[c-nsp] OC3 Throughput

[Euan Galloway]

On Mon, Nov 26, 2007 at 08:47:49AM -0500, Robert E. Seastrom wrote:

> My experience with both NPE-G1s and 7301s does not parallel this.
> Could be an artifact of the load we were running, the fact that we
> were running on 802.11 vlans on gigabit ethernet rather than ATM, or
> some kind of local configuration requirement that I haven't thought of
> today.  Wasn't running unicast RPF on the NPEs, preferring to run it
> on the edge-facing interfaces in the GSRs.  Was running dynamically
> loaded outbound SMTP filters via radius attribute 242.  Yes, I did
> have CEF turned on.  :-)
> 
> My experience is that once you were in the 120-130 Mbit range (sum of
> in and out), you were pretty much at the 80+% range and should be
> looking at a new router to add to the mix.  Of course, it probably
> depends even more on packets per second than it does on data moved.

Just to throw some random other numbers for people to find with google 
on day...

~220Mbit/s ~45kpps (sum in+out). <50% CPU as an LNS on 12.3 mainline.
Onboard GE to onboard GE (no subints).
~4000-4500 sessions over < 10 tunnels.
uRPF on every session / interface.
in/out ACL on every session / interface.
PBR every session - outbound SMTP (locally configured not from RADIUS)
Some classification/marking.

However - No fragmentation/reassembly (nor any MSS clamping requirement)
due to large MTU facing provider (bless BT for something).
frag/reassembly does destroy preformance quite nicely.

As you said, really all about the pps and features.

Similar to above, but with ATM in the LNS (G1), subints on the GE and 
some additional CPU expensive features, and the G1 hits more like 80% 
with single STM-1/OC3.

"Well done" or "zero features?" to the person getting 2* OC3 of L2TP 
tunnels terminated on one (or is that really "through one", tunnel 
switching rather than LNS?), and "I wonder what feature broke it" to 
your 130meg of GE-GE (unless significantly higher pps than above).

-- 
Euan Galloway