[NoCatNet] Re: [Community] Point-multipoint throughput loss

[Eric Johanson]

I've seen this same behavior on my node, doing similar things. 

In my case, I'm running rg1000's with linux on them.  This is using
separate 'PCs' for each link.

Random photos here: http://www.vilos.com/new_mast
http://vilos.com/new_mast/images/thumb/PICT0034-1-0.html

When I had two p2p links up, I'd see the same degradation of speed.  It
didn't help to change channels around (one on ch 1, the other on ch 11).

One theory on this is the 40mhz carrier that exists on most wifi cards is
causing interference between them.  Adding shielding or increasing the 
distance between them should help. I was going to propose using different 
brands of wifi cards, but it looks like you are doing this today 
(orinoco + Senao).

Something else that can cause this is layer 2 and layer 3 retries.  If you 
have one connection that has a low SNR, it forces retries. Under this 
model:

My Internet Internet
|
|
router
|
|
rg1000
!
!
dlink pci card
[pc]
dlink pci card
!
!
some wireless card
[MattW's place]

When Matt was downloading p0rn, he was seeing like 3k/s or less.  It was
causing 1.5mbit of retries on my internet connection.  The problem was
from one of the wireless links had a _lot_ of noise on one end, so packets
had to be retransmitted.  In this case, they came from the internet, thus
causing lots of thrashing about.  Interestingly, both 802.11 frames AND
tcp frames where being retransmitted (sometimes both, but typically 802
frames, followed by tcp frames).

You may be hitting this on your network.  Get mrtg running or sniff 
packets on the wire(less) looking for retries. 

Finally, remember that 802.11 is single duplex. Forcing your wired 
segments into single duplex will help with packet cols.

Let me know if any of this helps, or if you find a different solution!

-Eric

On Tue, 22 Jul 2003, Roger Weeks wrote:

> Here in NoCat land, we've built out a small but widely distributed network.
> Due to the hilly nature of Sonoma County, we've had to choose sites for relay points that are on hilltops.
> My own connection to this network, for example, takes radio out of the valley I live in across 2 miles to a ridgetop, where we have a Soekris 4521, with two Senao 200mw cards.  The signal from this point flows via radio to another ridgetop a mile south, through a Mac G3 running Yellow Dog Linux, and again down into another valley closer to town where there is DSL.
> 
> We are running into a real problem: any node in our network that is a relay - in other words, a node with more than one radio - significantly impacts the throughput of signal going through it.
> 
> Some examples:
> 
> >From the site with DSL, set up with a Stylistic using ethernet cards, and an AP-1000, testing using nttcp shows that sustained TCP transfers to the next hop flow at 4.5+ megabits/second.
> 
> At this next hop, there are two orinoco radio cards in the Mac G3, and a third radio in an external orinoco AP connected via ethernet.  Testing shows that on the ethernet side, TCP transfers remain at 4.5+ megabits/second.  If you connect to the radio of the external AP, throughput drops to 3 megabits/second.
> 
> If you follow this path on from the external AP to another hop, a VIA EPIA-M with two Senao 200mw cards and running pebble linux, you continue to maintain 3 megabits/second throughput into the ethernet port on the Via.  If you connect to the second radio, running in hostAP mode, your throughput drops to 1.2 megabits/second.
> 
> Take this link to the final destination, another Via with an orinoco card and a Senao card installed to test this scenario, and you maintain 1.2 megabits/second on the ethernet of the Via.  Connect to the Senao, running in hostap mode, and your throughput drops to 700kilobits/second.
> 
> Go back to the Mac G3, and follow the other radio link across the valley.  nttcp testing shows 4.5 megabits between these nodes.  But if you connect to the second radio on this node, a soekris 4521 with two senao cards, running pebble linux, your throughput drops to a maximum of 2.5 megabits/second.
> 
> This is very perplexing to a number of us here.  Are we doing something wrong?  We're seeing similar problems on a number of hardware/software platforms, leading me to believe that this is a radio issue.
> 
> But I don't understand why having two radios in the same box would cause throughput to drop in half.
> 
> Any suggestions/comments/test scenarios appreciated.
> 
> Roger Weeks
> _______________________________________________
> Community mailing list
> Community@freenetworks.org
> http://shoveler.ipl31.net/mailman/listinfo/community
> 
>