[e2e] Re: crippled Internet
kc at ipn.caida.org
Fri Apr 27 16:28:25 PDT 2001
Looking at the NLANR data (http://watt.nlanr.net/active/) we see
measurements such as "average RTT of 60 ms plus standard deviation of 5
ms", one may be tempted to conclude that life is good -- average one way
delay plus two std sums to 40 ms, much less than 150. However, there are
two caveats. First, a significant portion of the 150 ms budget must be
reserved for the processing function -- maybe half of it in the current
state of the art; this leaves a target of 75 ms for the network delay.
Second, measuring the std is a very poor approximation of the behavior
of jitter compensation algorithms. We need a much finer grain
understanding of the structure of the delay.
Stanislav mentions that "Surveyor shows 0th, 50th, and 90th percentile
graphs by default." It is a good start, but there are two issues. First,
surveyor uses the IPPM defined methodology, sending probes at
pseudo-random intervals; this is not a good modeling of the VoIP
traffic, which mostly consists of "talk spurts", trains of packets sent
at a 20 ms or 30 ms interval; the interesting measurement is the maximum
cumulative jitter within a packet train. Second, even if we accepted
random sampling, we would have to check for the 99th percentile, rather
than the 90th percentile. We could only derive this percentile from the
surveyor data if we new the structure of the delay distribution: if the
distribution is Poisson, then the 99th percentile will be about 2.4 time
farther from the mean than the 90th percentile; if it is a power law
with a 1.5 exponent, it will be 6.5 times farther.
Can we get better data?
can you (or someone) be more specific?
More information about the end2end-interest