[Tinyos-devel] Re: from longer CTP/LQI experiments

[Omprakash Gnawali]

On Dec 22, 2007 9:29 AM, Philip Levis <pal at cs.stanford.edu> wrote:
>
>
> On Dec 21, 2007, at 8:28 PM, Omprakash Gnawali wrote:
>
> > I did several experiments yesterday - most of them longer runs than
> > the ones I have been doing the last few weeks. And one interesting
> > case - lqi30: LQI with maximum retransmission of 30. The question is,
> > if we let lqi retransmit as many times as ctp, what will lqi
> > performance look like? The answer seems to be still not good enough
> > and much worse than ctp.
> >
> > Here are the graphs:
> > http://enl.usc.edu/~om_p/net2/ctp4bitle/figures.html
> >
> > First I did four runs back to back, each run 4050 seconds: lqi30,
> > lqi5, a8, and a9. Based on these four experiments, lqi30 delivers more
> > pkts than lqi5 but at a much higher expense so a8 and a9 are looking
> > good.
> >
> > Then I did two long runs - a9 for more than nine hrs (labeled aa9) and
> > lqi5 for about 2.5 hrs (labeled lqi5b). Looking at the high thl graph,
> > I don't see it settling down permanently. It settles down for a bit
> > and starts climbing again. But maybe this is because the threshold of
> > 10 is too low for a pkt to classify as a high thl - 10 should be
> > enough!  The trend is more clear in the high goodcost distribution
> > graph. Again you can look at THL(t) and goodcost(t) to try to make
> > sense.
> >
>
> Given how well CTP is doing in terms of cost and churn, I'm not so
> concerned with the high THL. In effect, CTP is using data packets to
> probe the topology when it changes. If it just used control packets
> instead, we would not be worrying about it.
>
> The leveling off and then climbing is interesting: I wonder if it is
> due to a significant change in the environment, such as the stargate
> OLSR sending more traffic.
>
> It's crazy that LQI still does so badly with a max retx of 30. If you
> look at full vs. fail, the packets are still mostly being lost due to
> retx timeouts, rather than, say, queue overflows. So this really does
> point out the limitations of layer 1 information.
>
> It's heartening that aa9 is 0.997 delivery, given how rough Tutornet
> is. It will be good to see Rodrigo's results from Mirage, to see if
> alpha=9 performs as well as alpha=5 has done there in the past (0.99
> - 0.999). If so, we might have a winner.
>
> Phil
>

Here are the graphs that describe the outcome of the experiments on
Mirage. Rodrigo did experiments with a5,a6,a7 (two runs), and a9:
http://enl.usc.edu/~om_p/net2/ctp4bitle/mirage-12212007/figures.html

The usual detailed plots are accessible from the links at the bottom
of the page. The subdirectories for time series plots are named
as usual as well.

It seems like alpha does not change the metrics on Mirage: note the
narrow range on the y-axis for most of the metrics graphs. My conclusion
is they work equally well and for lack of a trend, if we try to draw more
conclusion, it might be reading too much into the data.

- om_p