Interesting observations in the data

These are the plots shown by Daniel Karrenberg, with some notes added.

Plot 1

Introduction of the plot format. Nice nominal delay pattern. "As we like it."

Plot 2

Another nice one, but with (working)day patterns.

Plot 3

Generally nice, but some losses visible on Jan 15th. This is good to be able to look up in case customers complain.

Plot 4

Zoom in on Jan 15th of the previous relation. Note also the routing change between 1500 and 1600 with the corresponding delay change. This demonstrates that you can look at interesing bits using "plots on demand". Now you can tell when exactly the problem happened.

Plot 5

Now *really* zooming in on the detail around the packet loss period. It appears this hit out-of-the-blue with no warning. Another noteworthy detail is that the bin size of the histograms is roughly one minute. On the bottom left one can clearly see the stochastic nature of the traffic generation. In moste one-minute intervals we send 1 or two packets but there are minutes with 3 and with 0.

Plot 6

Another measurement. This direction looks good...

Plot 7

... but the reverse direction is really bad. ...

Plot 8

With plots on demand you can change the scale to see the whole picture. Note how the loss patterns follow the delays. Also look at the quite pronounced distribution graph that shows that the composite is really the sum of three distributions.

Plot 9

Zoom in on a particular day of the previous measurement. I cannot really explain this shape with the minimum delay being as reqular and the destribution of delays staying so narrow during the cycle. Any ideas welcome.

Plot 10

Lots of different paths. I assume some terrestrial and some sattellite. On Jan 19th it became really lossy too.

Plot 11

On the first glance this looks OK but the distribution has two peaks...

Plot 12

Zooming in shows that again there is an overlay of two distributions. The baseline difference of 5ms is remarkable. I expect this to limit TCP throughput. Using the traceroute database this could be traced to a particular ISP. According to them this is due to loadsharing across multiple physical links. We are still trying to explain the large (5ms) baseline difference in the measurements.

Plot 13

This is another one with barely visible double peak distribution.

Plot 14

Zoom in on the previous one. This shows a "square wave" change pattern with an amplitude of ~200us. This measures traffic across two load shared 2Mbit/s circuits between the RIPE NCC office and the Amsterdam Internet Exchange. The geographic dstance is approximately 4km. One line is engineered with 4 baseband modems, two of them back-to-back. The other line is engineered through an SDH transmission system. Comparing the delay data with the traceroute database confirms that the SDH line shows the 200us higher delay. The point of this is not to compare baseband transmission with SDH, there is more to that than delay. The point is that we can reliably measure these eeffects.

Page created by Henk Uijterwaal, Last Update: $Date: 2001/01/31 09:20:08 $.