On The Reliability of Voice Over IP (VoIP) Telephony
Sumalya Pal, Raviteja Gadde and Haniph A. Latchman
University Of Florida
Department of Electrical and Computer engineering
Abstract
Voice over Internet Protocol (VoIP) for telephony is
becoming important and widespread with the use of the
Internet for multimedia traffic and particularly voice
communication. The Public Switched Telephone Network
(PSTN) system is widely recognized as having set a very
high bar in expected telephony performance - an overall
availability of 0.99999 (“five nines”). For VoIP telephony
to replace traditional systems they should be able to
provide a comparable level of availability to the users as
on a PSTN system - a formidable challenge for VoIP
systems. This paper reviews the “five nines” availability
for PSTN systems and explores how emerging VoIP
systems may achieve this level of performance. We
propose a Kamailio and Freeswich combination (an
open-source SIP Router and VoIP application server,
respectively), operating in a Linux environment in
conjunction with a load balancer and ENUM to provide
high availability VoIP service.
Key Words: Alpine Linux, Availability, reliability,
redundancy, Five-nines, Kamailio, FreeSwitch, LVS,
Ultra Monkey, Virtual Servers. VoIP.
1. INTRODUCTION
IN THE world of traditional telephony, the providers of
the particular telephone services often quote a service
level of “five nines” or an availability of 0.99999 as a
standard signifying quality and assurance. This is
sometimes used as a bragging point about their individual
products and service offerings. It also means that any new
competitors in the same field have to achieve this goal in
order to compete effectively. Hence emerging Voice over
Internet Protocol (VoIP) telephony – both for long
distance or toll calls, as well as for local inter- or intra-
office calls (an alternative for traditional telephone
switches or Private Branch exchanges (PBX‟s)) – faces
the “five nines” hurdle.
Reliability, a concept that generally deals with the
continuous operation of a service, depends on the
hardware and software elements of the system, while
availability, is a measure of fraction of time that the
service is usable. Reliability can be defined as the
calculated value, which represents how often the system
fails as compared to the percentage of time the system, is
available. Availability on the other hand depends largely
on the probability of the failure of a hardware component.
It‟s calculated by counting the number of components in a
system and then calculating the overall mean time between
failures (MTBF). It is represented by this formula [1]:
Availability = MTBF / (MTBF+MTTR)
Where MTTR = mean time to repair. Typically MTBF
increases as the number of components in the system
increases. The most important factor is the MTTR [1],
which literally decides the availability of a certain IP
telephony connection. For example let‟s take MTBF =
400,000 Hours and MTTR = 1 Hour. Then the
Availability will be approximately 99.9997%, which is
excellent from any industrial standard. But let‟s take
MTTR to be 96 hours then the Availability will be 99.7%
approximately. .
Traditional telephony achieved the impressive “five-
nines” level of availability mainly from hardware
redundancy and reliability of components. Several
organizations have been working towards making the
availability of IP telephony approach that of the PSTN.
Some of the most prominent IP telephony companies
working towards this goal are Cisco, ShoreTel, Nokia and
AT&T.
In Section 2 we explore in greater detail the
implications of the nines‟ criteria, while in Section 3 we
will discuss some reasons why VoIP still lags behind the
PSTN when it comes to availability of the communication
system. Section 4 presents several recent positive
developments in IP telephony systems from some of the
companies working in this field and in Section 5 we
examine the VoIP frameworks that are being created for
providing high availability. Section 6 identifies some of
deficiencies in present approaches to VoIP systems and in
Section 7 we propose a framework for highly availably
VoIP systems based on open-source and an Internet
inspired approach. Section 9 presents our conclusion and
anticipates further work towards high availability open-
source based VoIP systems.
2. THE NINES’ CRITERIA
There are different figures of „nines‟ that are used in
demarcating the availability of a device. The one nine
represents a 9.0 % availability, which translates on an
annual basis to 332 days of downtime in a year, which
means that the system will function properly on average
for only one month in a year. The two-nine reliability or
99%, availability implies 3 days and 15 hours seconds of
downtime in a year. The three-nines or 99.9% availability,
signifies downtime of eight hours and forty-six minutes in
a year. The four-nines criteria or 99.99% availability
means that there will be a downtime of fifty two minutes
and thirty six seconds in a year. Now in today‟s age of
telecommunications the five-nines criteria has become the
holy grail of telecommunications industries.
Five nine criteria is generally used to represent a target
availability figure. Some people even go to the extent of
calling it a catchall since with high availability we get
high reliability and vice versa. Five-nine availability (or
99.999%) means a downtime of five minutes and fifteen
seconds or less per year. The PSTN service is now
shooting for six nines availability, which means an
availability of 99.9999%. This means a downtime of
thirty-two seconds or even less, per year.
3. REASONS WHY VOIP STILL LAGS
BEHIND PSTN
Consumers expect same level of reliability as they expect
from traditional PSTN service in VoIP. PSTN services
have matured into a highly reliable system providing
99.999% of availability. It has been noted that the PSTN
system can be characterized as an upper bound of
dependability that a distributed computing system can
achieve.
The major factors that affect the VoIP system are network
outages and SIP server outages. The service availability of
the system drops to 98% when network outages interrupt
calls [12]. In the following table we can see the service
availability of VoIP systems in various networks.
Network/path type
Call success probability
All
99.53%
Internet2
99.52%
Internet2+
99.56%
Commercial
99.51%
Domestic (US)
99.45%
International
99.58%
Domestic Commercial
99.39%
International Commercial
99.59%
Table 1 CALL SUCCESS PROBABILITY ON FIRST CALL
ATTEMPT WITH RESPECT TO NETWORK/PATH TYPE [12]
Also Packet Loss causes considerable quality
degradations for the users, which can be equated to a
dropped call. Even when the system has packet loss below
0%, it was observed that service availability that can be
achieved is a maximum of 97.7%. [12]
Network Outages are not a fleeting occurrence but a
common place in present Internet. The advantage the
PSTN is that once a call is established it is ensured quality
but on other hand Internet provides a best effort service.
There is also a overall call abortion probability of 1.5%
giving 98% service availability which is falls far short of
achieving 99.999% availability [12]. Given all the above
reasons we can see that achieving five-nines in a normal
Internet conditions is difficult.
Many VoIP systems often feature a central Session
Initiation Server (SIP) or other type of VoIP Server.
Server outages may occur because a software problem or
a failure in the server hardware. To achieve Five Nines
availability the system needs to have MTBF of 2,400,000
hours and MTTR of 24 hours. Even if MTTR is reduced
to 4 hours, we need to have a MTBF of 400,000 hours.
The state of the art system provides a MTBF of 100,000
hours and MTTR of 24 hours [1].
This means we cannot design a high availability system
with a single component. Highly available systems are
designed by adding a redundant similar component as a
stand-by. When the primary fails, the system falls back to
secondary. A similar approach is needed for high
availability VoIP.
4. IP TELEPHONY RELIABILITY
FRAMEWORKS
Accessibility, continuity and fulfillment are the main
factors in IP telephony. Accessibility is the ability to
initiate a voice call when desired; continuity is the ability
to finish the call successfully without jitter upon
successful access; and fulfillment is the desired call
quality by the customer upon successfully establishing the
call [3].
The basic function of IP network is to carry data and
traffic without experiencing disruptions due to increased
delay and packet loss. IP backbone routers inherently lack
carrier grade reliability [3]. In order to overcome this
weakness in IP telephony networks, a complete redundant
connection between backbone routers, which are grouped
in pairs, needs to be established.
A network architecture having a fully redundant
backbone guarantees network stability and minimum
delay for the rerouted traffic. This is because in the event
when one backbone router fails, the other backbone
routers, which are grouped together, will be able to
reroute any traffic. However such network architecture
involves high expenditure, which is a tradeoff that the
architecture imposes for high network availability. In
essence the high availability is to a large extent dependent
on the design of the network infrastructure.
The architectures that we use for network applications
follows IP distribution routing protocols like that of OSPF
(open shortest path first). Under such conditions the
rerouting delay is pretty high, and so these are not ideal
for robust Internet telephony applications. Cognitive
networking is becoming a necessity and newer ways to
reroute traffic are already under research and
development. One such protocol is MPLS (multiprotocol
label switching). In this protocol architecture, rerouting
paths are assessed in advance and whenever there is a
failure in network, immediately this alternate path is
invoked [6].
However further investigation is needed to verify that
such protocols can be used in a more cost effective way.
This is because it might not be possible to implement such
fast routing protocols in every class of networks due to
cost or may be due to security considerations.
5. AN EVALUATION OF THE
APPROACHES TOWARDS “FIVE-NINE”
PERFORMANCE FOR VOIP
There have been several attempts by industries and
academic institutions to find a robust and cost effective
mechanism for making VoIP five-nines available. Some of
the approaches include server redundancy (1+1)
mechanism by industries, software based approaches
using SIP protocol and different software tools like Ultra
Monkey [25], Piranha and other similar technologies
based on Unix based systems or even using peer-to-peer
(P2P) as a mechanism for effective VoIP usage.
In the following discussion we will evaluate these
approaches as to how effective they are in the search for a
five-nine-availability VoIP solution.
Session Initiation Protocol (SIP) based approach
a. The SIP Architecture
The Session Initiation Protocol (SIP) is a more
flexible and simpler solution compared to H.323
architecture [23], for handling multimedia sessions of
VoIP. SIP is defined by IETF signaling protocol as a
text based protocol (that uses UTF-8 encoding) which
incorporates several elements of the Hyper Text
Transfer Protocol (HTTP) and Simple Mail Transfer
Protocol (SMTP) and is widely used for controlling
multimedia communications such as voice and video
calls over Internet Protocol (IP). SIP is an application
layer protocol and is independent of the underlying
transport layer. SIP can run efficiently on
Transmission Layer Protocol (TCP), User Datagram
Protocol (UDP) and also on Stream Control
Transmission Protocol (SCTP). It uses port 5060 for
UDP as well as for TCP [24]. SIP acts as an enabling
protocol for VoIP telephony services.
Fig -2 Basic SIP architecture for VoIP
As an alternative to classical PSTN, SIP based
telephony services are being considered since it offers
a number of advantages over the former.
There are many impressive features of SIP that made
it a possible choice for VoIP. Some of these features
are:
1. SIP makes sure that the call made by the user
reaches its destination, without the restriction
of the respective party‟s location.
2. This takes into account that not all party (in a
conference call) can support all the features,
for example video. Hence negotiating feature
is highly regarded in SIP.
3. Call cancellation and call hold is completely
supported by SIP. A user can add more users
to the call or terminate any user. Call
transferring is also supported in SIP.
4. Flexibility in call features is one of the major
advantages of SIP. The user can start using
another program while maintaining the call.
For example a user can start using video
while on phone.
5. Less advanced devices can be incorporated to
make a call since SIP allows usage of several
different codecs, which enable negotiation of
the media in a call.
b. SIP availability
Wenyu Jiang and Henry Schulzrinne of Columbia
University [12] conducted an experiment in which
they set up a group of test clients in different
networks and simulated automatically random calls
between these clients. The overall service availability
as calculated by them was 0.98.
They stated that in order to calculate the correct
measure of availability in VoIP system, we must
consider the call abortion probability of the users.
They concluded that although SIP does not really
provide five-nines availability like the PSTN, it does
come very close to mobile networks‟ availability,
which is 0.97 to 0.99.
N+1 Redundancy approach
This method is followed by several
networking/hardware industries in order to make their
products five-nines available. Ciscos, Nokia,
ShoreTel are some of the companies who have
invested considerable amounts in this technique.
Cisco has made considerable advancement in
achieving five-nines availability in their IP telephony
systems. Taking a leaf from the definition of legacy
PBX connections, which over all neglects the
influence of non-redundant components of an IP
telephony system, Cisco claims to have theoretically
proven that their IP telephony solution meets the five-
nines standard. They tested their IP telephony
solution for hardware reliability, software reliability,
link/carrier reliability, power/environment reliability
and network design reliability. For this they largely
included the parts count method by Telcordia, which
is used, for MTBF calculations [2]. Hardware
reliability was assessed using the CI infrastructure
model with redundant Catalyst 6509 chassis, power
supplies and supervisor modules for Cisco Call
Manager access switches, which were all redundant in
nature [2]. Based on the estimated software forced
reloads the theoretical software reliability for the IP
telephony product was measured [2].
Cisco typically enforced N+1 redundancy technique
to their IP telephony environment.
The overall availability has been calculated using the
equation:
System Availability = π
n
i=1
= availability (i)
Where i = number of components from 1 to n. This
number is then multiplied by the availability of each
component to give the overall system availability
[2][4].
Now in case there is another failure while the
repair for the first failure is still under way then
calculation of availability of a redundant parallel
system has to be conducted. Cisco performed this by
using this equation:
Parallel availability = 1 – [ π
n
i=1
(1 – component
availability(i))] [2]
In essence Cisco, Nokia and Shore-Tell calculated
the parallel availability/reliability of each component
in a redundant system thus ignoring the non-
redundant parts of the system and then calculated the
overall system reliability by adding up the parallel
reliability of the redundant system.
This process of redundancy is very effective in five-
nines availability assessment of any system, but on
the flip side, it increases the cost of manufacturing the
system.
6. PROBLEMS WITH CURRENTLY
AVAILABLE APPROACHES IN
ACHIEVING FIVE-NINES
AVAILABILITY
The redundant mechanism followed by most industries in
order to raise the service availability of traditional VoIP
by making use of some redundant service hosts
coordinated by central load balancer, in which each
component is coupled with a stand by partner is not a very
cost effective mechanism.
Another approach is to make use of a layer-4
switch in addition to the available switching device
between SIP clients and a set of SIP servers to dispatch
service load without any modifications to the SIP service
components. An example of such an approach is the
development of a hardware-based mechanism like BIG-IP
[18]. Such an approach might be able to improve the
availability of a system but will cost a lot in
implementation.
Software based approach developed on a Linux
Virtual Server (LVS) [19], can achieve the required
availability, but may cause delay as the SIP messages
traverse through intermediate nodes.
There exists a Peer-to-Peer (P2P) approach [20]
[21] towards making VoIP more accessible and available.
In such a mechanism, there exists no centralized manager
and the peers themselves act as servers as well as clients.
However there are numerous security concerns with such
an approach. Any peer with malicious intent can alter the
SIP messages by other peers. Then again there is a
problem of anonymity, as the peers might not want to
share SIP messages, which can be read by other peers.
Such a mechanism also suffers from SIP message
transversal delays, which might cause additional packet
loss and poor connectivity.
7. KAMAILIO – FREESWITCH BASED
VOIP FRAMEWORK
Here we propose a mechanism for attaining five-nines
availability of VoIP telephony using RTP and SIP. The
basic setup consists of Kamailio or SIP Router [26],
which is an open source SIP server, and FreeSwitch [27],
which is an open-source telephony platform, designed to
facilitate the creation of multimedia-messaging (voice
video and text) driven products.
The basic setup consists of two or more Kamailio virtual
servers and one or more FreeSwitch virtual server running
on Linux platform. Each of the Kamailio virtual servers
has a pre-determined number of supportable clients based
on the CPU and memory limits. The X-Lite [28]
softphone or any number of SIP Compliant IP telephones
are used as the primary IP Telephony devices. Call are
routed from the clients to the Kamailio servers using a
distributed DNS-based ENUM (E.164 Number to URI
Mapping) system with priority settings. If at any moment a
particular Kamailio virtual server is down or running at its
full capacity then it will reroute the calls to other Kamailio
virtual server 2 or to the FreeSwitch virtual server(s)
which will primarily serve the purpose of voice mail and
music on hold, an automated attendant.
Fig 4: Flowchart of our proposed setup.
To give hardware redundancy, it is proposed to use Ultra
Monkey as a load balancer on the Linux Virtual Server
(LVS), which is essential for creating highly available
network services.
Fig 5 – Ultra Monkey setup.
Ultra Monkey is an open source project to provide
flexible high availability frameworks. This can be used for
developing both single server based high availability
systems as well as multiple servers based high availability
systems. It uses the Heartbeat protocol [29] to monitor
the servers.
Heartbeat is a protocol that monitors messages sent at
regular interval between two servers and at any point of
time if messages are not received from one server, it is
assumed that the server in question has failed and some
form of evasive mechanism is followed in order to rectify
this. Heartbeat protocol can send heartbeat messages over
both serial links and Ethernet interfaces. It uses an IPfail
plugin [30] that helps in determining which nodes should
be active.
Fig 6 – Ultramonkey with ipfail plugin [30]
After configuring Heartbeat, we designate a master node
which when Heartbeat starts up, designates an interface
for a virtual IP address which may be accessed by external
end users. Under any situation if this node fails then in
order to ensure that this machine receives all traffic bound
for this address, another node in the heartbeat cluster will
start up an interface for this IP address. This process is
carried out using Gratuitous ARP [31]. This process is
called IP failover takeover.
Ultra monkey makes use of Heartbeat to manage IP
addresses on the host on which Linux Virtual Server runs.
This also monitors the ultimate destination of a connection
made to a virtual service using IPaddr2 resource.
FreeSwitch is used as a voicemail, music-on-hold,
conferencing and automated attendant server. From Figure
6 we can see how the whole setup works. The Kamailio
servers if unable to handle the calls at any time will
redirect the calls to the FreeSwitch server, which will
either offer an appropriate service.
8. CONCLUSION AND IMPLICATIONS
The main hurdle in providing five-nines-availability over
VoIP networks is that a server failure at any one point,
shuts down the system for a certain amount of time in
order to recover from the failure. This typically decreases
the MTBF. If we can somehow can take the load off the
failed server and divide it to several working servers, the
MTBF will certainly increase; thus increasing the
availability of the VoIP network.
Ultramonkey and ENUM provides load balancing to the
servers using heartbeat and priority routing. Hence at any
time when there is a load on one server (in our case
kamailio servers), it will aptly reallocate the load to the
server, which is comparatively free. Heartbeat
continuously monitors the messages sent between the two
servers, and at any point when it finds that there has been
a communication break down, it takes evasive action to
rectify the failure.
A working model of the Kamailio-FreeSwich Utltra
Monkey and ENUM-based VoIP system has been built
and work is continuing on formal theoretical study of the
overall availability as well as on generating empirical
results as the system is scaled in terms of the number of
supportable clients as well as in geographical and server
count scope.
9. Reference
[1] Ensuring Reliability in IP Telephony – Shore Tel-
IP Telephony from A-Z e-book
[2] IP Telephony: The Five Nine Story – Cisco systems
white paper
[3] VOIP Reliability: A Service provider’s perspective
- Carolyn R. Johnson, Yakov Kogan. Yonatan Levy,
Farhad Saheban and Percy Tarapore, AT&T Labs
[4] High Availability Solutions For SIP Enabled
Voice-Over-IP Networks – Cisco Systems whitepaper.
[5] Exploring the challenges to powering the future as
telecommunications transitions to IP based networks –
Nicholas Osifchin International power strategies, USA
[6] IP Telephony: Reliability You can count on – Shore
Tel white paper
[7] Power and Cooling for VOIP and IP telephony
Applications – Viswas Purani
[8] Convergence: the business case for IP Telephony –
Bob Emmerson.
[9] VOIP and IP Telephony: Planning for convergence
in state government – Nascio Whitepaper
[10] Self-Admission Control for IP Telephony using
Early Quality Estimation - Olof Hagsand1, Ignacio
M´as1, Ian Marsh2, and Gunnar Karlsson1-1 Department
of Microelectronics and Information Technology Royal
Institute of Technology (KTH) S-16440 Kista, Sweden 2
Swedish Institute of Computer Science Box 1263 SE-164
29 Kista, Sweden
[11] IP Telephony Security: an overview -Cisco
Systems.
[12] Assessment of VoIP Service Availability in the
Current Internet –Wenyu Jiang, Henning Schulzrinne
[13] Design and Implementation of a Low Cost DNS-
based Load Balancing Solution for the SIP-based VoIP
Service - Jenq-Shiou Leu, Hui-Ching Hsieh, Yen-Chiu
Chen, and Yuan-Po Chi
[14] High-Availability Solutions for SIP Enabled Voice-
over-IP Networks –CISCO
[15] Design and Implementation of a High Availability
SIP Server Architecture- Diplomarbeit , Nils Ohlmeier
[16] Lessons from the PSTN for Dependable Computing
-Patricia Enriquez, Aaron Brown, David Patterson
[17] Network Performance Analysis of Internet
Telephony on SIP in ENUM Implementation - Yudha
Indah Prihatini, Adi Permadi, Wahyu Novian Condro
Murwanto, Rendy Munadi
[18] “BIG-IP” - http://www.f5.com/products/big-ip/
[19] “IP Virtual Server” -
http://www.linuxvirtualserver.org/
[20] D. Bryan, B. Lowekamp, C. Jennings, “A P2P
Approach to SIPRegistration and Resource Location,”
draft-bryan-sipping-p2p-02, IETF, March 5, 2006.
[21] D. Bryan, B. Lowekamp, C. Jennings, “SOSIMPLE:
A Serverless, Standards-based, P2P SIP Communication
System,”International Workshop on Advanced
Architectures and Algorithms for Internet Delivery and
Applications (AAA-IDEA), June 2005.
[22] http://www.networkdictionary.com/Telecom/VOIP-
Architecture-based-SIP.php
[23] SIP- Session Initiation Protocol - J. Rosenberg, H.
Schulzrinne , G. Camarillo, A. R. Johnston ,J. Peterson,
R. Sparks , M.Handley, and E. Schooler.
[24] http://www.voip-info.org/wiki/view/SIP
[25] http://www.ultramonkey.org/
[26] http://www.Kamailio.org/w/
[27] http://www.freeswitch.org/
[28] http://www.counterpath.com/x-lite.html&active=4
[39] http://www.linux-ha.org/wiki/Heartbeat
[30] http://www.ultramonkey.org/3/ipfail.html
[31] http://wiki.wireshark.org/Gratuitous_ARP
