Wireless network design principles
CSAIL wireless network design principles
The CSAIL wireless network has evolved quite a lot as wireless technology has changed in the 17 years since the original network was specified for the building. This document is an attempt to set out the current (as of 2021) design principles that we apply when making changes to the wireless infrastructure.
To the extent possible, we make changes on the basis of on-the-spot surveys conducted using Ekahau Pro, an industry-standard software package for wireless network analysis. Ekahau Pro has excellent facilities for surveying wireless network coverage and visualizing the results; however, localization of the survey equipment is performed manually by an operator. Thus, in order for a location to be surveyed, and therefore taken into account for coverage planning, we must be able to both physically access it and locate it on a map by dead reckoning. Parts of offices to which access is blocked by large piles of junk, oversized desks, or other obstructions cannot be surveyed and will not be considered in coverage planning. (We normally use Ekahau’s visualization in a mode that will interpolate from measured points but the architecture of Stata is such that this sometimes gives erroneous results.)
All of our wireless access points have a 1000BASE-T wired backhaul; we do not (at least not intentionally) operate any “repeaters” or “mesh mode” access points, because this significantly reduces the capacity of the network and exacerbates the already severe “hidden terminal” situation caused by the architecture of the building. As a consequence, when considering when an access point can be located, we must identify how the backhaul connection can be provided. Sometimes, it may not be feasible to provide wireless service in a particular location, because there is no CSAIL network wiring nearby. While we can install wiring, this is a very expensive proposition in Stata (about $1,600 per location not including the access point itself) and we must budget for this expense.
Most of today’s wireless network technology was designed to serve large open-plan offices with only free-space attenuation between an access point and all of its clients. This is emphatically not the situation in Stata. Unless we were to put an access point in every office, which is both impractical and expensive, most clients will be separated by multiple sheetrock walls from where their nearest access points are located. Corner offices in particular will often need to receive some signals through the masonry- and metal-clad exterior walls of the building. This implies that we must have a greater density of access points than is typical, in order to deliver sufficient signal strength in enough locations to penetrate every office.
As of 2021, the CSAIL wireless network consists of 82 access points:
- 1 in the Toyota lab on P2
- 1 in the basement
- 11 on the second floor
- 12 on the third floor
- 16 on the fourth floor
- 4 on D5
- 7 on G5
- 8 on G6
- 8 on G7
- 5 on G8
- 5 on G9
(The number of access points naturally decreases with floor area, hence the largest CSAIL-occupied floor has the most access points.)
These access points are of four different hardware types:
- MR14 (very old, 802.11n only, limited channel selection)
- MR34 (somewhat old, 802.11n and 802.11ac “Wave 1”)
- MR42 (newer, 802.11n and 802.11ac “Wave 2”)
- MR46 (newest, 802.11ax)
The different models are not distributed in any systematic fashion, but simply what was current at the time they were purchased. The old MR14 access points are only in very low-volume areas and are being phased out by cascading when newer devices are upgraded. We are slowly replacing the highest-use MR34 access points with higher-performance MR46 devices as the opportunity presents itself.
Limited service on 2.45 GHz
A major downside of needing so many access points for office coverage is that the access points interfere with each other in the more open areas of the building, especially in the two- and four-story open spaces. This is a particularly serious problem for the 2.45-GHz frequency band, used by many older devices, which has both greater signal penetration and far fewer available channels (only three, as compared with 22 for the newer 5-GHz band). The interference problem on 2.45 GHz is so severe, in fact, that we provide at most “best effort” design for it: there is no possible 3-coloring of any space below the 9th floor of Stata, due to adjacencies between floors caused by the double-height spaces. Thus, if a user has a wireless issue and their device is only capable of using the older IEEE 802.11g standard, or it supports 802.11n but only on the 2.45-GHz band, we will advise the user to update their equipment. We configure a minimum data rate of 18 Mbit/s, so devices that support only 802.11b or “classic” 802.11 (which would be old enough to count as “retrocomputing” these days) cannot connect to the network at all: at least 802.11g is required. (This is necessary to make the 2.45-GHz band usable at all, because there is a lot of broadcast traffic on the CSAIL network and broadcasts are transmitted at the lowest supported data rate, leaving less air time available for full-speed communications.)
Design considerations for 5 GHz
While the 5 GHz frequency band does have significantly more channels available, the band is split into multiple sub-bands which have different legal restrictions. For the U-NII-1 and U-NII-3 bands, which are the only bands supported by consumer-grade wireless access points, there are different power limits for each sub-band (channels 36–48 and 157–161). Many consumer devices can operate on channel 165, which overlaps the U-NII-3 and ISM bands, but our access points cannot, so it’s a great choice if you need a private wireless network.
For the remaining bands (collectively referred to as U-NII-2, channels 52–64 and 100–144), federal law designates wireless networks as a secondary use, with the primary use being radar systems; access points operating on this band must implement “dynamic frequency selection”, continuously monitoring for any sign of a protected radar signal. If an access point detects a radar signal, it must immediately (within milliseconds) cease operating on its configured channel and switch to an unrestricted channel in the U-NII-1 or U-NII-3 band. The certification tests for this are sufficiently expensive and onerous that manufacturers generally choose not to get certification for consumer-grade equipment, which is why low-end access points usually do not support operation on U-NII-2 channels.
The Boston Logan Airport Terminal Doppler Weather Radar (TDWR) operates on 5.610 GHz, which is wireless channel 120. CSAIL’s wireless equipment cannot use channels 120 or 124 due to FCC restrictions. (MIT campus wireless does use both channels, I am not sure by what legal authority — campus wireless uses a different hardware vendor who may have received different certifications.) In addition, channel 144 overlaps the boundary between the U-NII-2C and U-NII-3 bands; while our access points support this channel, none of our monitoring and survey equipment supports it, so we do not use it.
In the 5-GHz band, the 802.11ac (“Wi-Fi 5”) and 802.11ax (“Wi-Fi 6”) standards allow for bandwidths of 20 MHz (single-channel), 40 MHz (two channels), 80 MHz (four channels), or 160 MHz (eight channels). Many consumer devices come by default with 80 MHz bandwidth configured; in some cases (e.g., Apple devices) this cannot be configured at all. Due to the high density of access points, both CSAIL and MIT networks in Stata use 20-MHz bandwidth exclusively; private access points as well must be both configurable, and in practice configured, to use 20 MHz — there simply are not sufficient channels available for use of any other bandwidth.
This does imply that clients cannot get anything close to the speed advertised by consumer-grade wireless equipment (which is intended for use in a detached single-family home with no other nearby access points). With 20 MHz bandwidth, the maximum speed of an older 802.11n client, or an 802.11ac client with a single “spatial stream” like a phone, is about 75 Mbit/s. A newer 802.11ac Wave 2 or 802.11ax client with three spatial streams (like a typical modern laptop) can achieve about 225 Mbit/s in practice.
(There is one significant exception: the Bliel conference room (32-G601) is located far from any CSAIL or MIT wireless network, and has only a single access point. We configure that access point to use 80-MHz bandwidth since it cannot interfere with other users.)
Objectives for network planning
When planning whether to add or relocate an access point, or whether to reconfigure radio parameters, we keep the following objectives in mind:
- Avoid co-channel interference. As noted above, the architecture of Stata means that in many places, two access points may each have clients in a zone where their signals overlap, but not be able to hear each other — resulting in a so-called “hidden terminal” problem, wherein both access points try to transmit to different clients at the same time, but a client in the overlap region can’t decode the signal intended for it because of the interference. We try to the greatest extent possible to ensure that no client can hear two access points transmitting on the same channel. This effectively requires us to use all 20 of the usable channels in the 5-GHz band (and, as noted above, is completely impossible in the 2.45-GHz band).
- Provide at least two usable signals in every location. Between outages caused by radar detection, hardware failures, firmware updates, other maintenance, and simple capacity limitations, it will sometimes be the case that the closest access point to a given location will be unable to service all clients. We try to ensure that there is a second access point within range of every office, although due to the geometry of Stata this isn’t always feasible. When examining survey results and considering access-point placement, we look at the received signal strength from the second- and third-strongest access points at every location.
- Ensure that at least one signal in every neighborhood is on a U-NII-1 or U-NII-3 channel. This ensures that, even in the case of a radar hit that takes out multiple U-NII-2 channels, at least one channel is available that is not subject to the Dynamic Frequency Selection rules and will not have to change channels.
- Avoid adjacent-channel interference. To the extent possible, we try to avoid having the strongest and second-strongest signals in any location be on immediately adjacent channels (i.e., channel numbers 4 apart). This is impossible for the 2.45-GHz band, so we only consider this factor when choosing channel assignments for the 5-GHz band.