Predicting Offered Traffic Load

The way we use the Internet has a number of interesting implications for offered traffic
distribution. The average holding time for an Internet session is 30 times longer
than a traditional voice call. Fortuitously, Internet busy hour is not the same as voice
busy hour. Session call length peaks at 4:00 A.M., at over 80 minutes. (It is probably best
not to speculate what people might be doing on the Internet at four in the morning.) If
we can fill the whole of our loading box in Figure 14.7, we have doubled our network
bandwidth utilization. We might not have doubled our revenue, but we will definitely
have increased our revenue and margin.
We need AMPU (average margin per user) rather than ARPU (average revenue per
user). SMS is good because it costs relatively little to send but can be billed at a disproportionately
high rate. SMS accounts for 10 percent of revenues but 20 percent of margin
in a typical European network.
Delay-tolerant content and delay-tolerant applications are useful for quiet-hour loading.
Application or software downloads to handsets and image uploads from surveillance
devices are both examples of delay-tolerant content. The loading shift can,
however, only be achieved by having buffered bandwidth available both in the device
(handset or surveillance device) and the network. It may even be worth subsidizing
additional handset-resident/device-resident memory to encourage this load shift effect.
Offered traffic has the habit of being in the wrong place at the wrong time. We can
support the offered traffic load by buffering. The buffers can by milliseconds, seconds, or hours, depending on how delay-tolerant our data is and what we are trying to
achieve. It may be that our highest-value offered traffic is very delay-sensitive. We may
want to completely avoid buffering—effectively to circuit-switch or virtual circuitswitch
the data (or rather, session-switch the data).
We can only dimension the network accurately if we can predict what the future traffic
mix will be and what traffic quality will be required. We can predict the likely traffic
mix if we can predict the likely mix of devices producing the offered traffic load. Take a
million subscribers and determine what the likely hardware and software device mix
will be in 3 to 5 years’ time. How many subscribers will have digital cameras in their
phones? Will the digital cameras be CCD or CMOS? How many subscribers will have
65,000-color screen displays, and display drivers capable of supporting 12, 14, 16, or 24
frames a second? The subscriber product mix will determine uplink loading distribution
and downlink load distribution. The subscriber product mix will determine uplink
loading traffic properties and downlink traffic properties.
The traffic properties will determine radio and network bandwidth quality requirements,
which, in turn, will determine radio and network resource requirements—how
much network density is needed, how much backhaul is needed. Defining the subscriber
hardware/software product mix also helps us to define how bursty the bandwidth
will be, which, in turn, determines how much signaling bandwidth we need and
how much additional traffic bandwidth we need.
Handset hardware and software determines offered traffic loading. Offered traffic
loading determines future network value and future network cost, and enables us to
calculate future network margin.
Summary
Traffic is coming from an ever-increasing number of sources—computer products,
consumer electronics products, IT products—and may come from or go to an everincreasing
choice of wireless, wireline, or (digital) TV networks. The traffic is becoming
increasingly asynchronous, which means burstier bandwidth. Traffic properties (and
traffic value) have to be preserved as traffic is moved into and through wireless, wireline,
and TV or radio networks.
An increasing percentage of this traffic may pass over the Internet. This means that
we need to qualify the impact of bursty bandwidth on Internet protocols—the impact
of Internet protocols on bursty bandwidth. The Internet is a queued network. Queued
networks do not like bursty bandwidth. Bursty bandwidth compresses buffer bandwidth.
Compressed buffer bandwidth triggers transmission retries, which, in turn,
absorb radio and network bandwidth and introduce delay and delay variability
(which compromises session value). Cellular networks have conferred the gift of
mobility to the user experience. Unfortunately, users now take mobility for granted
and expect wireless networks to perform identically to wireline networks.
The performance available from wireline networks, particularly circuit-switched
wireline networks, provides a benchmark against which wireless networks are judged.
The benchmark is determined by bit rate and bit quality. Bit quality is determined by metrics such as delay and delay variability, and bit error rate. These quality metrics
have a direct impact on session quality. Session quality is determined by the quality of
the network (and the radio access layer) and how well the network can accommodate
rapid changes in session amplitude. Session amplitude is determined by the dynamic
range of each user’s offered traffic and the aggregated dynamic range of multiple users
when multiplexed together (the per-user and multiuser multiplex).
Inconveniently, bursty bandwidth does not smooth when users are multiplexed
together and may become more bursty. This tends to put network components (RF
components and router buffers) into compression, which produces packet loss and
retries. One answer is to overprovision the radio access network and core network, but
this increases delivery cost.
Ultimately, we need to reassure ourselves that the additional delivery overheads
implicit in managing the rich media mix can be recovered from higher tariffs. We want
to avoid overprovisioning, because this adds unnecessary cost. We want to avoid
underprovisioning, because this compromises content value. We can only dimension
networks accurately if we can predict the product mix:
 How many handsets or devices there are in our network?
 What is their image capture capability (image bandwidth)?
 What is their video capture capability (video bandwidth)?
 What is their audio capture capability (audio bandwidth)?
This determines uplink offered traffic. Display capabilities (display and display driver,
and audio driver bandwidth) determine downlink offered traffic.
A view also has to be taken on the peak-to-mean loading on the network and
whether or not we provision (overprovision) for peak loading. These loading issues
become particularly complex when we consider trigger moments on the downlink that
create a substantial instantaneous peak in demand for uplink bandwidth.
Network planning still tends to be focused on downlink bandwidth provision. In
reality, network value is substantially moving toward uplink-generated value. Uplink
quality (the ability to handle highly bursty uplink offered traffic) will be a key future
requirement. 359