Introduction
This is a study of the differences (if any) of the measurements from pinger.slac.stanford.edu (a single (bare hardware non-virtual) Linux host (Dell PowerEdge 2650) running the PingER Measurement Agent (MA) and pingervm (AKA dhcp-nebula-124-66.slac.stanford.edu a Nebula KVM Virtual Machine with floating addresses) MA measuring the ping Round Trip Times (RTTs) to hosts worldwide and between pinger and pingervm. Hencefiorth we refer to ping requests from pinger to pingervm as pinger to pingervm or pinger>pingervm and ping requests from pingervm to pinger as pinger to pingervm or pingervm>pinger. Both pinger and pingervm are located on the second floor of Building 50 (the computer center) at SLAC. The traceroutes between the two machines are seen here.
The measurements were made from February 26th 2015 and March 3rd 2015. Every ~30 minutes a burst of 10-30 (10 if no loss else up to 30 until 10 ping responses were received) ping requests was sent by the MA to the target. The ping packet size was 100Bytes, the ping requests in the burst were separated by 1 second. The number of pings between pinger and pingervm was ~6000.
Below we look at a statistical analysis of the results for pings from pinger and pingervm to the World, to N. America, Europe and between pinger and pingervm.
To World (excluding SLAC targets).
The following tables are from pingtable.pl
PingER non VM measurement agent
Tick | min | 25th% | avg | median | 75th% | 90th% | 95th% | max | iqr | std dev | # pairs |
---|---|---|---|---|---|---|---|---|---|---|---|
Mar2015 | 23.312 | 178.346 | 239.329 | 211.152 | 293.499 | 339.425 | 379.947 | 780.659 | 115.153 | 104.510 | 107 |
Feb2015 | 22.689 | 176.167 | 241.608 | 212.435 | 295.448 | 326.637 | 362.350 | 803.815 | 119.281 | 110.212 | 112 |
NB. February data is incomplete for the VM, so leave out
PingER VM measurement agent
Tick | min | 25th% | avg | median | 75th% | 90th% | 95th% | max | iqr | std dev | # pairs |
---|---|---|---|---|---|---|---|---|---|---|---|
Mar2015 | 23.698 | 176.221 | 238.352 | 213.278 | 294.279 | 340.938 | 381.412 | 796.140 | 118.058 | 109.834 | 107 |
Feb2015 | 22.515 | 178.790 | 239.368 | 210.327 | 301.033 | 329.510 | 369.066 | 807.798 | 122.243 | 110.401 | 110 |
To Europe
PingER non VM Measurement agent
Tick | min | 25th% | avg | median | 75th% | 90th% | 95th% | max | iqr | std dev | # pairs |
---|---|---|---|---|---|---|---|---|---|---|---|
Mar2015 | 150.996 | 162.944 | 175.858 | 173.632 | 186.939 | 201.843 | 204.698 | 204.698 | 23.995 | 17.131 | 14 |
PingER VM Measurement agent
To Europe
Tick | # pairs | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
Mar2015 | 151.017 | 162.965 | 176.121 | 173.904 | 185.921 | 199.620 | 208.584 | 208.584 | 22.956 | 17.586 | 14 |
To N. America
PingER non VM measurement agent
Tick | min | 25th% | avg | median | 75th% | 90th% | 95th% | max | iqr | std dev | # pairs |
---|---|---|---|---|---|---|---|---|---|---|---|
Mar2015 | 23.312 | . | 54.890 | 62.042 | 79.315 | 79.315 | 79.315 | 79.315 | . | 28.678 | 3 |
PingER VM measurement agent
Tick | min | 25th% | avg | median | 75th% | 90th% | 95th% | max | iqr | std dev | # pairs |
---|---|---|---|---|---|---|---|---|---|---|---|
Mar2015 | 23.095 | 41.357 | 42.888 | 62.077 | 79.350 | 79.350 | 79.350 | 61.775 | 4 |
Correlation plots of pinger vs pingervm for hourly measurements
Below are correlation plots of hourly PingER measurements between pinger.slac.stanford.edu and pingervm.slac.stanford.edu between Feb 26 and March 3rd, 2015.
Time series
If one compares the average and median statistics for the tow sets of data, one gets the tables below:
pinger>pingervm | min_RTT(ms) | avg_RTT(ms) | max_RTT(ms) |
Average | 0.304368 | 0.385107 | 0.50461 |
standard deviation | 0.045313 | 0.054924 | 0.272844 |
median | 0.297 | 0.385 | 0.471 |
IqR | 0.07625 | 0.0635 | 0.08075 |
Count | 272 | 272 | 272 |
pingervm>pinger | min_RTT(ms) | avg_RTT(ms) | max_RTT(ms) |
Average | 0.245923 | 0.310827 | 0.392096 |
stdev | 0.040275 | 0.03718 | 0.052863 |
Median | 0.237 | 0.315 | 0.386 |
IQR | 0.06425 | 0.04925 | 0.077 |
Count | 272 | 272 | 272 |
And plots of the time series appears as below (spreadsheet)
Looking at the above manually scaled plots of pinger>pingervm and pingervm>pinger it is apparent the RTTs from pinger to pingervm are > pingervm to pinger. Looking at the average, standard deviation, median and IQR tables and the differences between pinger as the monitoring site and pingervm as the monitoring site, we see in tabular form (where the average errors are +- (stdev(pinger>pingervm)+stdev(pingervm>pinger)) and the median errors are +- (IQR(pinger>pingervm) + IQR(pingervm>pinger)). the Probabilities are those that the pinger>pingervm and pingervm>pinger distributions are the same (assuming normal distributions).
Diff (pinger-pingervm) | min_RTT(ms) | +- | Probability | avg_RTT(ms) | +- | Probability | max_RTT(ms) | +- | Probability |
Average | 0.058445 | 0.085588 | 0.682689 | 0.074279 | 0.092104 | 0.682689 | 0.112515 | 0.325707 | -0.99405 |
stdev | 0.005037 | 0.017744 | 0.21998 | ||||||
median | 0.06 | 0.1405 | 0.07 | 0.11275 | 0.085 | 0.15775 | |||
IqR | 0.012 | 0.01425 | 0.00375 |
It appears that the min_rtts, avg_rtts and max_rtts are within 1 standard deviation (better than 68% assuming a normal distribution) of one another.
By looking at the cumulative distributions we can get the Median differences probability.
Minimum RTT distributions between pinger and pingervm
We focus on the minimum RTT since that is least affected by queuing etc.