WiFi performance and quirks

Hal Murray hmurray at megapathdsl.net
Thu Mar 26 06:43:13 EDT 2009

wmb at laptop.org said:
> I can easily explain the 1.4 MBytes/sec figure - that's exactly what
> you'll get if you send maximum-size Ethernet frames at a rate of one
> frame per millisecond.  The 1/ms rate limitation is easily explained
> by the USB macro-frame time. 

Thanks.  That matches exactly what I'm seeing.

1.4 megabytes/sec is 11.2 megabits/sec.  If the packet format is ethernet 
(and I counted right) the per-packet overhead would drop the throughput down 
to 10.9 megabits.  My tests get up to 10.8.  That's close enough for me.

> I did a lot of work to make the NANDblaster send as fast as possible,
> using the mesh in a restricted mode to avoid retransmissions.  I got
> to 2.8 Mbytes/sec.  Ricardo Carrano told me that the fastest rate
> that has ever been measured using Linux is 1.4 MBytes/sec.

2.8 megabytes/sec is 22.4 megabits/sec.  I found one web blurb that counts 
all the overhead.  It said that the best 802.11g can get is 22.  I didn't 
check anything, but it seemed reasonable and gave enough info that I probably 
could check things.
The killer is a long preamble to be backward compatible with 802.11b.


Thinking that if one XO is limited to 11 megabits, I did the obvious thing of 
running 2 TCP streams in parallel, one to each XO.

My Linksys AP seems to crap out at slightly over 7 megabits/sec.  That's what 
I get with one TCP stream from PC->AP->XO.  If I run 2 tests to separate XOs, 
I get roughly half the bandwidth on each.

Sending in the XO->AP->PC direction gets "interesting".  One TCP stream gets 
10 megabits.  Two streams get under 1 megabit/sec each.  If I chop the socket 
buffers down to 4K, things work much better.  They get about 5 megabits each. 
 What a rathole.  I think I'll stop digging myself in deeper.


Does anybody know of a handy database of throughput for USB WiFi dongles 
and/or APs?  I poked around a bit but didn't find anything.

These are my opinions, not necessarily my employer's.  I hate spam.

More information about the Devel mailing list