[Battlemesh] The 'blowing up the network' results
toke at toke.dk
Wed Aug 12 14:29:28 CEST 2015
So I managed to find the USB stick with the results from the 'blowing up
the network' data sets. Below is a description of the test run, and a
link to the results.
We ran five different tests on the test setup (same setup as the other
tests; all credit for setting up the test bed goes to the testing team,
I just piggy-backed on their efforts): Two variants of the RRUL test,
TCP upload and download tests, and a 8-stream download test designed to
mimic the dslreports.net speedtest.
The RRUL test consists of running four simultaneous bulk TCP streams in
both the upstream and downstream directions (so eight streams total),
while simultaneously running UDP and ICMP latency measurements. The test
runs for 60 seconds, to make sure the network is fully saturated. There
are two variants of the test: A best-effort only (called RRUL_BE) in
which no diffserv marking is employed. And the straight RRUL test, which
marks each of the four TCP streams with different diffserv markings
matching the mapping into hardware queues in the WiFi stack.
The TCP upload and download tests are just a single TCP stream with
simultaneous latency measurement, while the 8-stream download test runs
(as the name implies) eight simultaneous download streams while also
The results of the test runs are available here:
That directory contains the Flent data files for all of the test runs,
as well as CDF and box plots of the two variants of the RRUL test. As
you can see, performance is fairly abysmal, with latency going up into
the several seconds range. As I said on Saturday, the point here is not
to pick a winning protocol (we only have the one sample, which is wayyy
too little data to conclude anything meaningfully), but to show that
they all break down in such conditions.
For those interested in exploring the data further, I'd suggest
installing Flent (https://flent.org) and loading the data files
themselves into the GUI. There are several other plots available for the
RRUL tests, and of course the other tests have plots available as well.
Oh, and as a final note, I'd recommend adding these (or similarly
disruptive) tests to next year's test regimen. Coming up with a test
that actually causes sufficient disruption that the routing protocols
start dropping links would be particularly interesting, in my
(Did I mention that I like blowing up networks? ;))
Feel free to ask any questions about the results or the tests themselves.
More information about the Battlemesh