[Soekris] PCMCIA + MiniPCI cards in net 4511
jim at netgate.com
Thu Sep 23 02:28:26 UTC 2004
Lutful Khan wrote:
>>>The practical problem is that there simply isn't enough bandwidth
>>>available if 1000 people are all trying to run 802.11 in a single
> The past year. we have been forced to delve deep into custom 802.11
> and 802.16 MAC design in FPGAs rather than using
> Prism/Atheros/Broadcom et al.
There is another path. Use commodity hardware (e.g. Atheros, or
potentially Broadcom) to implement a different (somewhat 802.16-like)
MAC while retaining 802.11's PHYs.
I call this "WiFiMax". :-)
Now you should openly wonder why I moved my *ss to Hawaii. :-)
OK, here is some Soekris content, the resulting software-MAC may, or may
not run on a 45xx series board. Its likely to run on a 48xx though.
> It may be possible to tremendously increase the bandwidth (1Gbps+) and
> number of users (10K+) if all 11 channels are digitized together at
> one AP which could run multiple SSIDs.
Sure, you could just use the entire bandwith of 2.4GHz and schedule the
MAC. The various (proprietary) "turbo" modes come close to doing the
first half of this.
The problem here is that you won't be able to build the inexpensive
clientseveryone expects. I won't go into the receiver structural issues
(this *is* the Soekris list, after all.) And there are no standards to
do this with, so attracting the diverse market that 802.11 enjoyes (with
all of its investment) is unlikely.
> A company called Engim is working on a whole spectrum RF front-end,
> but it can be done even now ($$$) using the latest AD/DA converters,
> wideband receivers, and FPGA-DSP.
Ah yes, Engim... their problem is fundaamentally that even if they can
recover all three "non-overlapping" channels via DSPs (they don't use
FPGAs, btw, far too slow. Nope, they have full-custom silicon, fabbed
at IBM) the transmits from the AP (say, on channel 6) will still cause
interference at the clients that aren't co-channel.
Read: its the same problem we've been talking about.
Read: "Short Engim" (except they're not public. If they were a short
would be a fine bet to make.)
Engim also have a pretty severe near-far problem (so they limit the
antenna gain you can run).
So much for that.
I'm sure that most of the list thinks that slapping a card into a 4526
and supplying pigtails and antennas will get it done. Its a fine
aproximation, for sure.
Quick question: How many of you put both antenna leads on an 802.11g
card? Much as I love you to buy extra pigtails, you're wasting money.
Oh sure, your card vendor's datasheet says "diversity",
("Dual MHF antenna connectors (also known as Hirose U.FL) for diversity
but they dont' say "per-packet" diversity, now do they? Consider how
short the OFDM preamble is. (10 short symbols in 8us, followed by 2 long
symbols of 4us each.)
The first 10 symbols (the "short" symbols) are used for establishing AGC
and the coarse frequency estimate of the carrier signal. Typically five
to six of the ten are consumed during RSSI, AGC, and timing recovery.
The receiver uses the long symbols for fine-tuning.
This leaves perhaps 3 short symbols (in the first of the incoming frame)
which can be "dropped" while the baseband searches the antennas for the
"best" signal. In that time (2.4us) you have to sample, switch, settle,
sample, compare and potentially switch and settle again.
Now compare this to 802.11b's 128 bits (out of the 144-bit PLCP) of
sync, which are (always) sent at 1Mbps. This is 128us in which to
sample, switch, settle, sample, compare and (potentially) switch and
settle again. Compared to 2.4us, 128us is pure luxury.
Note that I was not vendor-specific, so please don't read this as my
knocking any particular vendor's implementation.
(*) to be perfectly fair, even the netgate website gets it wrong.
More information about the Soekris-tech