Finding And Fixing VoIP Problems
BY TOM GALLATIN
VoIP quality must be measured to ensure appropriate customer
satisfaction. Recent state-of-the-art technology is significantly advanced
by VoIP quality monitoring and testing systems based on the E-Model, an ITU
G.708 standard scheme whose key contribution is that it passively measures
each call more accurately without having to generate costly and intrusive
test calls. Administrators may use systems that are E-Model based to
proactively measure the behavior of their VoIP network, baseline the
performance, and understanding the human perception of the call quality,
thus predict the MOS score. Passive techniques are best for bringing the
problems to light, providing accurate information that enables a rapid fix.
Everyone wants the option of having high-quality phone calls with high
availability at low cost. Voice over IP networks promise this, but are still
far from delivering fully on that promise. Jitter and packet drops that can
be tolerated in an IP data network are key contributors to poor quality in
VoIP. To absorb most jitter, buffering is often employed but buffers can
overflow and cause drops plus significant delay that is also a cause of poor
perceived quality. High bandwidth (the ï¿½big pipesï¿½ solution) can eliminate
much of the buffering and drops, however, bandwidth is not inexpensive and
is not a panacea for all that can occur on even the most robust of IP
networks. It only masks root causes leaving problems to grow. They must be
found and fixed. Persistent quality problems will lead to customer
dissatisfaction. It is essential to pro-actively monitor the quality of
every call made in a network in order that quality issues can be detected
and resolved immediately.
Before you can fix the problem, you have to identify and find it.
Compounding the problem is that analog voice tools and tests donï¿½t analyze
VoIP in Ethernet networks. Some voice quality monitoring and analysis tools
are based on traditional analog tests and depend on generating an active
load on the network, which can give misleading results in IP networks
thereby making problems difficult to find.
How can network architects get an accurate reading of the voice quality
Passively measuring each call has been made possible with the ITU G.708
standard E model with extensions to recognize the role of human behavior in
assessing call quality. It has become a popular standard for voice over IP
quality measurements. The R-factor is a dimensionless metric based on
E-Model, which can be calculated, based on observed traffic flow for every
phone call on the network. The R-factor, with human behavior factored in,
can be correlated to a subjective Mean Opinion Score (MOS), giving the
network administrator an accurate, unbiased, independent accounting of the
quality of the calls as they would be assessed by human listeners.
R-factor measurement helps solve the four major problems with existing voice
quality measurement techniques:
1. Measurements can be made on the actual calls in the network there by
eliminating false positives on test calls in a non-deterministic environment
2. PSQM, PESQ, and PAMS tests are based on analog network tests, and are
intrusive, which means that special test calls must be made through the
network in order to make measurements. This uses network bandwidth and
actually reduces quality. It is only feasible to make a small number of such
test calls and to get occasional snapshots of network performance. This is
often not effective for non-deterministic IP networks.
3. Typical network impairments are time varying, which means that voice
quality varies during a call and the perceived quality during a transition
differs significantly than the actual measured quality. Active voice quality
measurement tools are not able to predict the impact that time varying
quality has on the human listener. Extended versions of the E-model which
consider human psychological reaction to time varying impairments are able
to accurately measure the human userï¿½s experience.
4. Packet Loss and Jitter are recorded as averages as opposed to being used
to actually measure the perceived quality of the call by the human listener.
PACKET LOSS BURSTINESS
Packet loss will increase when jitter buffers are minimized to eliminate
delay. Uniform packet loss on a call usually will result in very little
quality impact, even if the packet loss is a relatively high percentage of
the total packets involved in the call. However, the degree of ï¿½burstinessï¿½
has a major impact on voice quality. This is significant because packets
will usually occur in bursts due to rerouting or sudden network spikes.
As a simple example, consider a three-minute Voice over IP phone call during
which there is a two-second period of time during which five percent of the
packets are lost. Many voice quality analysis systems would average this
packet loss over the call, assuming that a uniform five percent of packets
are being lost, which would not be apparent to a listener. This would result
in the incorrect prediction that the subscriber would not hear any
degradation in quality. The R-factor can take the ï¿½burstinessï¿½ of the packet
loss into account and make a more accurate prediction of the actual MOS
perceived by the listener.
THE HUMAN PERSPECTIVE -- SHORT TERM MEMORY
The call quality reported by a listener depends on how near the end of a
test call the degradation occurred. This is believed to be due to the
ï¿½recency effect,ï¿½ which relates to a human characteristic of tending to
remember and penalize only the most recent events. If an event was early in
the test call, and the quality improved, then the listener tends to forgive
and forget, resulting in a higher test score. Moving a garbled signal from
the beginning to the end of a 60-second call results in a change to the
reported MOS score from 3.82 to 3.18.
COMPARISON TO PSQM / PAMS
One important measure of the effectiveness of voice quality measurement
algorithms is their ability to predict the subjective quality score that
people would assign.
A recent study by Telchemy Corporation
compared the R-factor score generated by E-Model based systems with Human
Factors extensions against the widely used PSQM and PAMS active, objective
measurement algorithms. Six sets of audio files were used in the study, each
containing five impaired files. Listeners were asked to rank each set of
five files from best to worst and then the average rank was calculated for
each file. The three voice quality measurement algorithms were then used to
perform the same ranking.
The mean rank distance was calculated for each file set and for each
measurement algorithm. This gives a measure of how closely the measurement
algorithms predicted the ranking given by listeners. As an example, if the
order given by the human listeners was 1 3 2 5 4 and the order given by one
of the measurement algorithms was the same the mean rank distance would be
0; if the order given by the measurement algorithm was 1 2 3 5 4 then the
mean rank distance would be 0.4.
INSTRUMENT THE NETWORK
Large enterprise networks require distributed solutions for centralized
management. In these scenarios, network costs can be mitigated through a
device called a tap. Taps enable monitoring, analysis, or security
instrumentation to be dynamically inserted into data network links without
affecting the network. They also enable use of distributed monitoring
devices on different key backbones without requiring the purchase of a
device for each segment.
A tap goes beyond the access granted by a switch span or mirrored port
because it provides access to the actual network and to any physical errors
without impacting the performance of the switch. Taps are fault-tolerant,
passive to the network, and allow for the monitoring, capture, and analysis
of physical errors on individual segments, or in the case of matrix
switching taps, have the ability to rove between segments. Taps may provide
advantages over span or port mirroring, which usually filter traffic and
When problems occur, network administrators do the basics pretty easily.
They investigate the change management system to find out ï¿½who changed
what,ï¿½ ï¿½pingï¿½ the usual suspects, and look at management information bases
of the VoIP equipment vendor. They may even swap gear to troubleshoot the
problem. If the problem is still not solved, the network administrator will
be grateful for non-invasive test tools to give him information he can use.
Most minimally instrumented networks will have one system deployed
strategically on the critical links, which the technicians can access from
anywhere. No flights, no guessing at vendor logs, just pure data from the
network itself to help technicians troubleshoot the problem.
With most E-Model based test tools, measurements can be performed on each
channel -- audio and video -- within every call on the IP network. These
measurements go beyond what is provided by the IP gateway, switch, and
router vendors. The accuracy of the QoS measurements from third-party test
tools is unbiased. This measurement information can be exported as a
historical baseline into a centralized repository where it can be warehoused
for capacity planning, network service auditing, and customer-service
applications. When trouble strikes, the network administrator may compare
current traffic with baseline traffic patterns and will know if the problem
was within the network rather than some customer-premises equipment or rogue
application. Given the inherent chaos of IP, it is important for the network
administrator to know what is really happening.
If you need to know the voice quality experienced by your customers, then
you need accurate metrics. If you wait until your customers are letting you
know about voice quality problems then it is too late. You need real time
monitoring and analysis of voice quality to let you manage your service
Tom Gallatin is Product Marketing Manager at Finisar Corporation. Finisar
Corporation is a technology leader for fiber optic subsystems and SAN/LAN
performance tools, which enable high-speed data communications over Gigabit
Ethernet LANs, Fibre Channel SANs, and MANs. For more information, please
To The May 2003 Table Of Contents ]