There are commonly defined 3 types of DiffServ tunnelling methods over MPLS:  the Pipe Model, Short Pipe Model and Uniform Model (QoS for IP MPLS Networks, Santiago Alvarez 2006).

The Pipe model ensure that there are no changes to the DSCP marking through the Label Switched Path (LSP) – even if a Provider router remarks the traffic (useful if connecting different DSCP domains).  The ingress LSR applies a Per-Hop Behavior (PHB marking), and applies to the pushed MPLS Labels (it may look at the IP Packets marking for this).  The egress LSR does not modify the tunnelled PHB marking.   Therefore the Pipe Model cannot use the Penultimate Hop Popping (PHP) of the label, but instead uses Ultimate Hop Popping (UHP).

The Short-Pipe model does use Penultimate Hop Popping (PHP), but infers the packet PHB from the tunnelled PHB marking in the IP Packet.  This could have challenges for supporting multiple customers with differing markings, hence the Pipe Model is more common.

The Uniform Model ensure that the DSCP domain of the original packet (ie. the Customer marking) is maintained throughout, so LSR maintain the PHB marking of the original packet when they perform a POP operation.  Any packet remarking is reflected in the packet marking when it leaves the LSR.  So the Penultimate Hop Popping performs it’s usual operation, and the Uniform method propagates the PHB marking to the exposed packet.

Tunnelling methods without PHP: