5GArch 2017: 4th International Workshop on 5G Architecture

Analytical Modeling for

Virtualized Network
Functions
Jonathan Prados-Garzon, Pablo Ameigeiras, Juan J. Ramos-Munoz,
Pilar Andres-Maldonado, and Juan M. Lopez-Soler
Department of Signal Theory, Telematics and Communications
University of Granada, Spain

Presenter: Jonathan Prados-Garzon

jpg@ugr.es
Departamento de Teoría de la Señal,
Telemática y Comunicaciones
Outline
1. Introduction 6. Experimental Setup
2. Related Works 7. Results
3. Assumptions 8. Conclusions
4. Analytical Model 9. Future Work
5. Use case: a vMME with a
three-tier design

University of Granada - Analytical Modeling for Virtualized

Network Functions
Introduction
NFV is a key enabler of the future mobile 5G networks.
In NFV, commodity hardware executes network functions, so
called as Virtualized Network Functions (VNFs).
VNFs might consist of several VNF Components (VNFCs)
working together and they admit varying internal designs and
operations.
NFV entails numerous advantages:
Lower costs.
Greater flexibility and scalability.
Easier and accelerated deployment of new services.
University of Granada - Analytical Modeling for Virtualized Network Functions 3/19
Introduction
Motivation

There is a lack of analytical models for VNFs and chains of

VNFs.
An analytical model for assesing the performance metrics of the
VNFs has the following and interesting applications:
Dynamic Resource Provisioning (DRP).
Performance prediction.
Identifying bottlenecks.
Request policing.

University of Granada - Analytical Modeling for Virtualized Network Functions 4/19

Introduction
Objective

Our objective is to propose a mathematical model for assessing

the performance metrics of a VNF.
This work is a first step towards the above goal ↦ In this work:
We propose an analytical model based on an open queueing network
of G/G/m queues to model VNFs and validate it by simulation.
We compare our model with other baseline models in terms of
computational complexity and estimation error.

University of Granada - Analytical Modeling for Virtualized Network Functions 5/19

Related Works
Work(s) Description Pros Cons

Urgaonkar et al. [5] • Tailored for Multi-tier ✓ High flexibility and/or  Its computational
Internet Services generality. complexity depends on the
• Closed queueing network, ✓ Reduced set of input number of active sessions
MVA based model paramters. O(K NU).
 A packet flow utilizes the
resources at exactly one
tier [5].
Bi et al. [14] • DRP for virtualized Multi- ✓ Its computational  Limited flexibility and/or
tier Internet Apps. complexity is O(K). generality.
• Open queueing network, ✓ Reduced set of input
Jackson’s network based parameters.
model ✓ Resources held
simultaneously at multiple
Prados et al. [6] • Dimensioning for a vMME
tiers.
with a three-tier design.
• Open queueing network,
Jackson’s network based
model

University of Granada - Analytical Modeling for Virtualized Network Functions 6/19

Assumptions
All VMs, hosted on the same physical machine do not share any computational resource.
Two instances of the same VNFC might offer distinct performance (heterogeneity of
hardware).
The packets are served following a First-Come, First-Served (FCFS) discipline.
The mean and variance of the service time (𝑠 = 𝜇−1 and 𝜎𝑠2 ) for a VNFC instance is given
by its computational resource acting as the bottleneck.
Generic arrival and service processes.

VNFC2
VNFC3
VNFC1 VNFC2 FE
VNFC3 DB
VNFC2 W
VNF
University of Granada - Analytical Modeling for Virtualized Network Functions 7/19
Analytical model
Queueing model
To model a VNF we employ an open
network of K G/G/m queues
𝑄1 , 𝑄2 , … , 𝑄𝐾 .
Each queue represents a VNFC
instance running on a VM.
Each server of a queue represents an
instance of the computational resource
acting as the bottleneck.
For the transitions between queues, we
assume probabilistic routing. This
approach allows us to model the path
of any packet flow. Figure: A VNF with a three-tier design.

University of Granada - Analytical Modeling for Virtualized Network Functions 8/19

Analytical model
Methodology of analysis: QNA

QNA (Queueing Network Analyzer),

Notation Description proposed by Whitt in [7], is an
λ0𝑘 Mean external arrival rate at queue Qk
approximate method to estimate
2
the mean and variance of the
𝑐0𝑘 SCV of the external arrival process at queue Qk
system response time of a
m𝑘 Number of servers at queue Qk queueing network.
μ𝑘 Average service rate at queue Qk QNA method can be seen as a
2
𝑐𝑠𝑘 SCV of the service process at queue Qk generalization of the open
𝐾 (𝑗) Number of instances of the j-th VNFC
Jackson’s network of M/M/m
queues to an open Jackson’s
𝑃 = [𝑝𝑖𝑘 ] Routing probability matrix network of G/G/m queues [7].
ν𝑘 multiplicative factor for the flow leaving Qk
In fact, QNA provides exact results
Table I: Model Input parameters. for Jackson’s networks.

University of Granada - Analytical Modeling for Virtualized Network Functions 9/19

Analytical model
Methodology of analysis: QNA

Step 1: Internal flow parameters computation.

To compute the mean arrival rates and SCVs of the arrival process to
each queue. 𝐾 𝐾
λ𝑘 = 𝜆0𝑘 + ෍ 𝜆𝑖 𝜈𝑖 𝑝𝑖𝑘 2 2
𝑐𝑎𝑘 = 𝑎𝑘 + ෍ 𝑐𝑎𝑖 𝑏𝑖𝑘
𝑖=1 𝑖=1

Step 2: Analyze the different queues in isolation as standard

G/G/m queues. 𝜌 𝑐 +𝑐 𝛽 𝑐 +𝑐 𝐶 𝑚2,
𝜆
𝜇 2 2 2 𝑘
𝑘
𝑘 𝑎𝑘 𝑠𝑘 𝑎𝑘 𝑠𝑘 𝑘
𝐼𝑓 𝑄𝑘 𝑖𝑠 𝐺/𝐺/1 𝑊𝑘 = 𝐼𝑓 𝑄𝑘 𝑖𝑠 𝐺/𝐺/𝑚 𝑊𝑘 =
2𝜇𝑘 (1 − 𝜌𝑘 ) 2 𝑚𝑘 𝜇𝑘 − 𝜆𝑘

Step 3: Compute the global performance parameters assuming

queues stochastically independent. 1 𝜆
𝐾
𝑘
𝑇 = ෍ 𝑊𝑘 + 𝑉𝑘 𝑉𝑘 = 𝐾
𝜇𝑘 σ𝑖=1 𝜆0𝑘
𝑖=1

University of Granada - Analytical Modeling for Virtualized Network Functions 10/19

Use case: a vMME with a three-tier
design
We only consider Service Request (SR), S1-Release
(SRR), and X2 Handover (HR) LTE control procedures.
Then, the average number of packets per control packet to be
processed by the MME is given by:
3𝜆𝑆𝑅 + 3𝜆𝑆𝑅𝑅 + 2𝜆𝐻𝑅
ഥ𝑝𝑝 =
𝑁
𝜆𝑆𝑅 + 𝜆𝑆𝑅𝑅 + 𝜆𝐻𝑅
Let us assume a Poisson arrival process. Let 𝑁𝑈 denote the
number of UEs generating signaling workload:
2
𝜆0𝐹𝐸 = 𝑁𝑈 3𝜆𝑆𝑅 + 3𝜆𝑆𝑅𝑅 + 2𝜆𝐻𝑅 𝑐𝑎𝐹𝐸 =1
We consider deterministic service processes for each VNFC.
2 2
c𝑠𝐹𝐸 =0 𝑐𝑠𝐷𝐵 =0
However, the service time might be different for each control
message type at Worker VNFC .
2
−1 (𝐻)
𝑁𝐼 (𝐻) 𝑁𝐼 𝐻
2
𝜎𝑠𝑊
𝑠𝑊 = 𝜇𝑊 = ෍𝑓 2
2
𝜎sW = ෍𝑓 𝐻
− 𝑠𝑊 c𝑠𝑊 = 2
𝑟𝑐𝑝𝑢 𝑟𝑐𝑝𝑢 𝑠𝑊
𝐻 𝐻 Figure: A vMME with a three-tier
design.
University of Granada - Analytical Modeling for Virtualized Network Functions 11/19
Use case: a vMME with a three-tier
design
Regarding the packet flow modeling, which is given by the VNF operation:
There is no multiple packets creation or packet dropping at any VNFC, then:
𝜈𝑘 = 1 ∀ 𝑘 ∈ [1, 𝐾]

For simplicity, we assume only one instance for the FE and DB:
1−𝑝0𝐹𝐸 1−𝑝0𝐹𝐸
0 ⋯ 0
𝐾𝑊 𝐾𝑊
𝑁ഥ𝑝𝑝 2
ഥ𝑝𝑝 +2
0 … 0 ഥ𝑝𝑝 +2
𝑁 𝑁
1
𝑃= ⋮ ⋮ ⋮ ⋮ ⋮ 𝑝0𝐹𝐸 = 2 or, equivalently, ഥ𝑝𝑝 ; 𝑉𝑊 =
𝑉𝐹𝐸 = 2; 𝑉𝐷𝐵 = 2/𝑁 ഥ𝑝𝑝 )
(1 + 2/𝑁
ഥ𝑝𝑝 𝐾𝑊
𝑁 2
ഥ𝑝𝑝 +2
0 … 0 ഥ
𝑁 𝑁𝑝𝑝 +2
1 1
0 … 0
𝐾𝑊 𝐾𝑊

University of Granada - Analytical Modeling for Virtualized Network Functions 12/19

Experimental Setup
Signaling Rates per UE To validate our model, we use two
λ𝑆𝑅 (𝑃𝑜𝑖𝑠𝑠𝑠𝑜𝑛) 0,0045 procs/sec software tools both implemented in the ns-
3 simulation environment:
λ𝑆𝑅𝑅 (𝑃𝑜𝑖𝑠𝑠𝑠𝑜𝑛) 0,0045 procs/sec A generator of LTE control procedure calls [6].
λ𝐻𝑅 (𝑃𝑜𝑖𝑠𝑠𝑠𝑜𝑛) 0,0012 procs/sec A simulator of a vMME with a three tier design
and its interactions with other network entities.
EPC Delays For all the simulations, the vMME has to
eNB ↦ vMME 4,5 ms process 4 ∙ 106 control packets as the stop
condition.
vMME ⇌ [eNB | S-GW] 9,0 ms
All the experiments were carried out on a
Service Rates machine with 4 GB RAM and Intel Xeon
𝜇𝐹𝐸 120000 pkts/sec CPU X3330 @ 2.66GHz x 4 processor.
𝜇𝐷𝐵 100000 pkts/sec
The machine ran Linux Ubuntu 12.04 LTS
OS.
𝜇𝑊 28497 pkts/sec
Table II: Parameters configuration.
University of Granada - Analytical Modeling for Virtualized Network Functions 13/19
Results
Model’s computational complexity

NU K Jackson QNA MVA Simulation

100000 3 6.667 ms 8.000 ms 9.740 s 224.280 s
500000 4 6.000 ms 9.340 ms 48.691 s 241.100 s
1000000 5 5.333 ms 10.667 ms 97.162 s 252.880 s

2000000 8 7.000 ms 21.335 ms 194.884 s 251.660 s

Table III: Execution times Vs the number of users and queues.

QNA has a time complexity O(K) (similar to Jackson’s networks analysis), where K is the
number of queues in the network.
For cellular scenarios, where there might be a large number of active sessions, MVA
might take longer than the simulation.

University of Granada - Analytical Modeling for Virtualized Network Functions 14/19

Results
Model Validation

Our model exhibits an estimation error approximately equal to 10%,

whereas for the models based on Jackson’s networks and MVA it ranges
from 60% to 90%
ห𝑇𝑠𝑖𝑚 − 𝑇𝑡ℎ𝑒𝑜 ȁ
𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑒𝑟𝑟𝑜𝑟 ≡ 𝜀 =
𝑇𝑠𝑖𝑚

University of Granada - Analytical Modeling for Virtualized Network Functions 15/19

Conclusions
In this work, we propose an analytical model based on an open queueing network to
estimate the mean response time of a VNF.
This model can be easily extended to be applied to chains of VNFs.
The model is sufficiently general to capture the complex behavior of VNFs:
Arbitrary packet flows through VNFCs, caching effects and different load balancing strategies at any VNFC.
Resources held simultaneously at multiple tiers [5].
Creation or combination of packets at any VNFC instance (e.g., to model a firewall dropping packets or a video transcoder).
Heterogeneity of the physical hardware.
Arbitrary external arrival processes and service processes.

We validate our model by simulation and compare it with other baseline models based on
queueing theory.
Our model exhibits an estimation error approximately equal to 10%, whereas for the considered baseline
models it ranges roughly from 60% to 90%.
The computational complexity of our model is similar to the method for analyzing Jackson’s networks.
However, further work is required…

University of Granada - Analytical Modeling for Virtualized Network Functions 16/19

Future Work
To conduct a field evaluation of the proposed model (stronger validation).
To extend the model:
Performance degradation due to the sharing of physical resources.
To consider the utilization of several computational resources (CPU, Memory, I/O disk) at a VNFC [5].
VNFs implemented using microservices architectures.
VNF characterization.

University of Granada - Analytical Modeling for Virtualized Network Functions 17/19

References
[5] B. Urgaonkar, G. Pacifici, P. Shenoy, M. Spreitzer, and A. Tantawi, “An
Analytical Model for Multi-tier Internet Services and Its Applications,” ACM
SIGMETRICS Performance Evaluation Review, vol. 33, no. 1, pp. 291–302,
Jun. 2005
[6] J. Prados-Garzon, J. Ramos-Munoz, P. Ameigeiras, P. Andres-
Maldonado, and J. Lopez-Soler, “Modeling and Dimensioning of a
Virtualized MME for 5G Mobile Networks,” IEEE Trans. Veh. Technol., vol.
PP, no. 99, pp. 1–1, 2016
[7] W. Whitt, “The queueing network analyzer,” Bell System Tech. J.,
vol. 62, no. 9, pp. 2779–2815, Nov. 1983.
[14] J. Bi, Z. Zhu, R. Tian, and Q. Wang, “Dynamic provisioning modeling
for virtualized multi-tier applications in cloud data center,” in IEEE 3rd
Int. Conf. on Cloud Computing (CloudCom), 2010, pp. 370–377.
University of Granada - Analytical Modeling for Virtualized Network Functions 18/19
University of Granada - Analytical Modeling for Virtualized
Network Functions