intelligent cach ing at the mobile edge...

3
Intelligent Caching at the Mobile Edge Network Kyi Thar, Nguyen H. Tran, and Choong Seon Hong Department of Computer Science and Engineering, Kyung Hee University. Email: {kyithar, nguyenth, cshong}@khu.ac.kr Abstract Caching popular contents at edge nodes (base stations) becomes a promising solution to improve the user s quality of services, as well as to reduce the network traffic. However, to i) correctly predict future popularity of contents, and ii) to efficiently store popular contents are challenging issues. So, in this paper, we proposed learning based caching scheme which is two folds: i) predict the future popularity score of contents using deep learning, and ii) cache the popular video with proactive caching. Then, we develop and train the prediction scheme using Tensorflow libraries, and test the performance of caching scheme with python based simulator. The simulation results show that the proposed scheme outperforms existing algorithms in terms of content hit probability, access delay and etc. 1. Introduction According to the Cisco Visual Networking Index, watching videos from wireless devices has been generating most of the Internet traffic and is forecast to continue to grow exponentially [1]. In order to handle the overwhelming Internet traffic, several future Internet network architectures have been proposed with in-network caching capability[2]. With the help of in-network caching, edge nodes (Base Stations (BSs) and Small-cell Base Stations (SBSs)) temporarily store video contents in their cache to satisfy user requests in the near future, whereas cache capacity of the edge node is limited. So, each edge node should store only popular video contents. Videos popularity can be defined as the ratio of the number of requests for the particular content to the total number of requests from users, usually obtained for a certain region during a given period of time. In reality, it is not easy to know whether the video content is popular or not because the video popularity can temporally and/or spatially vary. Therefore, contents popularity prediction becomes one of the most challenging issues to design an efficient caching scheme. Caching scheme can be classified into two categories; i) reactive caching, and ii) proactive caching. Reactive caching: On the request arrives, the edge node makes a decision whether to store the requested content or not is known as reactive caching [3]. Proactive caching: In the proactive caching scheme, every time period, the edge node stores the set of contents, depending on predicted contents popularity, before requested by users [4]. Our contributions are summarized as follows: We design the system to jointly work with deep neural network based prediction system and mobile edge computing architecture. We utilize Deep Recurrent Neural Network to predict the popularity of video contents with multidimensional sequences of information. We use MovieLens 1M dataset to train and test our proposed scheme. Test the performance of proposed caching scheme using the python based simulator. 2. System Model The system model of proposed scheme is shown in Fig.1, where BSs/SBSs are attached with the cache storage. BSs/SBSs collects the information as log files and send these log files to controller, which is located at the data center, to train the prediction model. Popular videos are stored at BSs/SBs with the help of prediction model. When request arrive, BS/SBS checks the requested video is located in its cache or not. If it is located, serve the content to the user. Otherwise, BS/SBS downloads the content directly from original servers. 2017년 한국소프트웨어종합학술대회 논문집 1197

Upload: others

Post on 18-Jul-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Intelligent Cach ing at the Mobile Edge Networknetworking.khu.ac.kr/layouts/net/publications/data/KSC2017/Intellige… · Intelligent Cach ing at the Mobile Edge Network Kyi Thar,

Intelligent Caching at the Mobile Edge Network

Kyi Thar, Nguyen H. Tran, and Choong Seon Hong

Department of Computer Science and Engineering, Kyung Hee University.

Email: {kyithar, nguyenth, cshong}@khu.ac.kr

Abstract

Caching popular contents at edge nodes (base stations) becomes a promising solution to improve the user’s quality

of services, as well as to reduce the network traffic. However, to i) correctly predict future popularity of contents, and ii)

to efficiently store popular contents are challenging issues. So, in this paper, we proposed learning based caching

scheme which is two folds: i) predict the future popularity score of contents using deep learning, and ii) cache the

popular video with proactive caching. Then, we develop and train the prediction scheme using Tensorflow libraries,

and test the performance of caching scheme with python based simulator. The simulation results show that the

proposed scheme outperforms existing algorithms in terms of content hit probability, access delay and etc.

1. Introduction

According to the Cisco Visual Networking Index,

watching videos from wireless devices has been

generating most of the Internet traffic and is forecast to

continue to grow exponentially [1]. In order to handle the

overwhelming Internet traffic, several future Internet

network architectures have been proposed with in-network

caching capability[2]. With the help of in-network caching,

edge nodes (Base Stations (BSs) and Small-cell Base

Stations (SBSs)) temporarily store video contents in their

cache to satisfy user requests in the near future, whereas

cache capacity of the edge node is limited. So, each edge

node should store only popular video contents. Video’s

popularity can be defined as the ratio of the number of

requests for the particular content to the total number of

requests from users, usually obtained for a certain region

during a given period of time. In reality, it is not easy to

know whether the video content is popular or not because

the video popularity can temporally and/or spatially vary.

Therefore, content’s popularity prediction becomes one of

the most challenging issues to design an efficient caching

scheme. Caching scheme can be classified into two

categories; i) reactive caching, and ii) proactive caching.

Reactive caching: On the request arrives, the edge node

makes a decision whether to store the requested content

or not is known as reactive caching [3]. Proactive caching:

In the proactive caching scheme, every time period, the

edge node stores the set of contents, depending on

predicted content’s popularity, before requested by users

[4].

Our contributions are summarized as follows:

We design the system to jointly work with deep

neural network based prediction system and mobile

edge computing architecture.

We utilize Deep Recurrent Neural Network to

predict the popularity of video contents with

multidimensional sequences of information.

We use MovieLens 1M dataset to train and test our

proposed scheme. Test the performance of

proposed caching scheme using the python based

simulator.

2. System Model

The system model of proposed scheme is shown in Fig.1,

where BSs/SBSs are attached with the cache storage.

BSs/SBSs collects the information as log files and send

these log files to controller, which is located at the data

center, to train the prediction model. Popular videos are

stored at BSs/SBs with the help of prediction model.

When request arrive, BS/SBS checks the requested video

is located in its cache or not. If it is located, serve the

content to the user. Otherwise, BS/SBS downloads the

content directly from original servers.

2017년 한국소프트웨어종합학술대회 논문집

1197

Page 2: Intelligent Cach ing at the Mobile Edge Networknetworking.khu.ac.kr/layouts/net/publications/data/KSC2017/Intellige… · Intelligent Cach ing at the Mobile Edge Network Kyi Thar,

Figure 1 System Model

3. Popularity Score Prediction System Design

The high level prediction system design is shown in Fig.

2, where dotted lines show the training process and dark

lines show the real time prediction process. The prediction

system consists of two parts i) Prediction module for

controller, and ii) Prediction module for BSs/SBSs. The

prediction module for controller is responsible for training

the popularity score prediction and located at the high-end

computing node (e.g. Data Center) because BSs/SBSs do

not have enough capacity to train the deep learning

models. Then, trained model is distributed to all nodes.

The prediction module for node is responsible for

popularity score prediction and cache decision, located at

the BSs/SBSs.

Prediction module for controller: consists of i) Data

Collector and ii) prediction module. Data controller collects

log files which included related information of the contents

and users from BSs/SBSs at every time period t. The

prediction module preprocesses the data from the log files

to get dataset to train the prediction model. Prediction

model can be different type of deep learning model such

as Deep Belief Network (DBN), Deep Convolutional

Neural Network (DCNN) and Deep Recurrent Neural

Network (DRNN). For the popularity score prediction, we

utilize the Recurrent Neural Network (RNN) because the

incoming requests for videos are in a sequential manner.

Among the variants of RNN such as Long Short-Term

Memory (LSTM), Gated Recurrent Unit (GRU), we choose

LSTM [5] as the RNN’s cell for distributed training, which

uses asynchronous stochastic gradient descent optimizing

on large clusters. Finally, the optimized trained model will

be stored in the Trained Model module and transfers the

trained model to the all of the nodes, when training is

finished.

Prediction module for BSs/SBSs: consists of i) Data

Collector, ii) Prediction Module, and iii) Cache Decision.

Prediction Module extracts information and feed those

information to the trained model and gets predicted

popularity scores of videos. Then, the Cache Decision

module utilizes predicted popularity scores to make cache

decision, whether the video content should store or not on

its cache.

4. Performance Evaluations

We use many to one RNN model which has 50 hidden

layers. We choose sigmoid function as the activation

Figure 2. Popularity Score Prediction System Model

2017년 한국소프트웨어종합학술대회 논문집

1198

Page 3: Intelligent Cach ing at the Mobile Edge Networknetworking.khu.ac.kr/layouts/net/publications/data/KSC2017/Intellige… · Intelligent Cach ing at the Mobile Edge Network Kyi Thar,

function. We choose root mean square error as loss

function and Adam optimizer to minimize the training loss.

We developed the python based chunk level simulator to

simulate our proposed scheme. Fig.3 shows the root

mean square error comparison between LSTM, GRU and

Simple RNN, where the LSTM gives the smallest root

mean square error. Then, in Fig.4, we compare the

caching performance of proposed scheme with LSTM,

GRU and Simple RNN, where LSTM gives the highest

probability of cache hit among others.

5. Conclusion

In this paper, we proposed the scheme to predict the

popularity score of each content by utilizing deep

recurrent neural network. Then, we proposed the caching

scheme which utilizes output (popularity score) of the

deep recurrent neural network to make a cache decision,

where each node their own cache decision independently.

Finally, we developed the prediction model by using

Tensorflow and train the model with MoveLens dataset.

As for the future work, we will apply the reinforcement

learning with the recurrent neural network for the video

contents predicting.

Acknowledgment

This work was supported by the National Research

Foundation of Korea(NRF) grant funded by the Korea

government(MSIT) (NRF- 2017R1A2A2A05000995). *Dr.

CS Hong is the corresponding author.

This work was supported by Institute for Information &

communications Technology Promotion(IITP) grant

funded by the Korea government(MSIT) (No. 2013-0-

00409, Research and Development of 5G Mobile

Communications Technologies using CCN-based Multi-

dimensional Scalability) *Dr. CS Hong is the

corresponding author.

References

[1] http://www.cisco.com/c/en/us/solutions/serviceprovider/visual-

networking-index-vni/index.html.

[2] Thar, Kyi, et al. "Hybrid caching and requests forwarding in

information centric networking." Network Operations and

Management Symposium (APNOMS), 2015 17th Asia-Pacific.

IEEE, 2015.

[3] Suoheng Li, Jie Xu, Mihaela van der Schaar, and Weiping Li.

Trendaware video caching through online learning. IEEE

Transactions on Multimedia, 18(12):2503–2516, 2016.

[4] Mingzhe Chen, Walid Saad, Changchuan Yin, and Merouane

Deb bah. Echo state networks for proactive caching in cloud-

based radio access networks with mobile users. IEEE

Transactions on Wireless Communications, 16(6):3520–3535,

2017.

[5] Has¸im Sak, Andrew Senior, and Franc¸oise Beaufays. Long

short-term memory recurrent neural network architectures for

large scale acoustic modeling. In Fifteenth Annual Conference of

the International Speech Communication Association, 2014.

Figure 3. Root Mean Squared Error Comparison Figure 4. Probability of Cache Hit Comparison

2017년 한국소프트웨어종합학술대회 논문집

1199