Ping Unpredictability

This is an extension of the ping predictability metric developed by Hans-Werner Braun of NLANR. We define:

ping success = 1 - # 100 data byte ping packets lost / total number of 100 data byte ping packets

ping data rate = 2000 bytes / (average response time of 10 consecutive 1000 data byte ping packets)

Then for a set of pings (e.g. all the pings to a given host in one day) we calculate:

s = average(ping success) / max(ping success)

and

r = average(ping data rate) / max(ping data rate)

Then we define ping unpredictability (u) as the distance of the coordinate (s,r) from (1,1) or:

u = sqrt((1-r)**2 + (1-s)**2)

Finally in order to normalize the distance u to a maximum of 1, we derive the normalized unpredictability:

U = u / sqrt(2) (0 <= U <= 1)

As an example, the ping data for Rutherford Appleton Laboratory (RAL) for January 30, 1997 (see below for the response time and packet loss data) yields:

	Success %	Data Rate (kB/s)
Max	100	10.15
Avg	91	4.75
Ratios	s	r
	0.91	0.47

and this yields u=0.54, U=38.1%.

The Offsite Unpredictability table shows the normalized unpredictability (U) for sites seen from SLAC starting in Jan-95.

One can also group hosts with some affinity together and average the unpredicatbility for those hosts for some period. This enables us to get an overall indicator of performance for the group of hosts. An example where we have averaged the ping responses for all the Esnet hosts, Western N. American hosts, Eastern N. American hosts and International hosts is seen below: