Copyright 2019 Thomas Schwengler. Significantly updated and completed in 2019 Edition available here.

Chapter 10
Wireless System Performance

We have seen a number of standards so far in this text, as well as techniques used to provide robust wireless communications. This chapter reports on the performance seen in tests conducted on several standards in different environments.

10.1 Wi-Fi System Performance

Wi-Fi systems and services are evolving from simple stand-alone access points (AP) to complex networks. We review here Wi-Fi data rates, and examine some extensions to larger systems.

10.1.1 Data Rates

Wi-Fi systems rely on IEEE 802.11 a, b, g, n, and ac. Early systems use frequency hopping schemes and direct spread spectrum, but most systems now focus on the OFDM physical layer. Their data rate of is determined by the type of modulation and coding: from BPSK 1/2 to QAM256 5/6. Physical data rates for 802.11a/g are quoted up to 54 Mbps, but maximum payload or user data throughput cannot exceed 36 Mbps, and actual measured throughput vary with suppliers (up to 30 Mbps); and interoperability between suppliers introduce even greater variations.

Table 10.1: Wi-Fi maximum rates, theoretical and actual measured throughput in indoor lab, with controlled environment and low interferences for 802.11g, 20 MHz channels at 2.4 GHz. (Actual rates are an average over different brands of access points and client devices).

Modulation20MHz sensitivitySNR11a/g PhyMax payloadActual
(dBm) (dB) (Mbps) (Mbps) (Mbps)

BPSK 1/2 -90.6 6.4 6 5.2 2.8
BPSK 3/4 -88.6 8.5 9 8.1 4.3
QPSK 1/2 -87.6 9.4 12 10.6 6.0
QPSK 3/4 -85.8 11.2 18 16.0 9.0
16QAM 1/2 -80.6 16.4 24 19.0 11.8
16QAM 3/4 -78.8 18.2 36 25.9 18.1
64QAM 2/3 -74.3 22.7 48 31.6 24.0
64QAM 3/4 -72.6 24.4 54 36.2 26.8

Table 10.1 summarizes typical data rates observed in a 20 MHz TDD 802.11g channel. These results were measured with low interferences and all client devices near the AP and within the same room, leading to system operations at maximum modulation.

802.11n and ac further increase throughput by various new schemes such as: lower guard interval, higher coding rate, greater channel width, and MIMO. Here again actual throughput are lower (of the order of half the PHY rate).


Figure 10.1: Wi-Fi aggregate throughput measured with 802.11n and 802.11ac access points with a number of client devices at close indoor range. Given the variations in multipath, we chose here to plot throughput as a function of RSSI, rather than distance estimate. 802.11n measurements were conducted in 40MHz channels (2x2 MIMO), 802.11ac measurements were made in 80MHz channels (3x3 MIMO).

10.1.2 Municipal Wi-Fi

Municipalities have tried a few years back to deploy affordable wireless networks free to use for their citizens. The common opinion is that Wi-Fi networks would be cheaper than other mobile wireless products because: 1) spectrum is free and 2) Wi-Fi client devices are free since they are part of laptops and tablets. Other arguments include 3) access points are cheap (when compared to cellular equipment) and 4) they are easier and cheaper to install (when compared to cellular towers)

The first two facts are a great advantage for Wi-Fi since they are an important part of the cost of providing service, but come at the cost of a fairly weak link budget, which greatly impacts operations. [154]

Case study: Longmont

The city of Longmont contracted Kite Network in 2006 to provide Wi-Fi Internet access throughout the city. The network was built to cover an ambitious 22 square miles of Longmont, and was completed in less than 90 days, which is impressive when compared to other initiatives. 1 [155] The city of Longmont was at the time also covered by several cellular networks, including an EV-DO network, which provides for interesting comparisons.

Network Cost Cost of rollout, spectrum, and recurring expenses for both EVDO and Wi-Fi can be estimated from various sources: EVDO rollout costs follow typical urban cell site costs (seen above) as if the network were rolled out in a greenfield manner.

Wi-Fi numbers in this case were published 2; it turns out that the build-out cost for both networks (5 EV-DO sites, and 600 Wi-Fi AP’s) are fairly similar. Cost of spectrum is estimated from published auction results in that area prorated for the Longmont population. Spectrum is of course a high cost for EVDO, while free for Wi-Fi (with its inherent uncertainties). Recurring cost such as power, real-estate, wired backhaul, is significant for both networks: EVDO needs a few high-cost, large towers; Wi-Fi needs for a similar coverage cheaper but more numerous AP locations.

Operational Cost The operational cost of maintaining wireless networks is the main difference here: Wi-Fi costs more than $4 million annually (per SEC filing), 3. EVDO operations are cheaper: the network has fewer sites and requires fewer field technicians for maintenance issues; we estimate that cost to $2.7 million for a city the size of Longmont. Several factors contribute to the operational cost differences between the two types of networks: costs of repair for instance: repairing a Wi-Fi access-point usually requires bucket-crews to reach traffic poles or street lamps while EVDO electronic equipment is on the ground. Trends also seem to show that the frequency of repair of a relatively cheap access-point is far higher than that of a more expensive EVDO cell site. Finally scaling up a network to much wider areas and many more cities is more manageable with a network requiring fewer sites. The comparisons are summarized in table 10.2.

Table 10.2: Network cost comparison between EVDO and Wi-Fi for the city of Longmont.

Item EVDO Wi-Fi

Spectrum $1.4 million $0
No. of cell sites 5 600
Tower $200k $0k
Site prep $50k $500
Equipment $120k $3k

Total Build-out $2 million$2 million

Power $15k $80
Real estate leases $3k $0
Backhaul $2.4k per site $14k tot

Monthly Recurring Cost $102,000 $62,000

Field techs 2 5
NOC techs 1 1
Tech support 2 3

Yearly Operations $2.7 million$4 million

Cost comparison is fairly close and does not show one technology roll-out as a clear winner. In practicality, however, lower operational expenses and clean spectrum insuring reliable service make a big difference that seem to outweigh the free aspects of client devices and spectrum. Since then, EV-DO evolved into LTE, while upgrading access points to 802.11n or ac has been very costly. In January 2008, Kite Network could no longer manage the Wi-Fi network, and opened it for free use. 4

Other Successes and Failures

Other municipal initiatives flourished between 2006 and 2008. Some targeted public safety respondent, including use in a specific spectrum block available for free for that purpose (4.9GHz). Some aimed for city coverage extending over hundreds of square miles. For instance in Albuquerque and surrounding cities, plans even included the build-out of areas where no homes were yet built: it was argued that the presence of wireless network would be an incentive for builders. Of course designing a wireless network before buildings are in place makes no sense. 5

Successful Wi-Fi initiatives seem to have a few facts in common: 1) a reasonable target build-out, such as campus, or city public buildings, and 2) an anchor tenant insuring a minimum revenue stream to sustain operational expenses.

Other Factors

There are other key cost factors of wireless networks, including: cost of acquisition of new customers, including activation, billing, and the distribution of client device. Other considerations include usage needs and patterns, especially around indoor coverage. The majority of data need is still indoor (at home or at work). Indoor coverage from outdoors access points are very difficult to build reliably (recall link budgets and indoor penetration).

Outdoor needs have been well met by mobile devices, but recent demand for capacity is pushing again for small cells (including Wi-Fi). Fixed outdoor usage is of course limited by weather and temperature, which limit revenue opportunities.

Table 10.3: Typical temperatures and weather conditions in major US cities show the seasonal nature of demand for capacity (gathered from

City Sunny daysRainy daysDays above 90FDays below 32F

Denver 115 89 34 155
Dallas 136 79 100 39
San Francisco 140 67 2 30
San Diego 147 42 4 0
Seattle 71 150 1 20

10.1.3 Radio Parameters Analysis and Modeling

Wi-Fi coverage and capacity considerations are similar to those of any cellular network, and even more stringent since only 3 non-overlapping channels are available for a reuse factor of 3 (Although reuse factors of 4 with channels 1, 4, 8, 11 have been rolled-out, e.g. in Longmont).


Figure 10.2: Wi-Fi access points in a Longmont residential area.

In an initial design phase, a simple one-slope model and low-resolution terrain data suffice for a rough estimate to qualify customers. As operations progress, actual measurements are compared to predictions and the process is refined further.

The variations in RSSI as a function of distance were measured in a residential neighborhood and are shown here. Note that measurements taken in the street may show a fairly low slope (due to the street corridor effect).


Figure 10.3: Received power level signal strength indicator (in dBm) of a Wi-Fi system as a function of distance (on a logarithmic scale).


Figure 10.4: Wi-Fi physical modulation rate as a function of distance (on a logarithmic scale).

10.1.4 Throughput Measurements

Throughput is affected by distance, shadowing, and interferences. In most cellular trials, mobile data is collected, which makes it impossible to quantify variations over long periods of time for a given location. In a population of fixed location, however, a measured standard deviation over a long period may be useful in predicting seasonal changes in the radio channel.

10.2 WiMAX System Performance

The WiMax standard appeared in 2006 as a cost-effective solution for fixed wireless access.

10.2.1 Data Rates

IEEE 802.16 and WiMAX profiles allow for several radio channel bandwidths, which lead to very different data rates. WiMAX systems have modulation and coding ranging from BPSK 1/2 to QAM64 3/4. Measured throughput vary with suppliers: a degradation of 40% to 50% is often observed compared to theoretical throughput. Table below summarizes typical data rates observed in a 3.5 MHz FDD channel (also see figure 10.9 on page ). Actual data results vary with suppliers, and interoperability between suppliers introduce even greater variations.

Table 10.4: WiMAX 3.5 MHz channel maximum theoretical and actual measured throughput (at 3.5 GHz).

Modulation3.5MHz sensitivitySNRTheoreticalActual
(dBm) (dB) (Mbps) (Mbps)

BPSK 1/2 -90.6 6.4 1.41 0.86
BPSK 3/4 -88.6 8.5 2.1 1.28
QPSK 1/2 -87.6 9.4 2.82 1.72
QPSK 3/4 -85.8 11.2 4.23 2.58
16QAM 1/2 -80.6 16.4 5.64 3.44
16QAM 3/4 -78.8 18.2 8.47 5.16
64QAM 2/3 -74.3 22.7 11.29 6.88
64QAM 3/4 -72.6 24.4 12.71 7.74

Of course different profiles and channel widths lead to different throughput results. An unlicensed TDD 10 MHz channel profile for instance has the advantage of adapting to asymmetrical data demand. Similar benchmark tests show that such a system is also capable of throughput around 8 Mbps (see 10.2.5 and specifically figure 10.10). Interferences from other cells (co-channel interferences) strongly impact actual rates, especially at cell edge [156] [157].

To estimate system performance, Stanford University Interim (SUI) models are used to create models that emulate different terrain types, Doppler shift, and delay spread as summarized in table. Terrain Types are (from [32]) defined as follows. A: the maximum path loss category, hilly terrain with moderate-to-heavy tree densities; B: the intermediate path loss category, hilly with light tree density or flat with moderate-to-heavy tree density; C: the minimum path loss category, mostly flat terrain with light tree densities. These terrain categories are also used to model obstructed urban, low-density suburban, and rural environments respectively.

Table 10.5: Typical parameters for SUI-1 to 6 channel models (delay spread values estimated for 30-degree antennas azimuthal beamwidths, and Ricean K-factors are for 90% cell coverage — from [33]).

ChannelTerrainRMS Delay SpreadDopplerRicean K
Model Type (μs) Shift (dB)

SUI-1 C 0.042 (Low) Low 14.0
SUI-2 C 0.069 (Low) Low 6.9
SUI-3 B 0.123 (Low) Low 2.2
SUI-4 B 0.563 (High) High 1.0
SUI-5 A 1.276 (High) Low 0.4
SUI-6 A 2.370 (High) High 0.4

10.2.2 Experimental Data

Fade emulators in lab environment can recreate the above SUI fading profiles to evaluate radio systems performance in different fading environments.


Figure 10.5: Throughput variations in time of a WiMAX radio system at 5.8 GHz in two different SUI fading models: SUI-1 (left) and SUI-3 (right). Radio system modulation was fixed to 64QAM-2/3.

Radio system under test comprises one base station (BS) and several subscriber stations (SS’s). Different fading channels are programmed in the channel emulator. The Fading emulator emulates two separate channels (forward and reverse links), each comprised of several multipaths, each of which is independently faded and delayed. Fade statistics for the direct path are either Rayleigh or Ricean, delayed paths are attenuated and Rayleigh faded as specified by SUI models. In this test, the unidirectional nature of the fade emulator forces us to separate transmit and receive paths (with circulators). Additional attenuation (pad) is added where necessary. Finally a traffic generator is wired to the BS and laptops are wired to SSs for data collection. Figure 10.6 shows the detailed setup.


Figure 10.6: Fade emulator setup between base and several clients.

10.2.3 Radio Parameters Analysis and Modeling

In an initial design phase, a simple one-slope model and low-resolution terrain data suffice for a rough estimate to qualify customers. As operations progress, actual measurements should be compared to predictions and the process should be refined further.

A path loss estimate is derived from RSSI measurement (and transmitted power level, which depends on base station power, cable loss, antenna pattern, and even on the deviation from boresight of the sector’s antenna. Path loss estimates are represented in figure 10.7.


Figure 10.7: Empirical path loss (in dB) as a function of distance (on a logarithmic scale), and comparison to prediction models.

Approximation of path loss to a one-slope model leads to the following equation:

P L (dB ) = 127.2 + 27.24× log(d∕d0)

with d0 = 1 km. The trial environment is compared to typical cellular models as discussed below.

A classic approach to compute slope estimate is a least square error estimate, in which a set of error terms {ei} is defined between each data point and a linear estimate. Minimizing the sum of these errors yields the slope and intercept, which intuitively gives a good approximation of the data set. That method also benefits from the following important properties [158]:

  1. Least square estimated slope and intercept are unbiased estimators.
  2. Standard deviations of the slope and intercept depend only on the known data points and the standard deviation of the error set {ei}.
  3. Estimated slope and intercept are linear combinations of errors {ei}.

From the last point, an assumption that the errors are independent normal random variables (as in a log-normal shadowing situation) leads to a slope and intercept also normally distributed. If we assume more generally that the data points are independent, the central limit theorem means that the estimated slope and intercept tend to be normally distributed. For the last assumption to be true, very low correlation of the wireless channel must exist between data points. This is the case when data points are measured at fixed locations tens or hundreds of meters apart, or, in a mobile cellular environment, from one cell to another.

The important point is that path loss exponent is approximated by a normal (or Gaussian) random variable. We also verify a few more key findings as in [32], for a 3.5 GHz fixed link:

  1. Free-space model (PL0 = 20×log(4 πd ∕λ)
    0) works well within 100 m.
  2. Path loss exponent depends strongly on height of transmitter (mobile height being more or less constant throughout).
  3. Variations around median path loss are Gaussian within a cell (Log-normal shadowing), with standard deviation σ 11.7 dB.
  4. Unfortunately, our limited number of cells does not allow us to quantify the nature of the variations of σ over the population of macro cells.

10.2.4 Throughput Measurements

Throughput is affected by distance, shadowing, and interferences. RSSI and SNR have a direct impact on the modulation and therefore on the throughput of the wireless link. That throughput is graphed as a function of distance in figure 10.8.


Figure 10.8: Throughput in Mbps measured at customer locations as a function of distance to base station, with ten point moving average, and logarithmic fit.

In fact, modulation and throughput change from time to time, their statistical distribution is illustrated in figure 10.9 (and 10.10). When compared to controlled fading environments described in 10.2.1, fading statistics in suburban areas show close correlation with SUI models 3 and 5, and throughput density functions near those of 16QAM 3/4 in such fading environments [159].

Finally we examine the standard deviation of measured signal strength over long periods of time for a given location: typical standard deviations in fixed links over several months vary between 1 and 6 dB, more as leaves come out if deciduous trees are present in the path. Trial data also show that the standard deviation tends to increase with distance, on average:

σfixed = 2.26 + 0.75 × log (d∕d0)

with d0 = 1 km.

Seasonal variations are moslty noticeable as leaves come out. The impact on the link budget has been reported for fixed wireless links [47] and in different wind conditions [44]. We measure some variations of the path loss exponent, the intercept, and the log-normal shadowing. In many cases the wireless system can adapt to these variations, but in some marginal locations where link budget nears the maximum allowable path loss, throughput is affected. As shown on figure 10.9, low bit rates are affected the most by changes in foliage.


Figure 10.9: Throughput cumulative distribution statistic measured in various foliage conditions in a 3.5 MHz FDD channel at 3.5 GHz.

10.2.5 Unlicensed WiMAX

Unlicensed WiMAX operates at 5.8 GHz, in 5 or 10 MHz TDD channels (using 256 or 512 FFT respectively). Their performance is illustrated in this section. Comparison between field and lab data shows that measurements made downtown Denver match a SUI-3 model, and that adaptive modulation seems to maintain a radio link between QPSK and 16QAM.


Figure 10.10: Throughput cumulative distribution statistic measured in a 10 MHz TDD channel at 5.8 GHz, for data measured in the field (downtown Denver) and data emulated in the lab.

Copyright 2018 Thomas Schwengler.