The 'Wi-Fi At Conferences' Problem

[David Farber <dave () farber net>]

Begin forwarded message:

From: Jim Thompson <jim () netgate com>
Date: October 11, 2009 10:08:21 AM EDT
To: dave () farber net
Cc: Brett Glass <brett () lariat net>
Subject: Re: [IP] Re: The 'Wi-Fi At Conferences' Problem

Dave, for IP, as you desire.

On Oct 10, 2009, at 3:47 AM, David Farber wrote:

> Begin forwarded message:
>
> From: Brett Glass <brett () lariat net>
>
> Because our ISP has developed BSD UNIX-based traffic shaping and  
> anti-bandwidth hogging software for its own networks, we have also  
> deployed it for conferences. It does a good job of keeping each  
> user's bandwidth use to a sane level, and is capable of  
> administratively blocking activities that have no place on a shared  
> conference LAN.
>
> We also use RF tricks. One of the best is to use 802.11a, which has  
> lots of non-overlapping channels. (You can even use alternate  
> channels, especially if you use the lower, indoor-only ones.) Also,  
> when they process 802.11a, the chipsets also tend to be better at  
> filtering out unwanted signals on nearby frequencies. 5 GHz signals  
> also penetrate walls poorly, which is a good thing; it allows  
> frequency reuse and helps to avoid interference from sources outside  
> the building or in other parts of the venue.

Hmm,

First, an apology.  My previous message was poorly-edited.  My  
concentration and memory recall seem somewhat less than I'd like since  
my late August "open heart" surgery.  I've tried to improve below.

Lets start with what was.

To my knowledge, none of the 802.11b radios formerly on the market had  
enough selectivity to eliminate in-band signals from a nearby radio  
operating on an adjacent (25Mhz away) or alternate (50MHz away) channel.

Minimum IEEE specs for an 802.11b receiver is 35dB of adjacent (25MHz  
away) channel rejection. IEEE doesn't publish specs for alternate  
channel rejection on 802.11b, but I can tell you that the best designs  
(in terms of ACR) are the "old" super-het receivers. Intersil  
(subsequently Connextant)'s Prism2/2.5 is good for about 41dB of ACR,  
the Lucent/Orinoco design was similar.  Alternate channel rejection is  
perhaps 20dB "better" with these chipset.  (Note that I have real- 
world experience with both courtesy of Vivato.)

This means that if you're operating (your Prism 2/2.5 or Orinoco  
chipset radio) on channel 11, a signal on channel 1 will be 60dB down  
from where it would be if your radio was on channel 1.

Free space path loss (**) in the first meter @ 2.4GHz is -41dB. Lets  
say you've got a garden-variety radio that puts up 32mW (15dBm) of tx  
power, and ignore antenna gain for now (so 0 dBi antennas on both  
radios). 15dBm - 41dB = -26dBm, and 15dBm - 60dB = -45dBm. These are  
the in-channel 'noise power' of the alternate channel radio for the  
chipsets detailed above, for adjacent and alternate channel operation.

Notice that it is at least 55dB (> 200,000:1) above the thermal noise  
floor (lets say -100dB for now.) Translated: you've significantly  
lowered your SINR. Moreover, this signal level is at least 40dB  
(10,000:1) higher than the 802.11b CCA threshold, so you'll quiesce  
the entire cell (no matter which radio was trying to operate) whenever  
one transmits.  Its worse than this, if one radio was receiving,  
you've destroyed the incoming packet(s).

Thus, in 802.11b, even if you pick "1 and 11", you're going to end up  
with significant in-channel power from any operation on the alternate  
channel by a close-by transmitter. This is why you can't "co-locate"  
APs in the same band on a rooftop or tower.  Now try, just *try* to  
discuss this in a rational manner with most "wireless ISPs".

Now, if you add any antenna gain to the equation, then, unless you  
manage to find really special antennas, (or use channel filters, etc)  
then the gain of both antennas affects the above in a negative way.   
At the end of the day, a 10m separation with small (2.2dBi) antennas  
was often practical with 802.11 (pre-b) and 802.11b.  By the time you  
move from 1m to 10m, the free space path loss has increased from -41dB  
to -60dB, which is "barely enough".  (do the math!)

So the minimum (and essentially optimal) 802.11b AP separation with  
2dBi antennas on channels 1 and 11, in free-space, is around 10m. Any  
closer and you're raising the noise floor in any AP or client that is  
receiving a packet.  Note that a client between the two APs may be on  
either channel.  (In 802.11, the client gets to pick which AP to  
associate with.)  Where there is one, there can be two, and these two  
might end up on different APs (and different channels), and as a  
result, interfere with each other's frames.   The answer here (and  
with 802.11g/802.11a networks) is not the little cells on a hexagon  
layout that you so-often see in the literature, but rather larger  
cells of co-channel APs, laid out on the same hexagonal pattern.  This  
minimizes the 'edge cases' just described and lets CCA (see below) do  
its job.

This situation is (much) worse today with 802.11g and 802.11a. Due to  
their (now universal) direct conversion archectures, and some  
interesting properties of OFDM, most 802.11g and 802.11a receivers  
can't muster even the 35dB of ACR that the IEEE specifies for 802.11b  
when operating in the 802.11g / 802.11a modes on adjacent channels,  
and for these reasons, the ACR specification was reduced for 802.11g  
and 802.11a receivers.

Brett is incorrect (to the limits of my knowledge) that 802.11a  
equipment has superior ACR compared to 802.11g or 802.11b gear.  In  
the IEEE specification, ACR varies by modulation rate, but the minimum  
ACR for 802.11g/a at 6Mbps is 16dB (alternate is 16dB more, for 32dB).  
This goes down to -1dB (yes, -1dB) ACR @ 54Mbps (with alternate  
channel rejection at 54Mbps down to 15dB.)  The chipset vendors I've  
spoken with during past jobs have stated that they might be able to  
get a dB or two better than the spec, (and we've not begun to discuss  
tolerance issues) but the point is moot, you have to deal with the  
clients that walk in the conference room door, and these are likely to  
be bottom-of-the-barrel minimum spec units, if for no other reason  
than the manufacturer was shaving pennies.

Brett is correct that there is much more opportunity for separating  
APs into different sub-bands of the spectrum allocated for 802.11a.   
This spectrum covers from 5.150 - 5.350 GHz and 5.470-5.825 GHz (U-NII  
rules) and/or 5.725 - 5.850 (ISM rules).   There are some restrictions  
here (5.150 - 5.250 GHz is for 'indoor-only' operations, 23dBm EIRP,  
and must use fixed antennas.  Operation between 5.250 - 5.350 and  
5.470-5.725 GHz is limited to 30dBm EIRP, and such operations must  
support both dynamic frequency selection (DFS) and transmit power  
control (TPC).  These restrictions serve to 'limit' the applicability  
of these frequencies to outdoor use, but most conferences are indoors.

So, assuming a conforming AP can be operated as low on the 5.150 -  
5.250 GHz band as possible (without violating the restriction that any  
out of band emission be no higher than -27dBm/MHz from the band edge  
(*)), and assuming that operation is under the FCC (USA) regulatory  
domain, the lowest available channel is "36" with a center frequency  
of 5.180GHz.  We've already seen (above) that alternate channel  
operation (50Mhz away) **at 6Mbps** results in a mere 32dB of ACR  
(IEEE minimum), and moving beyond a 60MHz spread is going to put the  
next AP in the middle of 5.250 - 5.350, operating on either "58" 5.280  
or "60" 5.300GHz (better).    Now we've used 2 'sub-bands' for two APs.

Assuming your (potential) clients all support operation in the 'new'  
5.470 - 5.725 GHz band, there are opportunities for 3, or possibly 4  
more APs that won't interfere with each other, or nearby clients on an  
adjacent 'sub-band'.  There is potential for one more 'cell' in the  
5.725-5.825 GHz band,
but don't let it interfere with operation in the band directly below,  
so likely operations will occur on 5.805 GHz (channel 161).  So yes,  
you can support 6 (perhaps 7) more "cells" using these bands, but  
first the gear that comes in the door (in customers hands) has to  
support 802.11a, and in particular, has to support the bands you're  
interested in furnishing.   Note that the iPhone does not support  
802.11a, and that, to date, no Android or Windows Mobile phone does,  
either.  Many of the "netbooks" that are so popular these days do not  
support 802.11a, either.

The thing you really have to understand is this: In wireless, it is a  
sin to throw what would have been an otherwise correctly-received  
packet on the floor (or destroy it as outlined above.) You really  
can't afford to have a near-by AP think the air is clear (because its  
off- channel) and then stomp on the reception of some near-by AP or  
client.

(As an aside, if you think about it, since "traffic shaping" works by  
dropping already-received frames, its use with 802.11 is also a sin.   
There are a plethora of other issues with "traffic shaping", and the  
IP list has been over the discontinuities in statements similar to  
Brett's, "administratively blocking activities that have no place on a  
shared conference LAN".)

In co-channel operations, there is a function called Clear Channel  
Assessment (CCA) that holds any given radio from transmitting if either
a) there is a strong signal using the wireless medium, or
b) there is a somewhat weaker signal, which is clearly an 802.11  
signal (that is, "this" radio has decoded the preamble), above some  
threshold.

We did a bit of work at Vivato to make several radios (13 in the 11b  
product, 6 in the 11g product) synchronize via the CCA signal, such  
that if any of them was receiving a packet, none of the others would  
transmit.  This is the key to making a "conference room" network  
really fly.  Multiple, independent APs, all on the same channel, which  
are able to co-ordinate operation between themselves in a highly- 
similar manner. Given the higher rates of 802.11g (54Mbps translates  
to about 27Mbps of throughput for a single client, but this degrades  
with increased loading by additional clients), or, better, 802.11n,  
its entirely possible to have the available 'in-room" bandwidth exceed  
the typical "Internet connectivity" bandwidth for most situations.    
(e.g. A single cell of 802.11g APs and a single cell of 802.11a APs  
could match the available bandwidth of a DS-3 exiting the conference  
center.)

Another issue at Vivato was that the 'antennas' for these radios were  
both highly directional and pointed in different directions, so the  
actual signal level might not be high enough to set CCA to a given  
client, but any the clients transmitting would ruin the reception (at  
the AP) of any other.  If you look at the patent art out of VIvato,  
one of the big ideas was "complementary" beam forming, which was a  
method to raise the signal level "high enough" for the rest of the  
clients to keep them from transmitting (and ruining the inbound  
packet.  See: <http://www.ll.mit.edu/asap/asap_03/6-page_Papers/tarokh_ASAP.pdf 
>

When you consider that channel models out of AT&T Bell Research and  
Harvard show that the 54Mbps 'rate' of 802.11g / 802.11a can only be  
achievable at a radius of 8 meters in a NLOS (non line-of-sight)  
environment, and that to maximize throughput to/from a room-full of  
conference attendees, you want to be able to send (and receive) at the  
highest rate possible, the desirability of an architecture designed to  
do same seems inevitable.  As a reminder, a 'radius of 8m' will cover  
a room where the walls are no farther apart than 37 feet.  However,  
with 802.11n, range can be as much as doubled (even through walls)  
compared to the same operation using conventional IEEE 802.11 OFDM, if  
the environment supports "sufficient" orthogonality in the channel  
matrix.  (You can think of this as 'scattering', see below.)    
However, 802.11n also supports an optional 40MHz wide 'channel' that  
can be used to increase bandwidth (and therefore modulation rate), and  
this plays against all the channel spacing discussion above.

I have ideas on how to make the above work (sans any back-end 'co- 
ordination' such as that found in several "Wi-Fi switch" products), if  
anyone wants to get in-touch.  It requires hardware modifications to  
the APs, Gigabit Ethernet uplinks, and a bit of software, but nothing  
else.

> It also helps to know your equipment and use quality gear. When I  
> attended David Isenberg's conference a couple of years back, the  
> folks who were running the network use Linksys routers, but with the  
> upstream network cables plugged into the 4 port switch rather than  
> into the upstream port of the NAT router.
>
> I believe they did this -- in a spirit of extreme "network purism"  
> -- so that they could avoid doing NAT and give every client a  
> public, static IP. (You can't turn NAT off on the Linksys routers,  
> so they couldn't just give each router a subnet.)

I don't know which Linksys model was used at Isenberg's conference,  
(and Brett doesn't say), but for at least some models of Linksys  
wireless APs, you can disable NAT.   Moreover, putting each router on  
a single subnet seems like either overkill or a waste to me.

> An interesting idea in theory, but a bad one in practice. Not only  
> did this architecture fail to prevent broadcast packets from being  
> relayed between the APs, jamming the network; it also triggered a  
> known bug in the Linksys routers that caused packets to be reflected  
> back from the access points.

Broadcast isn't necessarily bad, or even poor network design.  If your  
back-end network (hopefully a switched Ethernet) can't deal with the  
frames uplinked from the APs, then something is very wrong, or at a  
minimum, very misconfigured.

This paper from TI touches on the topic of ACR and gives additional  
detail to what I've written above. <http://focus.ti.com/pdfs/bcg/80211_acr_wp.pdf 
>  Meru also seems to have a clue about these subjects.

As for Brett's assertion that 802.11b/g signals "penetrate" walls more  
than 802.11a signals, the research doesn't tend to support him.  Yes,  
there is an approximate 3dB absorption loss difference between 2.4GHz  
and 5Ghz through walls, which is approximately linear with frequency,  
but scattering is at least as important as absorption, and the  
scattering due to the signal passing through walls is not  
appreciatively different at 2.4GHz than 5Ghz.  Further, 3dB, while "a  
lot" (2x or 1/2x, depending on how you wish to view it), its probably  
not enough to 'separate' rooms in the manner which Brett implies,  
other than in 'edge' cases.  Countering this is the simple fact that a  
given size antenna is "larger" at 5Ghz than 2.4GHz by the same  
(frequency-linear) amount, and the antenna gains at both ends add,  
though these 'larger' antennas are of necessity, more 'directive.

For more on this, see: <http://www.triquint.com/prodserv/tech_info/docs/WJ/Indoor_prop_and_80211.pdf 
> for more but note that the paper states in conclusion:

"We suggest that there is no intrinsic impairment of 5.2-5.8 GHz  
propagation vs. 2.4-2.5 GHz propagation in office/light manufacturing  
environments, and thus no intrinsic impediment to roughly equivalent  
deployment of 802.11a and 802.11b wireless LAN systems.  Note,  
however, that this does not imply equivalence of practical systems:	 
higher frequencies suffer higher losses in cables and circuit boards,  
and low-cost devices may suffer from reduced gain or lower output  
power at higher frequencies. Further, to achieve equivalent collecting  
area, higher-frequency antennas become more directional, which may be  
inconvenient for end-users."

Some of these factors may account for the effects Brett describes,  
rather than (lack of) "penetration".

Jim

(*) note to non-RF engineers, this is HARD work.
(**) yes, I'll live in my own state of sin, and speak about this as  
though it were reality rather than merely a model.





-------------------------------------------
Archives: https://www.listbox.com/member/archive/247/=now
RSS Feed: https://www.listbox.com/member/archive/rss/247/
Powered by Listbox: http://www.listbox.com