Re: Measure the latency for a specific throughput/datarate method

["Jeff W. Boote" <boote --at-- internet2.edu>]

Tiago Junqueira wrote:
>   Hi Jeff,
>
>  Thanks for the telling me about the program thrulay. I'll test it
> today. About the other question about the jitter, can you help me?
> Iperf uses the following formula to calculate the instantaneous
> Jitter:
> #        // TODO functionalize this
> #        // from RFC 1889, Real Time Protocol (RTP)
> #        // J = J + ( | D(i-1,i) | - J ) / 16
> #        transit = mPacketTime.subSec( sentTime );
> #        if ( lastTransit != 0.0 ) {
> #            deltaTransit = transit - lastTransit;
> #            if ( deltaTransit < 0.0 ) {
> #                deltaTransit = -deltaTransit;
> #            }
> #            mJitter += (1.0/16.0) * (deltaTransit - mJitter);
> #        }
> #     lastTransit = transit;
>
>  But it seems that the last samples have more weight than the earlier
> ones, so if the Jitter before the last is high like:
> J(Last)=J(Last-1)+(|D(i-1,i)|-J(Last-1))/16, the J(Last) will be high.
> I was making mistake while taking the J(Last) as the overall jitter,
> but it can't be in this way. Do you know how Iperf makes the average
> of all the Jitter[i]?

Iperf uses the RFC 1889 version of jitter which does in fact bias later packets. 
The RFC 1889 version is really more intended for instantaneous statistics during 
something like a video conference. The performance you can expect is more biased 
toward the most recent measurements - so for this application it makes sense.

For owamp*, I have chosen to provide a measure of delay variation instead (which 
many people call jitter). I use the difference of the 95th percentile of delay 
and the median delay. There is a draft going through the IETF/IPPM group that is 
advocating using inter-quartile delay for this. (difference of 75th percentile 
and 25th percentile) This seems like a very reasonable choice to me.

But, as to the answer to your question: Iperf does not return an average of all 
Jitter[i]. (And an average of a running average is not likely to make a lot of 
sense, is it?)

jeff

* http://e2epi.internet2.edu/owamp