[c-nsp] MPLS QoS

[Johannes Resch]

On Tue, June 23, 2009 05:12, Shine Joseph wrote:
> Hi,
>
> Can anyone with SP experience state why would one choose pipe model over
> short pipe model or vice versa for deploying MPLS QoS?

With full pipe you get some nice benefits like the ability to classify all
user traffic only based on MPLS EXP (since traffic will always be
transported labelled only), and giving your end customers the option to
transport their customer-set DSCP value transparently through your
network.
(you can fully decouple your core QoS handling from customer-set DSCP
values).

However, there are also a few issues here:

Full pipe model requires ultimate hop popping, which has a performance
penalty on many platforms. (e.g., on 7600 with SUP720/RSP720 it forces a
recirculation of the traffic on the egress PE, effectively halving the PPS
rate of the box). On that platform, supposedly EARL8 will have the HW
capabilities of doing a label pop operation plus IP lookup at the same
time without recirculation.

Also, be sure to test HW platforms you are using well before making this
decision. We found quite a few nasty bugs related to UHP on multiple
vendors' products in the process, which lead me to think that UHP is not
really commonly used (yet).

One example - due to a HW bug on 6500/7600 3B/3BXL cards, full pipe mode
does not work at all on any 6500/7600 chassis if it is in 3B/3BXL mode (if
activated regardless, it leads to wrong imposed EXP values for traffic
IP->MPLS). This has been fixed in 3C/3CXL HW.


HTH,
-jr