[c-nsp] bgp router

[bill fumerola]

On Fri, Jun 06, 2008 at 08:33:13AM +0200, Elmar K. Bins wrote:
> > My gut feeling is "go with a 7301 or 7200/NPE-G1".
> >
> > Why?  Because it can deliver the 200 Mbit/s bandwidth, and it's a 
> > "simple" architecture - everything is software, and there is lots less
> > hidden surprises than with the 6500/7600 platform.
> 
> That would depend on packet sizes. I know we're a bit extreme (most of
> our packets are around 64-128 Bytes), yet...we're hitting 50% CPU
> load on 7301s with like 60 Mbps of Traffic (in+out aggregated), which
> amounts to around 72kpps.

we experience the same. traffic is a little higher, but a large amount
of it is DNS packets, hence mostly <512 bytes.

> If your traffic consists of considerably larger packets, you may want
> to go with 7301s (G1) or 7201s (G2); if your packet sizes are small,
> you need to consider hardware forwarding platforms.

i know this may be heresy on this list, but look at juniper's J6350.
similar price to a c7301, more throughput (even at small packet sizes).

> Why is it, btw, that IOS doesn't use both CPU kernels there? Or did I miss
> an IOS version that started doing that? (still on 12.3T here)

i believe the 2nd CPU can only be enabled for some very specific features:
http://www.cisco.com/en/US/docs/routers/7300/install_and_upgrade/7301/7301_install_and_config_guide/5418c.html#wp1154543

%%%%%%
The Cisco 7301 includes a dual-CPU-core BCM 1250. All Cisco IOS images
for the Cisco 7301 platform use CPU-core 0. CPU-core 1 allows acceleration
of specific feature sets via separately purchased special software. As
of Cisco IOS Release 12.3(14)YM, multi-processor forwarding (MPF)
accelerates the following broadband features: L2TP Access Concentrator
(LAC), L2TP Network Server (LNS), and PPP Terminated Aggregation (PTA).
Port adapters are not supported in the multi-processor forwarding (MPF)
path on processor 1.
%%%%%%

wild-ass speculation follows:
i imagine the cost of data structure and code-path locking, IPIs and
other multi-processor primitives (or simply the fiscal cost of coding
same for this platform in 15+ year old code) negates any value to enabling
the 2nd CPU for code paths that run in interrupt context and/or run
through to delivery of the packet.  the aforementioned MPF features can
run independent of the IOS data structures that would need to be locked
if the entire IOS code ran in what we traditionally call SMP. they most
likely directly access the broadcom hardware over amd hypertransport,
hence the unavailability of port adapters for MPF.
</speculation>

there were murmurs of a team at cisco porting freebsd mips, which would
have given native SMP support. however, all the people who were supposedly
working on that no longer work for cisco (or now work in groups whose
bailiwick is clearly not core OS coding). read into that what you will.

-- bill