[j-nsp] flow sampling: what packets are chosen?

[sthaug at nethelp.no]

> > Run-length appears to be unsupported on newer (e.g. MX3D MPC) hardware.
> 
> ISTR reading that's because they're 1:1 hardware sampling inside Trio
> instead of just punts to the LC CPU or RE CPU, at least for IPFIX on
> Trio.  Probably from the O'Reilly Juniper MX ... yup found it.  (any
> typos are mine as I'm retyping this)
> 
> "When using inline IPFIX the only valid rate is 1.  The option
> run-length isn't configurable, because there's no need to sample data
> from the perspective of the microcode in the Trio Lookup Block.  Every
> packet will be inspected and is subject to flow export."

We're doing inline IPFIX with 1:100 sampling rate, and getting sensible
data. This is on an MPC-3D-16XGE-SFPP line card. Config snippet:

sampling {
    instance {
        inline-netflow {
            input {
                rate 100;
                run-length 0;
            }
...

Steinar Haug, AS 2116