CN109922007B - Load balancing method based on convolutional neural network - Google Patents
Load balancing method based on convolutional neural network Download PDFInfo
- Publication number
- CN109922007B CN109922007B CN201910036377.5A CN201910036377A CN109922007B CN 109922007 B CN109922007 B CN 109922007B CN 201910036377 A CN201910036377 A CN 201910036377A CN 109922007 B CN109922007 B CN 109922007B
- Authority
- CN
- China
- Prior art keywords
- neural network
- convolutional neural
- communication
- load balancing
- communication link
- 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
Abstract
The invention discloses a load balancing method based on a convolutional neural network, which comprises the following steps: acquiring time delay, speed and packet loss rate of a plurality of operator communication links, and respectively using the time delay, speed and packet loss rate as input samples of a convolutional neural network; constructing an input sample as a training set, and training and constructing a convolutional neural network by using the constructed training set; according to the trained convolutional neural network, the quality of each communication link in each time period is respectively judged, and the quality of the communication quality of each communication link is obtained; and carrying out load balancing on a plurality of communication links with known good and bad communication quality to realize the stability of communication. The invention combines the convolutional neural network with the load balancing, effectively judges the communication quality of a plurality of communication links, realizes the load balancing and improves the stability of communication.
Description
Technical Field
The invention belongs to the technical field of communication, and particularly relates to a load balancing method based on a convolutional neural network.
Background
With the rapid development of the internet, the user demand gradually increases, and the overload and overload of the network are common. Therefore, multilink transmission of data becomes inevitable. The use of multiple links to transmit data requires the study of load balancing methods for the multiple links.
Load balancing is established on the basis of the existing network structure, and an effective method is provided for expanding the bandwidth of network equipment and a server, increasing the throughput, strengthening the network data processing capacity and improving the flexibility and the usability of the network.
Currently, common load balancing processing methods include a polling method, a hashing method, a minimum connection method, and the like. The traditional load balancing method has low efficiency. The current research is based on other single variables such as time delay or throughput, and the like, and is in the face of one aspect. And the bottleneck of line lease bandwidth among different network operators causes that the network inter-access is not smooth and the user experience is not good due to prolonged time and high packet loss rate in the network inter-access process of the cross-operator.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a load balancing method based on a convolutional neural network, and solves the problem that communication is unstable in the network inter-access process of cross operators in the prior art.
A load balancing method based on a convolutional neural network comprises the following steps:
the method comprises the following steps: acquiring data of three quality parameters of a communication link which is carrying out communication data transmission for multiple times as a training set;
step two: training the convolutional neural network by using the obtained training set to obtain a trained convolutional neural network;
step three: acquiring data of three quality parameters of each communication link to be equalized in a plurality of communication links to be equalized which are in communication data transmission, training the trained convolutional neural network, and respectively obtaining probability values p of ith quality parameters of nth communication link to be equalizedni；
Step four: quality probability xi of nth communication link to be equalizednBy passing
And comparing the communication quality probabilities of the communication links to be balanced to distribute the data transmitted by the communication links to be balanced, so as to realize load balancing.
Further, the three quality parameters of the communication link include a time delay, a rate and a packet loss rate.
Further, the pooling layer of the convolutional neural network in the step two adopts maximum pooling
Further, in the second step, the output layer of the convolutional neural network adopts a softmax classifier, and the output is as follows:
y＝softmax(g*z+b0)
g is the weight to be learned, z is the feature vector, b0Is an offset.
Further, the probability value p of the ith quality parameter of the nth communication link to be equalized in the third stepniThe method comprises the following steps:
wherein y isiAnd an output vector y, i of an output layer of the convolutional neural network representing the ith quality parameter in the nth communication link to be equalized is 1, 2 and 3.
Further, the communication link is a 4G wireless communication link.
Furthermore, the convolutional neural network comprises an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer.
Compared with the prior art, the invention has the beneficial effects that:
(1) the invention realizes the load balance among a plurality of communication links of a plurality of operators through the convolutional neural network.
(2) The invention judges the load balance of the communication link through three link quality parameters of time delay, speed and packet loss rate of a plurality of links, thereby improving the stability of transmission.
(3) The invention trains the training set through the convolutional neural network, dynamically judges the transmission quality of a plurality of communication links and simultaneously realizes load balance.
Drawings
FIG. 1 is a schematic diagram of the structure of a convolutional neural network;
fig. 2 is a flow chart of a convolutional neural network-based load balancing method.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific examples described herein are intended to be illustrative only and are not intended to be limiting.
The communication link adopted in this embodiment refers to a 4G wireless communication link of an operator such as telecommunications, communications, and mobile, and the quality parameters include a time delay, a rate, and a packet loss rate. The adopted convolutional neural network model is essentially mapping from input to output, and mapping capability in the network is constructed by learning a large number of mapping relations between input and output. The convolutional neural network is combined with a load balancing method, and the convolutional neural network is constructed through a training set, so that the load balancing of multiple links of multiple operators is realized.
A load balancing method based on convolutional neural network, as shown in fig. 1, includes the following steps:
the method comprises the following steps: acquiring data of three quality parameters of a communication link which is carrying out communication data transmission for multiple times as a training set;
first, in a 4G wireless communication link in which communication data is being transmitted, an ICMP (Internet Control protocols) request message of a certain time length is sent to a destination host by a local host, and the local host obtains a response, so that data of quality parameters, such as time delay, rate and packet loss rate, of three communication links can be obtained, where the certain time length is set to 5 seconds in this embodiment.
The time delay and the packet loss rate can be directly obtained, and the calculation process of the rate upsilon is as follows:
send N to destination host1Byte ICMP echo request, recording the time T from sending request to receiving data returned from destination host1；
Send N to destination host2Byte ICMP echo request, recording the time T from sending request to receiving data returned from destination host2；
The rate v of the current communication link is:
υ＝(N2-N1)/(T2-T1) (1)
the obtained quality parameters of the three communication links, namely the time delay, the speed and the packet loss rate of each link respectively form a group of training samples. Training samples are obtained multiple times according to time within a period of time (such as 12 hours), and multiple groups of training samples of each communication link respectively form a training set.
Step two: training the convolutional neural network by using the obtained training set to obtain a trained convolutional neural network;
fig. 2 is a schematic structural diagram of a convolutional neural network, which is divided into an input layer, a convolutional layer, a pooling layer, a full-link layer, and an output layer.
The input layer is time delay, speed and packet loss rate data respectively acquired by any one of telecommunication, Unicom and mobile communication links. Simultaneously, inputting by adopting a 3 x 8640 sample format, wherein 3 represents three parameters of time delay, speed and packet loss rate, and 8640 represents the total number of data obtained in the acquisition time;
the number of convolutional layers and the number of convolutional kernels of each layer are determined according to actual conditions, the width of each convolutional kernel is the same as that of an input matrix of the layer, W is set, a sample matrix of a convolutional training set is X, the height of each convolutional kernel is h, R represents a real number, and the convolutional kernels are represented as g e RhWConvolving the feature a extracted at the sample matrix (x, y)jComprises the following steps:
aj＝s(g*Xx,y+b0) (2)
where s is a Sigmoid function and,
b0is an offset. Sliding the convolution kernel left and right on a sample matrix with the length of L, wherein the value of j is constantly changed, namely j is 1, 21,a2,...,aL-h+l]In this embodiment, L is 8640.
The above process is only one convolution kernel to generate one feature, and the actual model has a plurality of convolution layers in calculation, each convolution layer has a plurality of convolution kernels, and in this example, 15 convolution kernels with the size of 3 × 3 are adopted.
Wherein the pooling layer is used to reduce the eigenvectors output by the convolutional layer by pooling. In this embodiment, maximum pooling is employed, i.e.
Wherein, the full connection layer converts the two-dimensional characteristic diagram output after convolution pooling into a one-dimensional vector, and the generated characteristic vector is expressed as
Wherein, the output layer adopts a softmax classifier, wherein the neuron activation value of the output layer is 1, which indicates that the input data corresponds to the communication quality condition of the link, and when the neuron activation value is not activated, the neuron activation value is 0, and the output result is:
y＝softmax(g*z+b0) (3)
wherein g is the weight and feature vector of the learning
b0Is an offset.
Step three: acquiring data of three quality parameters of each communication link to be equalized in a plurality of communication links to be equalized which are in communication data transmission, training the trained convolutional neural network, and respectively obtaining probability values p of ith quality parameters of nth communication link to be equalizedni；
And sending an ICMP echo request message to a target host to acquire the time delay, the speed and the packet loss rate of the communication link to be equalized at the current time so as to obtain the input of the convolutional neural network.
Outputting the probability of the quality parameters of the three links of the nth communication link by the convolutional neural network model trained in the step two:
wherein p isniThe probability value of the ith quality parameter in the nth to-be-equalized communication link is represented, in this embodiment, the three quality parameters are time delay, rate and packet loss rate, and correspond to p respectively1，p2，p3，yiAn output vector y of the output layer of the convolutional neural network representing the ith quality parameter.
The larger the probability value is, the better the effect is on the communication link quality parameter. And dividing the communication quality of the communication link according to the obtained probability value. And mapping the quality of the communication link into a 3-dimensional vector, namely a good vector, a medium vector and a poor vector.
In this example, the communication quality of the link is determined based on the probability values of the quality parameters of the three communication links that are output. When the probability values of the three link quality parameters are all larger than or equal to 0.8, the communication quality of the communication link is excellent; when the probability values of the three link quality parameters are less than or equal to 0.3, indicating that the communication quality of the given communication link is poor; when the probability values of the three link quality parameters are judged not to be good or poor, the communication quality of the link is judged to be medium.
The probability value of the quality parameter of one communication link to be balanced and the link communication quality are obtained in the same way, and similarly, the link quality parameter data of the other two communication links can be obtained, and the probability values of the quality parameters of the other communication links to be balanced and the link communication quality are obtained through the training of the convolutional neural network.
Step four: quality probability xi of nth communication link to be equalizednBy passing
And comparing the communication quality probabilities of the communication links to be balanced to distribute the data transmitted by the communication links to be balanced, so as to realize load balancing.
Specifically, the probability values of the quality parameters of the three communication links are obtained through the third step, and the number of communication is countedAccording to the load balance, the communication quality probability xi of the nth link is calculatedn：
In which ξnRepresenting the communication quality probability, p, of the nth communication linkniThe probability values of the three communication quality parameters of the nth link are obtained according to the convolutional neural network, and i is 1, 2 and 3.
In the embodiment, three telecommunication, Unicom and mobile communication links, namely n ═ 1, 2 and 3, are adopted, and the sum S ═ xi of the communication quality probability is calculated according to the obtained communication quality probability xi of the three telecommunication, Unicom and mobile links1+ξ2+ξ3. Data of telecommunication, communication and moving three links is processed according to xinThe ratio of/S is distributed to three links. Thereby achieving load balancing of the three links. Wherein the data transmission is in units of data packets. Compared with the prior art, the bandwidth of the network equipment and the server is expanded, the throughput is increased, the network data processing capacity is enhanced, and the flexibility and the usability of the network are improved. In the process of network inter-access of cross operators, the network inter-access is smooth, and the user experience is improved.
The foregoing description is only an example of the present invention and is not intended to limit the present invention, and it will be apparent to those skilled in the art that various modifications and variations in form and detail can be made without departing from the principle and structure of the invention, but these modifications and variations are within the scope of the invention as defined in the appended claims.
Claims (6)
1. A load balancing method based on a convolutional neural network is characterized by comprising the following steps:
the method comprises the following steps: acquiring data of three quality parameters of a communication link which is carrying out communication data transmission for multiple times as a training set;
step two: training the convolutional neural network by using the obtained training set to obtain a trained convolutional neural network;
step three: acquiring data of three quality parameters of each communication link to be equalized in a plurality of communication links to be equalized which are in communication data transmission, training the trained convolutional neural network, and respectively obtaining probability values p of ith quality parameters of nth communication link to be equalizedni；
Probability value p of ith quality parameter of nth communication link to be balancedniThe method comprises the following steps:
wherein y isiAn output vector y, i of an output layer of the convolutional neural network representing the ith quality parameter in the nth communication link to be equalized is 1, 2 and 3;
step four: quality probability xi of nth communication link to be equalizednBy passing
And comparing the communication quality probabilities of the communication links to be balanced to distribute the data transmitted by the communication links to be balanced, so as to realize load balancing.
2. The convolutional neural network-based load balancing method of claim 1, wherein the three quality parameters of the communication link include delay, rate and packet loss rate.
3. The convolutional neural network-based load balancing method as claimed in claim 1, wherein the pooling layer of the convolutional neural network of step two employs maximum pooling
4. The convolutional neural network-based load balancing method according to claim 1, wherein in the second step, the output layer of the convolutional neural network adopts a softmax classifier, and the output is:
y＝softmax(g*z+b0)
g is the weight to be learned, z is the feature vector, b0Is an offset.
5. The convolutional neural network-based load balancing method of claim 1, wherein the communication link is a 4G wireless communication link.
6. The convolutional neural network-based load balancing method of claim 1, wherein the convolutional neural network comprises an input layer, a convolutional layer, a pooling layer, a fully-connected layer, and an output layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910036377.5A CN109922007B (en) | 2019-01-15 | 2019-01-15 | Load balancing method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910036377.5A CN109922007B (en) | 2019-01-15 | 2019-01-15 | Load balancing method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109922007A CN109922007A (en) | 2019-06-21 |
| CN109922007B true CN109922007B (en) | 2022-04-01 |
Family
ID=66960424
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910036377.5A Active CN109922007B (en) | 2019-01-15 | 2019-01-15 | Load balancing method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109922007B (en) |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101695050A (en) * | 2009-10-19 | 2010-04-14 | 浪潮电子信息产业股份有限公司 | Dynamic load balancing method based on self-adapting prediction of network flow |
| US11477666B2 (en) * | 2017-02-16 | 2022-10-18 | University College Dublin National University Of Ireland, Dublin | Methods and systems for network self-optimization using deep learning |
| US10795836B2 (en) * | 2017-04-17 | 2020-10-06 | Microsoft Technology Licensing, Llc | Data processing performance enhancement for neural networks using a virtualized data iterator |
| CN108809839B (en) * | 2018-07-17 | 2020-12-01 | 湖南理工学院 | Wireless Mesh backbone network flow control method and device |
-
2019
- 2019-01-15 CN CN201910036377.5A patent/CN109922007B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109922007A (en) | 2019-06-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2021227508A1 (en) | Deep reinforcement learning-based industrial 5g dynamic multi-priority multi-access method | |
| CN111447083A (en) | Federal learning framework under dynamic bandwidth and unreliable network and compression algorithm thereof | |
| CN111563275B (en) | Data desensitization method based on generation countermeasure network | |
| CN110460880B (en) | Industrial wireless streaming media self-adaptive transmission method based on particle swarm and neural network | |
| CN112532530B (en) | Method and device for adjusting congestion notification information | |
| JP7451689B2 (en) | Network congestion processing method, model update method, and related devices | |
| CN112040512B (en) | Mist computing task unloading method and system based on fairness | |
| CN113132490A (en) | MQTT protocol QoS mechanism selection scheme based on reinforcement learning | |
| CN107749827A (en) | Method for controlling network congestion, apparatus and system based on network state classification | |
| CN113676357B (en) | Decision method for edge data processing in power internet of things and application thereof | |
| CN113490239B (en) | Heterogeneous wireless link concurrent transmission control method based on adaptive network coding | |
| JP7007669B2 (en) | Communication system, traffic control device and traffic control method | |
| CN109922007B (en) | Load balancing method based on convolutional neural network | |
| CN114448899A (en) | Method for balancing network load of data center | |
| CN112529148A (en) | Intelligent QoS inference method based on graph neural network | |
| CN109672626B (en) | Service aggregation method based on queuing delay utilization | |
| CN115002031B (en) | Federal learning network flow classification model training method, model and classification method based on unbalanced data distribution | |
| CN113542121B (en) | Tree-shaped data center link layer load balancing routing method based on annealing method | |
| CN111669777B (en) | Mobile communication system intelligent prediction method based on improved convolutional neural network | |
| CN110177019B (en) | Control method for slowing down network congestion | |
| CN114116052A (en) | Edge calculation method and device | |
| CN114221897A (en) | Routing method, device, equipment and medium based on multi-attribute decision | |
| Wang et al. | A cross-layer adaptation scheme for improving IEEE 802.11 e QoS by learning | |
| CN116760777B (en) | Multipath congestion control method based on ABEA3C | |
| CN115102767B (en) | DDoS active defense system and method based on distributed collaborative 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 |