🌱Aadam's Garden

Search

Search IconIcon to open search

DeepSlice: A Deep Learning Approach towards an Efficient and Reliable Network Slicing in 5G Networks

Last updated Aug 10, 2022

# DeepSlice: A Deep Learning Approach towards an Efficient and Reliable Network Slicing in 5G Networks

Metadata

  • CiteKey:: thantharateDeepSliceDeepLearning2019
  • Type:: conferencePaper
  • Author:: Anurag Thantharate, Rahul Paropkari, Vijay Walunj, Cory Beard
  • Journal:: 2019 IEEE 10th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON)
  • Pages:: 0762-0767
  • Year:: 2019
  • DOI:: 10.1109/UEMCON47517.2019.8993066
  • Format:: PDF

Abstract

Existing cellular communications and the upcoming 5G mobile network requires meeting high-reliability standards, very low latency, higher capacity, more security, and high-speed user connectivity. Mobile operators are looking for a programmable solution that will allow them to accommodate multiple independent tenants on the same physical infrastructure and 5G networks allow for end-to-end network resource allocation using the concept of Network Slicing (NS). Data-driven decision making will be vital in future communication networks due to the traffic explosion and Artificial Intelligence (AI) will accelerate the 5G network performance. In this paper, we have developed a ‘DeepSlice’ model by implementing Deep Learning (DL) Neural Network to manage network load efficiency and network availability, utilizing in-network deep learning and prediction. We use available network Key Performance Indicators (KPIs) to train our model to analyze incoming traffic and predict the network slice for an unknown device type. Intelligent resource allocation allows us to use the available resources on existing network slices efficiently and offer load balancing. Our proposed DeepSlice model will be able to make smart decisions and select the most appropriate network slice, even in case of a network failure.

Files and Links

Tags and Collections

  • Keywords:: 5G, Intelligent Networks, Machine Learning, Network Slicing, Network Slicing Optimization
  • Collections:: UAEU RA

# Annotations

Annotations(6/23/2022, 4:32:25 PM)

==“5G networks allow for end-to-end network resource allocation using the concept of Network Slicing (NS).”== (p. 762)

==“Slicing would allow operators to efficiently run multiple instances of the network over a single infrastructure for serving various applications, use cases, and business services with superior Quality of Service (QoS).”== (p. 762)

==“Network slicing in 5G can cost-effectively deliver multiple logical networks over the same physical infrastructure. SDN and NFV together would allow us to manipulate these slices as and when needed without having to touch multiple different physical equipment in the network.”== (p. 763)

==“DL will perform real-time analysis for any given slice to determine the network performance, create a potential baseline for performance, be proactive in anticipating problems, inspect different network elements, and find out if anything is abnormal.”== (p. 763)

==“A typical consumer would request parameters like data rate, latency, mobility, isolation, power constraints, etc. Accordingly, a specific network slice type is provisioned if the existing network slice instance does not have enough capacity and associated network functions are initiated on demand.”== (p. 763)

==“The key parameters that are determined for network slicing are the slice type, bandwidth, throughput, latency, equipment type, mobility, reliability, isolation, power, etc.”== (p. 763)

==“Other than this work, no other work to our knowledge considers the easily overlooked but difficult problem of deciding which devices and connections should be assigned to which network slices.”== (p. 764)

==“Our dataset includes most relevant KPIs from both the network and the devices, including the type of device used to connect (Smartphone, IoT device, URLLC device, etc.), User Equipment (UE) category, QoS Class Identifier (QCI), packet delay budget, maximum packet loss, time and day of the week, etc. These KPIs can be captured from control packets between the UE and network.”== (p. 764)

General representation of the DeepSlice model, consisting of Network Slices

(p. 764)

Feature highlights of DeepSlice model

(p. 765)

==“Our pre-defined slice categories include enhanced Mobile Broad Band (eMBB), Ultra Reliable Low Latency Communication (URLLC), massive Machine Type Communication (mMTC) and the Master slice. The Master slice is the slice that will have network functions belonging to each of the other slices. It can always act as a back-up slice, in a hotstandby, and will be used depending on the load on other slices.”== (p. 765)

==“In our proposed model, we predict the network load on each network slice based on the previous information of incoming connections and keep track of which output ‘network slice’ is being utilized the most. We then allocate incoming traffic to the network by efficiently distributing them between all the slices as desired.”== (p. 765)

==“The main reason for selecting RF for our model over k-Nearest Neighbor, Naive Bayes, or Decision Tree is simply because of the nature and amount of data we have in our dataset. We have around 65K unique inputs, and all this data is well structured, so RF reduces the risk of overfitting by using multiple sub-trees.”== ==“Most importantly, it estimates any missing data and maintains the accuracy even when some input data is missing.”==

DeepSlice Model

(p. 765)

==“DeepSlice will eventually learn and understand what kind of a device goes to what slice and it will evolve over time to be able to predict future connections requiring a specific service or a network slice.”== (p. 766) I’m not really sure that this is a good enough model/description. Maybe an open-source implementation would’ve helped in understanding, but the current description glosses over many details.

==“Our DeepSlice can realize this overload”== (p. 766) How did they achieve this? Did they implement it in the model somehow by changing the error formula, or did they do it manually later on?

==“Now the DeepSlice will direct all new eMBB related traffic to the master slice and avoid any loss of traffic transmission in the network.”== (p. 767) Again, does the model predicts that? or does the model predicts eMBB but they then see that its down, so they just forward it to the Master Slice?

==“ https://github.com/adtmv7/DeepSlice”== (p. 767) Contains the model code as well.


# Notes

# Problem

To decide which devices and connections should be assigned to which network slices. ^3c219a

# Solution

They used a simple Deep Neural Network, trained on networking data, to predict which connection should be assigned to which network slice.

# Reaction

I don’t really understand/like what they’ve done/proposed here. It seems a very simple DNN model, trained on tabular data, predicting a single column. Based on the code available on the GitHub ^8aad90, it doesn’t seem vey sophisticated as well.

I’m not really sure about the quality of the data as well. It’s not very well-described. They don’t even outline how they gathered the data, using which device/software etc.

# Verdict

I should look for better approaches to understand how to tackle this problem. I should also find a reliable and well-documented dataset for this.

# Code

The code is available at https://github.com/adtmv7/DeepSlice.

By taking a quick glance at it, it doesn’t seem good. Feels like a half-assed work t