Impact of Network Function Virtualization: A Study

based on Real-Life Mobile Network Data

Ahmad Bilal∗ , Andras Vajda∗ and Taleb Tarik†
∗
Ericsson, Finland
†
Aalto University, Finland
bilal.x.ahmad@ericsson.com, andras.vajda@ericsson.com , talebtarik@ieee.org

Abstract—Mobile Operators are looking for new ways to cope same type of load is executed in a virtualized environment. We
with ever-increasing data traffic while improving the operational also analyzed different scenarios with various virtualization
and capital efficiency of their networks. Cloud computing and overheads for control and data-plane processing. To the best
network function virtualization (NFV) have emerged as key
enablers to optimize resource utilization and at the same time knowledge of the authors, this is the first measurement study
reduce network operational expenditure (OPEX). In virtualized of live network resources utilization which will give future
networks, network functions are delivered as software running directions to implementation of NFV in mobile networks.
on generic hardware allowing service providers to dynamically The remainder of this paper is organized as follows. Section
allocate resources based on traffic and service demands. In this II presents background literature related to NFV technologies.
paper, we analyze resource utilization using real-life data of two
different mobile networks and evaluate the impact virtualization Section III presents the networks and key nodes under study.
would have on these networks. Some conclusions are drawn based Section IV reports our analysis based on the measured data.
on the analysis. The simulation results with and without virtualization over-
Index Terms—NFV, Cloud Computing, Real-Life Mobile Net- heads for both networks are also reported therein. Finally, the
work Data, Resource utilization. paper concludes in Section V.
I. I NTRODUCTION II. NFV TECHNOLOGIES: RELATED WORK
Traditionally, network nodes are delivered pre-configured NFV allows a service provider to deliver network functions
in a highly optimized manner with specialized hardware as pure software running in a virtualized environment with
specific to node functionality. Deployment of new network reduced cost and high deployment efficiency [6][7]. This shift
services typically requires separate hardware with significant of hardware to software running in standard virtual machines
cost and lead-time of integration and operation. Network or containers (e.g., Docker or Googles Kubernetes) is expected
Function Virtualization (NFV) [1] promises to address these to reduce CAPEX and OPEX. Reduced cost, increased service
challenges through the decoupling of software from hardware deployment velocity, services introduced based on geogra-
with the introduction of a virtualization layer. NFV can be phy and costumers’ needs, ability to efficiently cope with
applied to most functions in the network whether these are emerging resource-intensive applications [8], reduced energy
control-plane or data-plane functions. The goal of NFV is consumption and several other benefits are expected to be
to run network functions as software in e.g. virtual machines achieved through NFV [1]. In short, NFV, along with Software
(VMs) on top of virtualization platforms deployed on generic Defined Networking, will enable the launch of anything as a
hardware. It is understood that running network functions on service in a more cost efficiency way while ensuring short time
general purpose hardware, rather than on dedicated hardware, of service to market [9]. However, this cannot be achieved
can impact performance. The virtualization layer introduces without addressing challenges spanning from system design
latency and an extra overhead that consumes extra capacity. to ensuring service resiliency [10].
Significant overhead may be required to implement software- NFV is also foreseen as an important technology to enable
based switches (commonly called virtual switches or vSwitch) the on-demand creation of cloud-based virtual mobile net-
that route packets to and from appropriate VMs. This CPU works. Here, an important challenge pertains to the placement
overhead can reduce maximum throughput and increase la- of Virtualized Network Functions (VNFs), within the same
tency on an I/O device. or across distributed datacenters, considering the performance
In this paper, real data from two different packet core constraints and functional relationship among VNFs that form
networks with dedicated network nodes were analyzed. The a single virtual network infrastructure. To this problem, differ-
two selected packet core networks represent mobile network ent solutions have been devised not only ensuring communi-
deployments in developed and developing countries, respec- cations efficiency for mobile users placing VNFs at strategic
tively. We created a load profile for each node that allowed positions but also ensuring cost efficiency for the operators;
us to evaluate how the overall load evolves over a period minimizing the cost associated with the instantiated VMs.
of time. Based on the load profile, the amount of general The work in [11] analyzes the impact of deployment strategy
purpose hardware needed was calculated, assuming that the on the overall performance of virtual network infrastructures

978-1-5090-0304-4/16/$31.00 ©2016 IEEE 541

deployed on the cloud. In [12], a fine granular resource-aware
VNF management is proposed for the initial deployment and
runtime management of virtual network infrastructures.
Cloud computing and industry-standard high-volume
servers are key enablers to achieve the goals of NFV. How-
ever, general purpose processors are not designed to process
modern, high speed protocols [2]. Historically, control plane
applications have been implemented on general purpose hard-
ware and the performance degradation is small enough to
be neglected. On the other hand, data plane applications are
mostly executed on specialized hardware to perform specific (a) West European country network (NW 1)
functions and to meet specific requirements in terms of Quality
of Experience (QoE) or Quality of Service (QoS) . Therefore,
realizing data plane application on general purpose hardware
to meet desired performance and cost requirements is an
industry challenge that needs to be tackled. In NFV, perfor-
mance degradation is due to the introduction of a virtualization
layer between hardware and Operating System (OS). This
abstraction layer, i.e., hypervisor, increases the overhead for
accessing resources, most importantly I/O resources, which are
the most difficult to partition and manage [3], due to mismatch
between CPU and I/O speeds; random nature of packet arrival;
higher number of virtual machines that need to be served; and
overhead due to switching done purely in software (i.e., the
(b) African country network (NW 2)
vSwitch component of hypervisor).
There are software-based, hardware-assisted and full Fig. 1: The two studied mobile networks
hardware-based solutions for sharing I/O devices. With the
software-based solutions, the hypervisor translates the I/O
requests it receives and allocates them serially to physical 1) SGSN-MME (Serving GPRS Support Node - Mobility
resources and packets are routed through the virtualization Management Entity): The SGSN-MME provides SGSN
layer, which adds performance overhead. Virtual machine and MME functionality. It is possible to use SGSN-
performance is greatly degraded if the processor is interrupted MME with SGSN functionality, MME functionality or
frequently for I/O [4]. both.
Some network devices use hardware-assisted approach to 2) EPG (Evolved Packet Gateway): The EPG provides
accelerate software-based I/O sharing but this solution still GGSN (Gateway GPRS Support Node), S-GW (Serving
involves a hypervisor in between, resulting in performance Gateway), and P-GW (Packet Date Network Gateway)
degradation. In order to overcome this performance overhead functionality. It is possible to use EPG with GGSN
the solution needs to avoid the involvement of hypervisors functionality, S-GW functionality, P-GW functionality,
in I/O operations. The PCI-SIG SR-IOV standard [5] offers S-GW and P-GW functionality, or all simultaneously.
an approach for I/O sharing that overcomes the performance
limitations as discussed above. SR-IOV enables I/O devices A. West European Country Network (NW 1)
to support multiple virtual functions, lightweight mechanisms
for transferring data to and from network adopter, which can There are three packet core sites in three different locations
be assigned directly and can communicate to VM bypassing consisting of three SGSN-MMEs and two EPGs as shown in
hypervisor [4]. Although this solution is highly efficient in Fig. 1(a). The three SGSN-MMEs are integrated in a triple ac-
terms of performance, it has important drawbacks [3], in terms cess pool network that provides redundancy to packet-switched
of scalability, outbound traffic and live migration limitations. network. 2G and 3G traffic from BSCs (Base Station Con-
Besides all these solutions, there is no study on real-life troller) and RNCs (Radio Network Controller) are distributed
networks that can quantify the impact of virtualization on in different ratios among three geographical areas. 4G traffic
overall system capacity and performance if virtualization and is disturbed equally among the SGSN-MMEs. The two EPGs
data plane overhead are taken into account. are also configured for triple access (2G/3G/LTE) at different
locations. Each EPG is divided into two logical nodes: S-GW
III. ANALYZED NETWORKS is only used for 4G traffic, P-GW used for 2G/3G and 4G
This section provides an overview of the studied mobile connections. For NW 1, we used high granularity data over
networks. The main nodes within the packet core network four months for SGSN-MME resources and three months for
under study are: EPG nodes.

542
B. African Country Network (NW 2)
There are two packet core sites in two different locations
consisting of two SGSN-MMEs and two EPGs as shown in
Fig. 1(b). This network also provides packet services to 2G/3G
and 4G users. However, resources utilization is small com-
pared to NW1 due to low number of PS users. Furthermore,
there is difference in resources dedicated to each network; for
instance the processor resources in NW 2 are one third of those
in NW 1. For this network, SGSN-MMEs data is analyzed for
seven months while EPG data is analyzed for five months with
one month excluded in between. The second EPG is not in use.
For the sake of resiliency, it is used as back up if running EPG (a) SGSN-MME CP & DP maximum CPU usage and daily variance
goes down. (NW 1).
The motivation behind studying these two networks is to
compare cost-effectiveness between virtualizing a network
with higher density of packet services and a network with
low density of packet services. For both networks, the SGSN-
MMEs statistics are captured with a five minutes granularity
while EPG statistics are captured with a fifteen minutes
granularity.
IV. ANALYSIS
This section is divided into two parts. The first part intro-
duces the actual measured data while the second part shows
simulation results when the actual measured data was to be
executed in a virtual environment.
(b) SGSN-MME Global maximum Total CPU usage and Daily variance
A. Measured Data (NW 1).

Load profiles were created for each node and on a global

level for both networks to evaluate how load evolves over
a period of time. Control Plane (CP) and Data Plane (DP)
resources utilization were analyzed separately from these
nodes. From Fig 2, it can be seen that the maximum CPU
utilization on a daily scale, for both node level and global
level, does not exceed 25 % of the total capacity. The main
reason behind this is that nodes are optimized in such a way
that in case of a failure, the load is taken by the other nodes
in the same pool. It was also noticed that the CPU usage
variance is almost constant with a slight decrease pattern on
weekends for the whole measurement period. These maximum
(c) EPG CP & DP maximum CPU usage and daily variance (NW 1).
CPU usages along with daily CPU variance are good indicators
to dimension nodes efficiently for failure cases. It was also
observed that the control plane usage was higher for NW 1
while the data plane usage dominates for SGSN-MME in NW
2 and that is due to limited 4G services in NW 2 as it can be
observed in Fig. 3. It was also observed that resource usage
follow a repetitive pattern on daily scale. After analyzing both
networks, we can draw the conclusion that days of month
are pretty much the same on the basis of resource utilization,
variation occurring primarily as a function of the time of the
day. However, resource usage patterns differ from a network
to another but repeat themselves on a daily basis within the
same network as shown in Fig. 3, where each line indicates
one day of a month. (d) EPG Global maximum Total CPU usage and daily variance (NW
1).
Fig. 4 plots the total CPU usage of SGSN-MMEs at global
level for both studied networks. From the figure, it can be Fig. 2: Maximum and total CPU usages and daily variances
for NW 1 Nodes.

543
 (a) SGSN-MME DP CPU Usage (NW 1) (b) SGSN-MME CP CPU Usage (NW 1)

(c) SGSN-MME DP CPU Usage (NW 2) (d) SGSN-MME CP CPU Usage (NW 2)
Fig. 3: DP and CP CPU usage of SGSN-MME for both studied networks.

(a) SGSN-MME Global CPU usage (NW 1) (b) SGSN-MME Global CPU usage (NW 2)
Fig. 4: SGSN-MME Global total CPU usage for both studied networks.

(a) EPG CP CPU usage on node level (NW 1) (b) EPG Global CP CPU usage (NW 1)
Fig. 5: EPG CP CPU usage on node and global level for NW 1.

544
noticed that in NW 2 the utilization is lower compared to NW predict how load will evolve will hence be an essential
1. It is because that NW 2 is upgraded lately for 4G services ingredient of any commercial NFV deployment [12][13]. In
and most of the control plane processors are not used at all. this study, we assumed that horizontal scaling or scale out (i.e.,
In NW 2, only one EPG is utilized and its utilization level adding or removing VMs based on load) is used. However,
is quite low, making it impossible to draw any observations for NW 2 vertical scaling (i.e., resizing VMs based on load)
from the respective graphs and they are thus not included in would be more interesting and cost effective [13]. In general,
this article. However for NW 1, EPG CP CPU usage on node horizontal scaling is more effective when amount of dedicated
and global level can be observed from Fig. 5. resources are high.
From the studied networks, it is worth underlining the
repetitive nature of the load as well as the fact that the best V. CONCLUSIONS
predictor of how load is likely to evolve is the time of day
Based on our analysis of data from real-life mobile net-
and the day of the week (e.g., weekend days) giving a further
works, the overall load is a fraction of installed capacity
refinement opportunity. These observations are important for
which applies both to global (all geo- redundant) and node
designing suitable scaling algorithms for networks deployed
level. Furthermore, resource utilization is different from a
in virtual environments [6][13].
network to another but is highly correlated with time of
B. Simulation Results day and follows the same pattern within the same network.
Based on real load data, we evaluated how much actual It is concluded that virtualization with dynamic scaling of
capacity (i.e., number of blades) would be required, if the same node size based on load is more cost efficient even if large
load would have been supported in a virtualized environment. virtualization overhead for data-plane is taken into account.
Three assumptions were made: This defines an interesting research area where future research
1) We can scale the allocated capacity according to actual shall focus. Additionally, for networks with small numbers of
load. In reality, one should assume a maximum 80 % dedicated resources, vertical scaling (resizing the VMs) will
load level (an industry-wide best practice) and take into be more cost-effective than horizontal scaling (adding more
account the granularity of allocated VMs. VMs) [13]. Future research work includes identifying any
2) There is a constant overhead due to virtualization both other possible correlation related to load to develop dynamic
for control and data plane traffic, modeled as the factor scaling algorithm for resources based on utilization and verify
α, with a value of 1.1 based on authors’ prior empirical it on real network conditions.
experience.
R EFERENCES
3) There is an additional overhead for data plane process-
ing, which we modeled as β, to quantify the impact on [1] ETSI, ”Network Functions Virtualisation: An Introduction, Benefits, En-
throughput and latency. ablers, Challenges & Call for Action,” NFV - Introductory white paper,
Oct. 2012
Same kinds of calculation were carried out both for node [2] K. Tan et al., ”Sora: High performance software radio using general
level and global level. It can be seen that even with high purpose multi-core processors,” in Proc. USENIX Int. Symp. NSDI, 2009,
β values the amount of generic hardware needed is still pp. 7590.
[3] R. Shah, ” Network I/O Virtualization : Challenges and Solution Ap-
significantly below its native counterpart as shown in Fig. 6. proaches, ” Seminar Report Indian Institute of Technology, Apr. 2014.
NW 2 plots are not included in this article as the number of [4] B. Johnson and R. Deshmukh, ”Boosting I/O performance for virtualized
dedicated resources are low and horizontal scaling, the focus servers,” Dell Power Solutions, No. 3, 2012.
[5] Intel, ”PCI-SIG SR-IOV Primer An Introduction to SR-IOV Technology,
of this study, does not yield interesting results. ” Revision 2.5, Jan. 2011.
A number of interesting conclusions can be drawn from [6] T. Taleb, M. Corici, C. Parada, A. Jamakovic, S. Ruffino, G. Karagiannis,
theses simulations. First, the overhead, introduced by vir- and T. Magedanz, ”EASE: EPC as a Service to Ease Mobile Core
Network,” in IEEE Network Magazine, Vol. 29, No. 2, Mar. 2015. pp.78
tualization, plays a smaller role than expected; even with 88.
a supposedly low performance implementation, NFV will [7] T. Taleb, ”Towards Carrier Cloud: Potential, Challenges, & Solutions,”
outperform native installations. This is due to a number of in IEEE Wireless Communications Magazine, Vol. 21, No. 3, Jun. 2014.
pp. 80-91.
factors: virtualization reduces the need for dedicated resources [8] T. Taleb, A. Ksentini, M. Chen, and R. Jantti ”NFV-based Dynamic
to achieve geographical redundancy (i.e., nodes can be deliv- Service Chaining for Emerging Mobile Social Media Applications”,to
ered at just two different locations datacenters to achieve appear in IEEE Trans. on Wireless Communications.
[9] T. Taleb, A. Ksentini, and R. Jantti, ” Anything as a Service for 5G
geographical redundancy instead of three or more as today) Mobile Systems” to appear in IEEE Network Magazine.
and dynamically scale them according to load conditions; at [10] T. Taleb, A. Ksentini, and B. Sericola, ”On Service Resilience in Cloud-
the same time resource utilization can be kept at the optimal Native 5G Mobile Systems” , to appear in IEEE J. Select. Areas in
80 % (which is a typical industry best practice level) most Communications.
[11] F.Z. Yousaf, P. Loreiro, F. Zdarsky, T. Taleb, and M. Leibsch, ”Cost
of the time, hence approximating optimal resource utilization Analysis of initial deployment strategies of a Virtual Network Infrastruc-
levels. ture in a Datacenter,” in IEEE Communications Magazine, Vol. 53, No.
Second, the most important technology to unlock the benefit 12, Dec. 2015, pp. 60 - 66.
[12] F.Z. Yousaf and T. Taleb, ”Fine Granular Resource-Aware Virtual
of virtualization is optimized scaling, enabling allocation of Network Function Management for 5G Carrier Cloud,” to appear in IEEE
resources on a need basis. An algorithm that can accurately Network Magazine.

545
 (a) SGSN-MME (Node Level) (b) SGSN-MME (Global Level)

(c) EPG ( Node Level) (d) EPG (Global Level)

Fig. 6: Simulation Results

[13] S. Dutta, T. Taleb, and A. Ksentini, ” QoE-aware Elasticity Support in

Cloud-Native 5G Systems,” in IEEE ICC ’16, Kuala Lumpur, Malaysia,
May 2016.

546