Hi! On Tue, 11 Aug 2009, Philip Levis wrote:
> > On Aug 7, 2009, at 1:27 PM, Razvan Musaloiu-E. wrote: > >> Hi! >> >> About 2 months ago, a commit [1] to CpmModelC.nc that eliminates the >> problem of receiving packets (non-zero PRR) for negative SNR also made a >> significant change to the hardcoded PRR/SNR curve. This was not adequately >> documented in the commit message and many people might not be aware of >> this. Here is a plot that shows the differences: >> http://bit.ly/znrwQ I made a mistake in the above graph. In Yin's data the SNR is actually SINR. After accounting for this the new graph looks like this: http://farm3.static.flickr.com/2456/3812438521_c08b3ec98e_o.png So the new curve is actually in agreement with Yin's data. :-) >> The points represent the data collected by Yin Chen, a member of our lab. >> In his tests the noise level was controlled by using a noise generator so >> it covers a wide range. Similar data was also reported by the following >> paper: >> http://portal.acm.org/citation.cfm?id=1460427 >> The conversion of SINR to SNR should also make this data in agreement with the new curve. >> So there are reasons to believe that even the current curve we have in the >> tree is not the most accurate one. Caution is advised! :-) Considering that multiple datasets collected in very different ways are in agreement now I think there are good reasons to believe the new curve is a good one. >> [1] >> http://hinrg.cs.jhu.edu/git/?p=tinyos-2.x.git;a=commit;h=afe96d7e2a3747bba450e3db88a89569a0ba53b6 > > The methodology Yin used to collect the data set in the above figure is not > appropriate for calculating SNR/PRR curves. I think it is. Here is a visual representation of the signal/noise space explored by Yin's data: http://farm4.static.flickr.com/3548/3812438643_bf272d9b56_o.png Only 12 receiving motes were used but multiple TX levels were used. Here are two other views: SNR vs Signal http://farm3.static.flickr.com/2448/3813253680_f72943d1a3_o.png SNR vs Noise http://farm3.static.flickr.com/2623/3813253658_12aa32c68a_o.png > There are three separate issues, all of which have a common concern: > averaging. While the data sets in the above paper are suitable for its > conclusions, they have problems when it comes to simulation. > > 1) It combines many node pairs: each pair can have a different SNR/PRR curve. > This is why an SNR of 3dB can have a PRR of 95% or 5%. Here is the SNR vs PRR for each node: http://farm3.static.flickr.com/2607/3812438555_0199f0591d_o.png It seems that the overall shape of the curve stays the same and is only shifted to the left or right. We should be able to simulate this using an appropriate random variable. More experiments are needed to clarify this. > 2) Signal is a not a controlled variable, and it is averaged. A lack of > control means that it will vary due to environmental conditions, and the fact > that it is averaged over received packets leads to sampling bias. It is not > clear in the paper if the plot averages the RSSI of packets along a link or > averages the per-packet SNR. Since N is considered to be constant (more on > that below), the two are equivalent. But if the signal strength has > variation, and you observe the SNR only of received packets, you can observe > an SNR higher than the PRR would suggest. > > 3) Noise/interference are not controlled. Taking the average of N assumes > that it is a gaussian variable; while N is, external interference I can > disrupt this measurement. This might explain why the SNR/PRR curve is lower. > E.g., if 10% of my packets have a high interference, my averaged N will go up > significantly in a way that will lead the SNR curve to lead to incorrect > conclusions. Yin has some plots that shows the variations of the signal and noise in his data. I agree that both are legitimate concerns. > There are ways to give a sense of the degree of these methodological > limitations. The spread of discrete points in the Figure is a good approach > for 1). A plot of the S distribution could help with 2). A plot of the N > distribution could help with 3). > > The current curve in TOSSIM was generated from a data set that Kannan > collected using a variable attenuator, shielded cabling, and 2 micaZ motes. > I've attached a plot of the associated data. The error bar along the X-axis > is the standard deviation of signal strength. Using a variable attenuator and > shielded cables leads these values to be very stable. The red bar shows the > noise floor, which he calculated as the mode of RSSI readings taken when > there were no transmissions. Since these were measured in a closed system, > they are not affected by external interference. I did not know how was the data collected by Kannan. From your description I understand that the noise was always constant (-96 dBm). Is this correct? It also seems that in the experiment tested only a very small number of distinct PRR values (I can only see only 4-5). It could be possible to test a few more? More points in the -90 and -93 dBm would paint a much clearer picture. > The prior curve was generated with an inferior data set, taken from an > uncontrolled environment (such as Yin's). I think the experiments did by Yin are complementary to the one did by Kannan because they explore the SNR when the noise is different from the noise floor. I think this is very important because it's not so obvious that the SNR vs PRR curve is the same for various noise levels (which is exactly what CPM is relying on :-)). To reiterate what I said somewhere in the middle of this reply: the (at least visual) agreement between three very different experiments (Kannan, the Sensys 2008 paper and Yin) is an excellent indication that the new curve is a good one. :D All the best! Razvan ME _______________________________________________ Tinyos-help mailing list [email protected] https://www.millennium.berkeley.edu/cgi-bin/mailman/listinfo/tinyos-help
