[nsp] NxT1 MLPPP
warren at kumari.net
Thu Feb 12 10:38:09 EST 2004
Adding fuel to the fire...
While this was not on the 2600 / 3700 series it may be useful info for
We had a 6xE1 bundle between 2x7507s (RSP4, VIP2-50), full of VoIP
(64byte packets). The E1s in the bundle were really bad (on average
each one would flap 200-300 times per day!) and sometimes the bundle
would just stop forwarding traffic (but OSPF would stay up!). Sometimes
removing a random E1 (we would usually take out the last one that
flapped but it seems that any one would do) from the bundle, changing
the encapsulation to HDLC, back to PPP and then putting it back in the
bundle would fix the entire bundle.
I kept getting reports from our NOC saying that the performance though
the box was terrible but that the CPU was really low. Eventually I
followed up and discovered that they were only checking the RSP CPU and
not no the VIP (why, oh why doesn't "sho proc cpu show VIP cpu too?!),
when I checked the VIP CPU it was around 80-95% on average. Removing
MLPPP and using ECMP fixed all of the problems, the CPU on the VIP is
now like 20% (we also don't have the problem of the bundle hanging!)
If you do decide to go the MLPPP route, remember that the bandwidth of
the bundle changes if a circuit goes down. This means that if you are
load balancing across 2 bundles and one of them loses a link you will
likely send all of our traffic across the other bundle. We fixed this
by setting the OSPF cost manually (and / or the bandwidth). That way if
we lost a single link we would still use the bundle. There is some
command that will make the entire bundle as down if there are not at
least N circuits up in it, but it escapes me at the moment.
Hope this helps,
On Feb 11, 2004, at 2:27 PM, Beprojects.com wrote:
> Does anybody have any performance numbers on NxT1 connectivity using
> on Cisco routers? I am specifically looking for throughput on
> like the 2600 and 3700 (any flavors is fine). How many T1's can you
> and how much throughput can it handle doing IP only with virtually no
> features turned on (cpu usage would be good as well). Real world
> and/or lab
> tests are both fine. Thanks.
> cisco-nsp mailing list cisco-nsp at puck.nether.net
> archive at http://puck.nether.net/pipermail/cisco-nsp/
Militant Agnostic--I don't know and you don't either!
With Feudalism, it's your Count that votes.
More information about the cisco-nsp