, virtualization and load balancing. The cloud computing virtualizes a large number of servers into computing resource nodes through a virtual machine technology, and users can quickly acquire needed resources only by purchasing computing resources at the cloud without concerning the realization and maintenance of hardware. Traditional cloud computing platforms are built from high performance servers. The high-performance server has high performance, but the high-performance server is expensive, high in power consumption and high in thermal power and needs a specially designed machine room. The operation and maintenance cost becomes a main part of the cloud computing cluster cost. Meanwhile, the cloud computing presents a centralized situation, and no matter where the user is located, the accessed computing resources are stored in the cloud computing data center in a centralized mode. This results in high network costs for the user. Edge calculation solves exactly the above problem. Edge computing is geographically distributed cloud computing, which is different from traditional cloud computing in that all computing resources are concentrated in one computer room, and cloud computing nodes are lowered to the user, and the nodes are not special high-performance servers but existing devices of the user, such as a mobile network base station, an intelligent router, an intelligent mobile phone and the like. The edge computing geographically distributed design not only reduces the network cost of users, but also reduces the operation and maintenance cost of computing resource providers.
The container is a novel software standard unit which contains a software body and various dependencies required by the operation of the software body and can be quickly and reliably deployed in various computing environments. Meanwhile, because the container does not virtualize the hardware, the container has no virtualization loss, and is more suitable for the situation that the performance of the computing node is not high compared with a virtual machine, and most of the computing nodes of edge computing are the same. Since edge computing is distributed, and in order to fully utilize computing resources, a computing task is divided into appropriate subtasks, each subtask is packaged into a container, and then deployed to each node for execution. Therefore, how to deploy a group of containers contained in a computing task to each computing node is very important to ensure the synchronous normal operation and the reasonable occupation of node resources. Various algorithms have been proposed in the industry. Most of the common container deployment algorithms at present balance the load of each node on the premise of meeting the container resource demand. Some algorithms take more indexes into consideration, such as node energy consumption, network time delay and the like, so as to achieve the purposes of saving energy, improving service quality and the like. The existing container deployment algorithm is still mostly specific to the special nodes, and the container deployment algorithm for the non-specific nodes still has a large blank. Meanwhile, the existing container deployment algorithm usually considers the current load of the node, and the container usually runs for a period of time, so that the problem of insufficient node resources may occur in the container running process, at this time, the container migration needs to be performed through a feedback mechanism, and the container migration causes unnecessary network and storage overhead, and increases the cost.
Disclosure of Invention
The invention aims to solve the technical problem that aiming at the defects in the prior art, the invention provides a load prediction-based container deployment method and system for an edge computing platform, aiming at the problem that the current container deployment algorithm only utilizes the current load information of a node to cause feedback delay, namely, the container is reasonably deployed to each computing node on the premise of ensuring the good operation of the original task of the node by predicting the resource occupancy rate of the node in the recent period of time.
The invention adopts the following technical scheme:
a load prediction based edge computing platform container deployment system, comprising:
the system comprises a plurality of computing nodes and a central node, wherein each computing node is provided with a task which runs for a long time and is not distributed by an edge computing platform, the task is called as an original task, the load of the original task is called as an original load, and the central node is special;
each computing node is provided with an original load monitoring system, and each original load monitoring system is connected with the node load prediction system and used for collecting the resource occupation condition of the original task of the node, uploading the resource occupation condition to a central server through the node load prediction system and used for prediction model training and node load prediction;
the central node is provided with a node load prediction system and a calculation task management system, the node load prediction system is provided with an LSTM model corresponding to each calculation node, and the node load prediction system receives original node load information and sends a prediction result to the calculation task management system;
the calculation task management system is responsible for deploying the containers, feeds back the node numbers and the task time to the node load prediction system according to the received information, and issues the containers to the available calculation nodes.
Specifically, the original load monitoring system runs an original load information collection daemon on each node without a calculation task, and the original load information collection daemon collects original load information of the node every 5 minutes to form a 4-dimensional original load vector l (t) ═ phi1,φ2,φ3,φ4) And storing the original load vector into a database, operating an original load information uploading daemon on each node without the operation of the computing task, reading the original load vector in 60 minutes from the database by the original load information uploading daemon every 60 minutes, and uploading the original load vector to a central server.
Further, the original load information includes a CPU utilization rate, a GPU utilization rate, a memory remaining capacity, and a network utilization rate.
Specifically, the network input layer of the LSTM model is three-dimensional data [ Samples, time _ steps, features ], where Samples is the number of training Samples; time _ steps is a time step, i.e. how many time points before each datum are related to the input datum; features are eigenvalues, i.e., load vectors l (t).
Further, using R pieces of original load information to predict original load information after k time; the prediction visual field is k, and k is an integral multiple of 5 minutes of the time interval s; f represents the model to be solved, L (t + k) represents the original load information of the node at the time t + k, and the predicted behavior is as follows:
L(t+k)＝f(L(t-R+1),L(t-R+2),L,L(t-1),L(t))
wherein, the input-output pair of the f function is the input and the output of the data set; an LSTM input-output data pair is:
<(L(t-R+1),L(t-R+2),L,L(t-1),L(t)),L(t+k)>
in particular, calculatingThe task management system is used for being responsible for the deployment of the containers, after the task management system receives a task, the task is divided into s subtasks suitable for running on the nodes according to the task input scale, and the subtasks are packaged into s containers; estimating maximum load information L 'in single container operation process'maxAnd an operating time T; then, the original load information L of each node in the next T time is predicted through the LSTM model of each nodei(t) assuming that there are n nodes, the current time is t1Then there is 0<i≤n,t1<t<t1+ T; and then selecting s L 'capable of meeting container load requirement'max+Li(t) < 1 and maximize resource utilization, namely max (L'max+Li(t)) is a usable node; and finally, respectively deploying the s containers to available nodes for operation, and acquiring subtask results after the operation is finished and combining the subtask results into a final result.
Another technical solution of the present invention is a working method of an edge computing platform container deployment system based on load prediction, including the steps of:
s1, the computing task management system divides the computing task into S subtasks which are respectively packaged into containers, and the maximum load information Lmax and the running time T in the running process of a single container are estimated according to the task configuration file submitted by the user;
s2, initializing the nodes by the computing task management system, predicting original load information L of the current node in T time by the node load prediction system, judging whether the residual resources of the current node in T time meet the requirements and whether the resource utilization is maximized, and adding an available node list if the requirements are met and the resource utilization is maximized;
s3, when the residual resources of the current node in the time T do not meet the requirement or do not maximize the resource utilization or S nodes do not exist in the available node list, adding 1 to the node number and returning to the step S2;
and S4, if all nodes are traversed, distributing the container to the available nodes, and ending the deployment.
Specifically, in step S2, if there are S nodes in the available node list, the container is distributed to the available nodes, and the deployment is ended.
Compared with the prior art, the invention has at least the following beneficial effects:
the edge computing platform container deployment method based on resource prediction fully considers the current situation that an edge computing node is a non-special node and the normal operation of an original task needs to be ensured. By predicting the resource occupation of the original task, the deployment of the container is reasonably carried out under the condition of ensuring the normal operation of the original task. Meanwhile, compared with the traditional container deployment method based on the current information, the container deployment based on the resource prediction is based on the future information, so that the problem of feedback lag of the traditional method is solved, and the container migration caused by insufficient node resources is reduced.
Furthermore, the non-special characteristics of the computing nodes are fully considered by adopting the original load monitoring system, and the load information of the original task is collected in a targeted manner, so that the method is more reasonable compared with the method of directly monitoring the load condition of the nodes.
Furthermore, the containers are deployed based on the predicted load information, compared with the method that only the load of the current time of the nodes is considered, the characteristic that the computing task can be executed for a period of time is fully considered, task migration caused by insufficient node resources in the task running process is avoided, and resource consumption caused by migration is avoided.
Furthermore, whether the residual resources of the current node meet the requirements within the time T or not is judged, and whether the resource utilization is maximized is judged, so that the node resources can be more fully utilized, and the waste is reduced compared with the method only for judging whether the residual resources of the node meet the requirements or not.
In summary, the edge computing platform container deployment method based on resource prediction provided by the invention considers the current situation of non-dedicated computing nodes in edge computing, and meanwhile, the resource prediction method is adopted to avoid the feedback hysteresis, so that containers can be more reasonably deployed to each computing node, and the cost of computing tasks is reduced.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
FIG. 1 is a block diagram of the system of the present invention;
FIG. 2 is a basic structure of the LSTM of the present invention;
FIG. 3 is a flow chart of a container deployment method of the present invention.
Detailed Description
Since the computing nodes of the edge computing are often non-dedicated, before the computing task is issued to the nodes, a long-running task exists on the nodes, and the task is called an original task, and the resource occupation caused by the task is called an original load.
Referring to fig. 1, the edge computing platform container deployment system based on load prediction according to the present invention includes an original load monitoring system, a node load prediction system, and a computing task management system, where the original load monitoring system operates on a computing node for a long time, the node load prediction system and the computing task management system operate on a central node for a long time, a plurality of computing nodes are respectively connected to the node load prediction system and configured to send original load information of the node to the node load prediction system, the node load prediction system receives the original load information of the node and returns a prediction result to the computing task management system according to a requirement of the computing task management system, and the computing task management system selects an available computing node according to the received prediction result and issues a container to the available computing node.
a. Raw load monitoring system
And the original load monitoring system collects the resource occupation condition of the original task of the node and uploads the resource occupation condition to the central server for prediction model training and node load prediction.
Running a primary load information collection daemon (called collectd) on each node without a computing task, wherein the collectd collects primary load information of the node every 5 minutes, and the primary load information comprises: CPU utilization rate, GPU utilization rate, residual memory capacity and network utilization rate. The original load information constitutes a 4-dimensional load vector l (t) ═ phi1,φ2,φ3,φ4) And storing the data into a database.
And running an original load information uploading daemon (called uploadd) on each node without a computing task, wherein the uploadd reads the load vectors in 60 minutes from the database every 60 minutes and uploads the load vectors to the central server.
b. Node load prediction system
The original load of the node has a periodic character because the original task is run on the node for a long time. Aiming at the characteristic, a node original load prediction method based on a long-short-term memory neural network (LSTM) is provided, and each node has an own LSTM model.
The LSTM network input layer is three-dimensional data [ samples, time _ steps, features ], which specifically is:
samples are training Samples;
time _ steps represents a time step, i.e. how many time points before each datum relate to the input datum;
features represent the characteristic value, i.e. the load vector l (t).
Predicting raw load information after k time using the latest R pieces of raw load information, i.e., time _ steps R; the prediction visual field is set as k, and the k is required to be an integral multiple of 5 minutes; that is, the original load information of the time point t + k is predicted using the time point t and the 12 latest original load information thereof.
The model to be solved is represented by f, L (t + k) represents the original load information of the node at the time t + k, and the predicted behavior is as follows:
L(t+k)＝f(L(t-R+1),L(t-R+2),L,L(t-1),L(t))
the input-output pair of the f-function constitutes the input and output of the desired data set.
Therefore, < (L (t-R +1), L (t-R +2), L, L (t-1), L (t)), L (t + k) > form an LSTM input-output data pair.
Referring to fig. 2, LSTM has three valves per base unit to adjust whether previous network memory states are used in the current network calculations.
The three gates are respectively: forgetting gate (forget gate), input gate (input gate) and output gate (output gate).
Forget to forgetHidden layer state h before gate controln-1How much to keep the current time hn；
Input gate (input gate) controls input X of the network at the present timenHow many states h are saved to the hidden layern；
Output gate (output gate) controlling hidden layer state hnHow much output value Y is output to the current momentn。
The three control valves form a basic unit of the LSTM, called a cell, and the lower diagram is the basic structure of one unit of the LSTM neural network, wherein fnIndicating forgetting gate, inDenotes an input gate, onDenotes the output gate, hnIndicating the current cell state. The basic structure of an LSTM neural network cell is shown in FIG. 2.
Forget door fnExpressed as:
fn＝δ(Wf,xXn+Wf,yYn-1+bf)
input door inExpressed as:
in＝δ(Wi,xXn+Wi,yYn-1+bi)
output gate onExpressed as:
on＝δ(Wo,xXn+Wo,yYn-1+bo)
current cell output hnExpressed as:
hn＝hn-1fn+intanh(WcXn+UcYn-1+bc)
current cell state:
yn＝ontanh(hn)
wherein, delta is sigmoid function, and acts on forgetting gate (forget gate), input gate (input gate) and output gate (output gate), and the output is [0,1 ]]Each value indicates whether or not the corresponding partial information should pass. A value of 0 indicates that information is not allowed to pass, and a value of 1 indicates that all information is allowed to pass; and the tanh function is used forStatus and output; w is a weight, e.g. Wf,xThe weight of the output information of the last time state corresponding to the forgetting gate is shown as b.
The LSTM network structure parameters are set as follows:
time _ steps is set to 12, i.e., each data is associated with data of the previous 12 time points;
the prediction horizon is set to 5, i.e. the original load vector after 5 minutes is predicted;
the number of hidden layer nodes is set to 10. Activation, i.e., the Activation function is set to 'tanh';
the current _ activation is set to 'hard _ sigmoid' for the activation function applied in loop step;
dropout is set to 0.2;
batch _ size is set to 128;
the Optimizer is set as Adam Optimizer;
the Loss function is set to MAE.
c. Computing task management system
The computing task management system is responsible for the deployment of the containers. After the management system receives a task, the task is divided into s subtasks suitable for running on the nodes according to the input scale of the task, and the subtasks are packaged into s containers. Then estimating maximum load information L 'in the operation process of the single container'maxAnd a running time T.
Then, the original load information L of each node in the next T time is predicted through the LSTM model of each nodei(t) setting n nodes in total, the current time is t1Then there is 0<i≤n,t1<t<t1+ T. And then selecting s L 'capable of meeting container load requirement'max+Li(t) < 1 and maximize resource utilization, namely max (L'max+Li(t)) of a node. And finally deploying the s containers to the nodes for operation, and acquiring and combining the subtask results into a final result after the operation is finished.
Referring to fig. 3, the container deployment method specifically includes:
s1, the computing task management system divides the computing task into S subtasks which are respectively packaged into containers, and the maximum load information Lmax and the running time T in the running process of a single container are estimated according to the task configuration file submitted by the user;
s2, initializing nodes by a computing task management system, predicting original load information L of a current node in T time by a node load prediction system, judging whether the residual resources of the current node in T time meet the requirements and whether the resource utilization is maximized, adding an available node list if the residual resources meet the requirements and the resource utilization is maximized, distributing containers to the available nodes if S nodes exist in the available node list, and ending the deployment;
s3, when the residual resources of the current node in the time T do not meet the requirement or do not maximize the resource utilization or S nodes do not exist in the available node list, adding 1 to the node number and returning to the step S2;
and S4, if all nodes are traversed, distributing the container to the available nodes, and ending the deployment.
The above-mentioned contents are only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited thereby, and any modification made on the basis of the technical idea of the present invention falls within the protection scope of the claims of the present invention.
Claims (8)
1. An edge computing platform container deployment system based on load forecasting, comprising:
the system comprises a plurality of computing nodes and a central node, wherein each computing node is provided with a task which runs for a long time and is not distributed by an edge computing platform, the task is called as an original task, the load of the original task is called as an original load, and the central node is special;
each computing node is provided with an original load monitoring system, and each original load monitoring system is connected with a node load prediction system respectively, is used for collecting the resource occupation condition of an original task of the node, and uploads the resource occupation condition to a central server for model prediction training and node load prediction;
the central node is provided with a node load prediction system and a calculation task management system, the node load prediction system is provided with an LSTM model corresponding to each calculation node, and the node load prediction system receives original node load information and sends a prediction result to the calculation task management system;
the calculation task management system is responsible for deploying the containers, feeds back the node numbers and the task time to the node load prediction system according to the received information, and issues the containers to the available calculation nodes.
2. The system of claim 1, wherein the raw load monitoring system runs a raw load information collection daemon on each node where no computing task is running, and the raw load information collection daemon collects the raw load information of the node every 5 minutes to form a 4-dimensional raw load vector l (t) ═ phi1,φ2,φ3,φ4) And storing the original load vector into a database, operating an original load information uploading daemon on each node without the operation of the computing task, reading the original load vector in 60 minutes from the database by the original load information uploading daemon every 60 minutes, and uploading the original load vector to a central server.
3. The system of claim 2, wherein the raw load information comprises CPU usage, GPU usage, remaining memory capacity, and network usage.
4. The system of claim 1, wherein the network input layer of the LSTM model is three-dimensional data [ Samples, time-steps, features ], Samples is the number of training Samples; time _ steps is a time step, i.e. how many time points before each datum are related to the input datum; features are eigenvalues, i.e., load vectors l (t).
5. The system of claim 4, wherein R raw load information is used to predict raw load information after k times; the prediction visual field is k, and k is an integral multiple of 5 minutes of the time interval s; f represents the model to be solved, L (t + k) represents the original load information of the node at the time t + k, and the predicted behavior is as follows:
L(t+k)＝f(L(t-R+1),L(t-R+2),L,L(t-1),L(t))
wherein, the input-output pair of the f function is the input and the output of the data set; an LSTM input-output data pair is:
<(L(t-R+1),L(t-R+2),L,L(t-1),L(t)),L(t+k)>。
6. the system of claim 1, wherein the computing task management system is configured to be responsible for deployment of the containers, and when receiving a task, the computing task management system divides the task into s subtasks suitable for running on the nodes according to the task input scale, and packages the s subtasks into s containers; estimating maximum load information L 'in single container operation process'maxAnd an operating time T; then, the original load information L of each node in the next T time is predicted through the LSTM model of each nodei(t) assuming that there are n nodes, the current time is t1Then there is 0<i≤n,t1<t<t1+ T; and then selecting s L 'capable of meeting container load requirement'max+Li(t) < 1 and maximize resource utilization, namely max (L'max+Li(t)) is a usable node; and finally, respectively deploying the s containers to available nodes for operation, and acquiring subtask results after the operation is finished and combining the subtask results into a final result.
7. The method for operating the load prediction based edge computing platform container deployment system according to any one of claims 1 to 6, comprising the steps of:
s1, the computing task management system divides the computing task into S subtasks which are respectively packaged into containers, and the maximum load information Lmax and the running time T in the running process of a single container are estimated according to the task configuration file submitted by the user;
s2, initializing the nodes by the computing task management system, predicting original load information L of the current node in T time by the node load prediction system, judging whether the residual resources of the current node in T time meet the requirements and whether the resource utilization is maximized, and adding an available node list if the requirements are met and the resource utilization is maximized;
s3, when the residual resources of the current node in the time T do not meet the requirement or do not maximize the resource utilization or S nodes do not exist in the available node list, adding 1 to the node number and returning to the step S2;
and S4, if all nodes are traversed, distributing the container to the available nodes, and ending the deployment.
8. The method according to claim 7, wherein in step S2, if there are already S nodes in the available node list, the container is distributed to the available nodes, and the deployment is ended.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910420328.1A CN110231976B (en) | 2019-05-20 | 2019-05-20 | Load prediction-based edge computing platform container deployment method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910420328.1A CN110231976B (en) | 2019-05-20 | 2019-05-20 | Load prediction-based edge computing platform container deployment method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110231976A CN110231976A (en) | 2019-09-13 |
| CN110231976B true CN110231976B (en) | 2021-04-20 |
Family
ID=67861431
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910420328.1A Active CN110231976B (en) | 2019-05-20 | 2019-05-20 | Load prediction-based edge computing platform container deployment method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110231976B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11868812B2 (en) | 2021-08-12 | 2024-01-09 | International Business Machines Corporation | Predictive scaling of container orchestration platforms |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110618870B (en) * | 2019-09-20 | 2021-11-19 | 广东浪潮大数据研究有限公司 | Working method and device for deep learning training task |
| CN111459617B (en) * | 2020-04-03 | 2023-07-11 | 南方电网科学研究院有限责任公司 | Automatic allocation optimization system and method for containerized application based on cloud platform |
| CN111488213B (en) * | 2020-04-16 | 2024-04-02 | 中国工商银行股份有限公司 | Container deployment method and device, electronic equipment and computer readable storage medium |
| WO2022067557A1 (en) * | 2020-09-29 | 2022-04-07 | 西门子股份公司 | Method and apparatus for designing edge computing solution, and computer-readable medium |
| CN112969157B (en) * | 2021-02-22 | 2022-10-18 | 重庆邮电大学 | Network load balancing method for unmanned aerial vehicle |
| CN112995636B (en) * | 2021-03-09 | 2022-03-25 | 浙江大学 | 360-degree virtual reality video transmission system based on edge calculation and active cache and parameter optimization method |
| US11928503B2 (en) | 2021-06-22 | 2024-03-12 | International Business Machines Corporation | Cognitive scheduler for Kubernetes |
| US11953972B2 (en) | 2022-04-06 | 2024-04-09 | International Business Machines Corporation | Selective privileged container augmentation |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106844051A (en) * | 2017-01-19 | 2017-06-13 | 河海大学 | The loading commissions migration algorithm of optimised power consumption in a kind of edge calculations environment |
| CN108062561A (en) * | 2017-12-05 | 2018-05-22 | 华南理工大学 | A kind of short time data stream Forecasting Methodology based on long memory network model 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 |
| CN108363478A (en) * | 2018-01-09 | 2018-08-03 | 北京大学 | For wearable device deep learning application model load sharing system and method |
| CN108809695A (en) * | 2018-04-28 | 2018-11-13 | 国网浙江省电力有限公司电力科学研究院 | A kind of distribution uplink unloading strategy towards mobile edge calculations |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101860481A (en) * | 2010-05-25 | 2010-10-13 | 北京邮电大学 | Service transport method for distinguishing priority in MPLS-TP over OTN multi-layer network and device thereof |
| US8612550B2 (en) * | 2011-02-07 | 2013-12-17 | Microsoft Corporation | Proxy-based cache content distribution and affinity |
| US9720738B2 (en) * | 2015-04-09 | 2017-08-01 | International Business Machines Corporation | Datacenter scheduling of applications using machine learning techniques |
| CN108809723B (en) * | 2018-06-14 | 2021-03-23 | 重庆邮电大学 | Edge server joint task unloading and convolutional neural network layer scheduling method |
-
2019
- 2019-05-20 CN CN201910420328.1A patent/CN110231976B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106844051A (en) * | 2017-01-19 | 2017-06-13 | 河海大学 | The loading commissions migration algorithm of optimised power consumption in a kind of edge calculations environment |
| CN108062561A (en) * | 2017-12-05 | 2018-05-22 | 华南理工大学 | A kind of short time data stream Forecasting Methodology based on long memory network model 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 |
| CN108363478A (en) * | 2018-01-09 | 2018-08-03 | 北京大学 | For wearable device deep learning application model load sharing system and method |
| CN108809695A (en) * | 2018-04-28 | 2018-11-13 | 国网浙江省电力有限公司电力科学研究院 | A kind of distribution uplink unloading strategy towards mobile edge calculations |
Non-Patent Citations (2)
| Title |
|---|
| Mathematical modeling and evolutionary generation of rule sets for;Liping Zhang, Qiuhua Tang, Zhengjia Wu, Fang Wang;《Energy》;20171231;全文 * |
| 基于负载感知和QoS的多中心作业调度算法;冯鸣夏，伍卫国，邸德海;《计算机技术与发展》;20181231;第28卷(第12期);全文 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11868812B2 (en) | 2021-08-12 | 2024-01-09 | International Business Machines Corporation | Predictive scaling of container orchestration platforms |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110231976A (en) | 2019-09-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110231976B (en) | Load prediction-based edge computing platform container deployment method and system | |
| CN108829494B (en) | Container cloud platform intelligent resource optimization method based on load prediction | |
| Kaur et al. | Container-as-a-service at the edge: Trade-off between energy efficiency and service availability at fog nano data centers | |
| CN107992353B (en) | Container dynamic migration method and system based on minimum migration volume | |
| CN110825520B (en) | Cluster extremely-fast elastic telescoping method for realizing efficient resource utilization | |
| CN108182105B (en) | Local dynamic migration method and control system based on Docker container technology | |
| CN111966453B (en) | Load balancing method, system, equipment and storage medium | |
| CN107404523A (en) | Cloud platform adaptive resource dispatches system and method | |
| CN104657221A (en) | Multi-queue peak-alternation scheduling model and multi-queue peak-alteration scheduling method based on task classification in cloud computing | |
| CN103152389A (en) | Method and system of responding peak access in cloud computer system | |
| CN113806018A (en) | Kubernetes cluster resource hybrid scheduling method based on neural network and distributed cache | |
| Xia | Resource scheduling for piano teaching system of internet of things based on mobile edge computing | |
| CN110086726A (en) | A method of automatically switching Kubernetes host node | |
| CN112799817A (en) | Micro-service resource scheduling system and method | |
| CN113822456A (en) | Service combination optimization deployment method based on deep reinforcement learning in cloud and mist mixed environment | |
| CN110442428A (en) | The coordination approach of Docker container | |
| CN110990160A (en) | Static security analysis container cloud elastic expansion method based on load prediction | |
| Liu et al. | K‐PSO: An improved PSO‐based container scheduling algorithm for big data applications | |
| Lan et al. | Task partitioning and orchestration on heterogeneous edge platforms: The case of vision applications | |
| CN113360245A (en) | Internet of things equipment task downloading method based on mobile cloud computing deep reinforcement learning | |
| CN116996941A (en) | Calculation force unloading method, device and system based on cooperation of cloud edge ends of distribution network | |
| CN114978913B (en) | Cross-domain deployment method and system for service function chains based on cut chains | |
| Patel et al. | An improved approach for load balancing among heterogeneous resources in computational grids | |
| Tang et al. | Edge computing energy-efficient resource scheduling based on deep reinforcement learning and imitation learning | |
| Du et al. | A combined priority scheduling method for distributed machine learning |
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 |