There are some VoIP metrics that you really need to focus on to manage VoIP, such as jitter buffer discards. But traditional methods of viewing network performance statistics are not so useful and often misleading when trying to understand the quality provided by and performance of VoIP networks.
The problem is that most statistics are provided as averages or percentages, such as "packet loss was 5%" or "the average jitter level was 5 milliseconds". These types of numbers paint a very misleading picture as one would naturally think of very regular rates, i.e. every 1 out of 20 in the loss example. The truth of the matter is that packet networks are bursty. That means they experience periods of high loss (and discards) and periods of low loss (and discards), i.e. periods when the loss and/or discard rate is well above 5% and periods when they are well below.
That level of understanding paints a very different picture that reveals problems in application or network performance as opposed to the implicit problem smoothing or concealing function averages perform.
This is especially true for real time applications such as VoIP. For example, most, it not all, vocoders or codecs employ packet loss-concealment (PLC) algorithms that are intended to moderate or nullify the effects loss and discards. The algorithms are very capable of handling random or regularly occurring low loss conditions, such as every 1 out of 20. PLC algorithms will generally replay the last packet, interpolate the next packet, insert comfort noise, etc., making the lost packet imperceptible to the listener. But suppose you had a call that lasted for 3.5 minutes, spanning 10,000 packets, and that lost 500 of them (5%) had been lost or discarded. If the packets had been lost fairly regular intervals, then the listener's opinion of the call quality would likely be very high. However, as we know packet networks are bursty, where the distribution of the loss and discards can be quite concentrated, such as 10 bursts of 50 lost or discarded packets each. That would likely annoy the listener considerably, and probably lead to termination of the call, because concealing the loss of more than a couple of packets in a short period of time is well beyond the capability of loss-concealment algorithms.
The important point is that for VOIP, quality assessments can be made only by understanding the actual distribution of impairments, not just the averages (especially the mean). Therefore, you should have tools that count the number of bursts, the number of packets lost or discarded in bursts, the number of periods of high loss, the burst length and burst density and so on.
About the author:
Bob Massad is a VP at Telchemy, Inc., a provider of voice quality management solutions.