How to implement Jitter?

[severindellsperger]

Hi @thombashi

For an application, I need to implement a jitter. In the case that the mean delay is enough large we can work with the delay-distro command. For example:
tcset eth0 --overwrite --delay 10ms --delay-distro 2ms
The result comes quite close to the selected delay-distro value which in this case represents the jitter value.
We face a problem if the mean delay is smaller than the jitter-value. In an example like the following the delay-distro approach isn't working successfully:
tcset eth0 --overwrite --delay 1ms --delay-distro 5ms
I understand that the problem is originated because the delay value can't be negative and therefore the delay-distro action can't take place correctly (+/- delay-distro value can't be applied).

Do you have an idea of how to solve the problem?
Especially the problem if the mean_delay is greater than the jitter value is a tough one...

Kind
severindellsperger

[severindellsperger]

I think the problem could be solved if another parameter is introduced to set the distribution of the latency. The problem is related because the delay distribution is set to normal distribution by default (which is not always the best solution). If I like to set the jitter to an exact value a uniform distribution would make more sense.

[thombashi]

Hi @severindellsperger,

As you pointed out, the current implementation of tcconfig only supports latency that normally distributed.
I will consider adding an option to set delay distribution other than normal for future release.

[thombashi]

You can now use --delay-distribution option with tcconfig 0.26.0.
Using --delay-distribution, You can use distributions {normal,pareto,paretonormal}.
However, uniform distribution can not be used, caused by tc. tc command does not accept uniform distribution even though tc-netem man describe that distribution.

https://man7.org/linux/man-pages/man8/tc-netem.8.html

[severindellsperger]

Awesome! Thanks for your work. I took a look at the implementation of tc and the source code of the Linux kernel. It looks like the network latency is normally distributed per default. However, the normal distribution would be possible if there exists some normally distributed correlation data on the localhost (see here for more information). I wrote my own implementation, which had worked with the normal distribution, and just raised an exception if the needed correlation table was missing. Maybe you can add this feature in a future release (if someone needs it).