I have not implemented such a setup, but I am just wondering why something specially would need to be taken care when using random-detect with MPLS. Can someone throw light on this?
I'm basically referencing that no other headers are looked at when doing label switching. Therefore, when applying random-detect to a queue we can not easily have TCP seperated from all other traffic to maximize our chances of dropping a TCP packet to slow down the source. It appears that the commonly accepted method is to set the mark probability denominator high enough and make sure that the majority of your traffic is TCP.