1. The document proposes a connectionist approach using graph neural networks for dynamic resource management in virtualized network functions to improve efficiency while ensuring reliability.
2. It aims to predict VNF resource requirements to avoid unnecessary resources being kept active/standby through topology-aware resource management using a GNN model.
3. The GNN model takes VNFC features and states as input and computes VNFC states over iterations before outputting resource forecasts to dynamically manage resources.
Introduction to Prompt Engineering (Focusing on ChatGPT)
A Connectionist Approach to Dynamic Resource Management for Virtualised Network Functions
1. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
A Connectionist Approach to Dynamic Resource
Management for Virtualised Network Functions
Rashid Mijumbi∗, Sidhant Hasija∗, Steven Davy∗, Alan Davy∗,
Brendan Jennings∗ and Raouf Boutaba†
∗Telecommunications Software and Systems Group, Waterford Institute of
Technology, Ireland
†D.R. Cheriton School of Computer Science, University of Waterloo, Waterloo,
Ontario, N2L 3G1, Canada
Montreal, Canada, November 1, 2016
2. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Presentation Outline
1 Introduction: Network Functions Virtualisation
2 Problem: Efficient vs Reliable Resource Management
3 Proposed Approach: Graph Neural Networks
4 Solution Model: GNN-based Dynamic Resource Management
5 Performance Evaluation
6 Conclusion and Future Work
3. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Network Functions Virtualisation
Increasing CAPEX and OPEX
The short lifetime of the NAs leads to increased Capital
Expenses (CAPEXs).
When NAs are specialised, they require specialised
maintenance and limits flexibility, leading to increased
Operating Expenses (OPEXs).
Declining Revenues
Competition with over-the-top providers
Inability to quickly provide new services
Separation between infrastructure and Service
Optimization of resource Usage and routing beyond BGP
4. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Network Functions Virtualisation
Physical Resources
Virtual Resources
Services
Network Function Virtualization Infrastructure
ManagementandOrchestration
Computing, Storage, Network Resources
Virtual Network Functions
Computing, Storage, Network Resources
ManagementandOrchestration
VNF 1 VNF 2 VNF 3 VNF n. . .
.
Source: R. Mijumbi, J. Serrat, J. L. Gorricho, N. Bouten, F. De Turck, R. Boutaba, ”Network Function
Virtualization: State-of-the-art and Research Challenges”, IEEE Communications Surveys and Tutorials. 2016.
5. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Problem: Efficiency vs Reliability
NFV Essential for 5G, Supporting Critical Applications
NFV will be an important building block for 5G
5G is expected to support critical infrastructure
Efficiency and reliability are important KPIs for 5G
Source: http://telematicswire.net/ec-plans-future-of-5g-for-automotive/
6. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
State-of-the-art
1 High VM provisioning time threatens reliability in critical
applications such as M2M
1
10
100
1000
1 2 4 8 16 32
Number of Virtual Machines
TotalProvisioningTime(s)
Eucalyptus OpenStack OpenNebula
.
Adapted from: Mike Jones et al. ”Scalability of VM Provisioning Systems”, 20th Annual IEEE High Performance
Extreme Computing Conference(HPEC), September 2016, Waltham, MA USA.
7. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Our Proposal
Objective
Predict VNF Resource Requirements so as:
To avoid resources are not unnecessarily kept active/standby
While ensuring reliable performance
Idea
Topology-aware Resource Management
Motivation: VNFC Dependencies
Virtualization
container such as a
VM
VNFC 1
VNF 1 VNF 2
VNFC 1
VNFC 2 VNFC 3
VNF 4
VNFC 1
VNF 3
Service Function Chain based on Virtualised Network Functions
VNFC 3 VNFC 4
VNFC 2 VNFC 5
VNFC 1VNFC 1
8. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Graph Neural Networks (GNN)
A supervised learning model aimed at solving problems in the
graphical domain.
Node, 𝑛4 Node, 𝑛3
Node, 𝑛1 Node, 𝑛2
Node, n
VNFC Features, 𝑓𝑛 Neighbourhood, , 𝑛∗
Using fn and n , a state sn, and an output on for each node n
are determined using equations (1) and (2) respectively.
sn =
m∈n
hw fn, fm, sm , ∀n (1)
on = gw sn, fn , ∀n (2)
9. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
GNN-based Dynamic Resource Management
Features 𝑓𝑛
of VNFC
Features 𝑓𝑚
of all VNFC’s Neighbours
ℎ 𝑤 𝑔 𝑤
VNFC State
𝑠 𝑛
Output
(Resource
Forecast)
FNN FNN
States 𝑠 𝑚
of all Neighbours
VNFC States
SFC Features
Output
Computation
State
Computation
3 4
1
2
Comprised of four main components: (1) SFC features, (2)
VNFC states, (3) state computation, and (4) output
computation.
10. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
SFC Features
Observations or monitoring data from the VNFCs
Include network parameters (such as CPU or RAM utilisation
levels) that can be measured.
fn =
cn
mn
dn
(3)
fnm =
bnm
dnm
(4)
-
SFC modelled as a directed graph G(N, L)
Virtualization
container such as a
VM
VNFC 1
VNF 1
VNFC 3
VNFC 2
𝑛2
𝑛3 𝑛4
𝑛5
𝑙32
𝑙21
𝑙14
𝑙46𝑙31
𝑙15
𝑙51
𝑛0
𝑛7
𝑛8
VNF 1
VNF 2
VNF 3
VNF 4
𝑛5
𝑙41
𝑛6𝑛0 𝑛1
𝑙12
𝑙13
𝑙23
𝑙03
VNFC 1
17. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Implementation Setup
Bono Sprout
Ralf Homer Homestead
HSS Mirror
cassandra
XDMS
cassandra
Rf CTF
memcached
I/S-CSCF BGCF
memcached
P-CSCF, WebRTC
Clearwater
virtualised IMS
SNMPUEs
SIPp
GNN-based Dynamic
Resource
Management
DNS
Heat Orchestration
SIP
CACTI
Monitoring
SUT
18. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Evaluation Details
Setup Parameters and Comparisons
1 100K Users, Call initiation/end based on Poisson/Exponential,
2 Each call transmits media extracted from real Skype traffic
traces
3 All VNFCs polled every 15s, History/Forecasting is 20
episodes,
4 Experiment 1: 10,000 data points for training FNNs
5 Experiment 2: Trained System used to determine accuracy on
1,000 measurements
6 Experiment 3: Predictions used to effect resource allocations
(Spin-up at 40%, Spin down at 20%)
7 Comparisons: Static, Manual, Automated
19. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Evaluations (1)
0
10
20
30
40
50
60
0 200 400 600 800 1000
RMSE
Training Iteration, each involving 10,000 examples
Ralf Bono Sprout
Homestead Homer Total
0.00
0.20
0.40
0.60
0.80
1.00
0 200 400 600 800 1000
%CPUUtlisation
Test Number
Actual Output Expected Output
0.00
0.50
1.00
1.50
2.00
2.50
3.00
0 200 400 600 800 1000
Delay(ms)
Test Number
Actual Output Expected Output
0.5
0.6
0.7
0.8
0.9
1
0 200 400 600 800 1000
%DelayPredictionError
Test Number
Error 100 period Mov. Avg.
20. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Evaluations (2)
0
0.2
0.4
0.6
0.8
1
100 400 700 1000
%CPUUtilisation
Test Number
Static Manual Automated
0.00
0.50
1.00
1.50
2.00
2.50
3.00
100 400 700 1000
Delay(ms)
Test Number
Static Manual Automated
0
2
4
6
8
100 400 700 1000
DroppedCallsThousands
Test Number
Static Manual Automated
0
10
20
30
40
100 400 700 1000
DroppedCallsThousands
Test Number
Static Manual Automated
21. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Conclusion and Future Work
Conclusion
Topology-aware approach to automated and dynamic
resource management approach for NFV environments.
Implemented in a real environment involving a virtualised
IMS, and using real VoIP traces,
Prediction accuracy of about 90%, and enhance the
processing delay and call drop rate by 29% and 27%
respectively.
Future Work
Improve generalisation accuracy by considering error functions
with different penalty terms.
More efficient ways of training the SFC encoding network.
22. Introduction Problem Proposed Approach Solution Model Evaluation Conclusion
Thank You
THANK YOU!
Contact: rmijumbi@tssg.org