Today's mesh testing.

[C. Scott Ananian]

On Sun, Mar 2, 2008 at 7:05 PM, Chris Ball <cjb at laptop.org> wrote:
>  broadcast, and they were all able to see and join a shared chat session
>  with each other.  The workload on the spectrum analyzer increased from
>  18% (no-one connected) to 26% (all connected).  The chat session is
>  consistent -- no-one is dropping out and new messages are seen by each
>  laptop, with a few seconds of lag.

I would interpret this as, "packets are being dropped like crazy, but
retransmits assure that the chat message gets through by the time a
few seconds elapse".  Is there any reason why chat would not be
instantaneous otherwise?

>  With the mass chat session still running, we shared a 500KiB PDF.  First
>  we joined the shared Read session with one laptop, and the download took
>  16 seconds to complete.  We then joined two more laptops at once, the
>  first download took 26 seconds and the second finished at 30 seconds.
>  Five more at once: all finished around 1m00s.  Ten more at once:  the
>  first finished at 2m18s, the last finished at 2m40s.  There were no
>  failures downloading the PDF.  The sharing was unicast TCP, with mesh
>  TTL set to 1, which explains the slightly worse than linear increase in
>  download time for more laptops downloading at once.

How did you get sharing to use unicast?  Was this a special patch, or
is this actually the default?  (My understanding is that sharing is
multicast.)

Just for reference, a 500kB file transmits over the "media lab 802.11"
network here in 1cc in just under 4 tenths of a second.  I tested and
it easily scaled up to 10 simultaneous downloads in 4-5 seconds.  So
somehow we're running 40-65 times slower than this on the mesh,
depending on the number of clients.

>  This is much more anecdotal than the full test plan, but we thought the
>  testers currently in Peru would want to know what they can expect from
>  the school server setup ASAP.  We don't have more laptops upgraded and
>  ready to join the network yet, but we don't have any reason to believe
>  we've saturated the network -- with the PDFs downloaded and Chat still
>  running, the duty cycle on the spectrum analyzer is now at 28%.  (In
>  general, wireless networks seem to start degrading around 40%.)

Based on my reading of the numbers above, it does seem like you're
pushing the network pretty hard.  But, as wad pointed out, if we can
do 32 clients on one channel of a school server, we can support the
100-laptop school scenario with three active antennas.
 --scott

-- 
                         ( http://cscott.net/ )