[atlas] Questions about paris traceroute

On Jul 17, 2014, at 6:49 PM, Randy Bush <randy at psg.com> wrote:

>>> how close an approximation?  chasing ecmp and lag is bad enough without
>>> other changes.  folk need reproducible results.
>> 
>> As far as I know, the actual hashing algorithms used by load balancers
>> are in general not known.
>> 
>> So the only thing paris-traceroute can do is to try to provide
>> consistent results for each variation and try to create enough
>> difference between variations that it likely that all paths are tried.
>> 
>> In this context, it doesn't really matter if you try different
>> variations in one measurement or run different measurements. In both
>> cases some bits will be different. But it not possible to predict in
>> advance what is going to happen.
>> 
>> Worth pointing out in this context (and it is also documented somewhere)
>> is that if a probe reboots, the bits it uses for paris-traceroute are
>> likely to be different. So consistent results with a single measurement
>> are only achieved for as long as a probe doesn't reboot.
> 
> a multi-id measurement needs to be closely clusetered in time or changes
> in routing or other network events can create ugly and unreproducible
> (in the someone else tries the same experiment) results.
> 
> don't know about others, but i don't really care about load balancers.
> it's ecmp and lag on big pipes.

With ecmp and lag, I hope that a routeur's hash stays consistent as long as the number of interfaces stays constant. The fact that a probe reboot may lead to a change in flow id is an issue.

Cristel