Re: [ath5k-devel] Unusually low speeds with ath5k and iwl3945

[Johannes Berg]

On Mon, 2008-11-24 at 10:43 +0100, Felix Fietkau wrote:
> Benoit PAPILLAULT wrote:
> > I did a similar test here and results is very strange. AP was my good
> > old linksys WRT54G running an iperf server. Client was a laptop running
> > either ath5k or madwifi/trunk and an iperf client. Channel is 5. Both
> > drivers show the same behaviour.
> > 
> > At the beginning, throughput was very low : 500 - 600 kbit/s. Suddenly
> > (after few minutes), it jumps to 15 - 17 Mbit/s and then few minutes
> > later (let's say 10 - 20 minutes maybe), it jumps back to 500 - 600
> > kbit/s. Using a fixed rate has no effect.
> > 
> > I used my latest wireless monitoring tools and I did not saw lost of
> > duplicates or lost packets. The only difference was the number of
> > packets sent by seconds....
> > 
> > Looking a my syslog, I just saw few messages, unrelated in time with the
> > throughput going up or down. They were:
> > - ath5k : unsupported jumbo
> > - switching to short barker preamble
> > - switching to long barker preamble
> > 
> > I can repeat the same test with iwl3945 as well, if needed.
> While reading the code for calculating the frame duration, i noticed
> something odd: It doesn't seem to be taking into account the short
> vs long preamble distinction for ERP rates. IMHO this might be causing
> issues like this. I've seen similar behaviour a long time ago when testing
> iwl3945 against a Broadcom AP with exactly the same throughput drop (500-
> 600 kbits/s).
> When I analyzed the problem with an extra monitor mode card, I found out
> that the throughput drop is caused by a huge number of retransmissions,
> and if I remember correctly (I didn't look for this specifically back then),
> the retransmissions went down the rate scaling table until they hit the
> first non-ERP rate and that one worked on the first try.
> 
> Johannes, does that sound like a probable cause? If so, it should be easy
> to fix.

I have no idea. Shouldn't that be easy to tell from the monitor? And if
the short/long switching was _unrelated_ in time it seems like it
wouldn't be a problem?

johannes

Attachment: signature.asc
Description: This is a digitally signed message part