[c-nsp] Cisco 8000

[adamv0025 at netconsultings.com]

> Saku Ytti
> Sent: Friday, December 20, 2019 8:59 AM
> 
> 
> So I fail to see vendor/merchant different here, I continue to see run-to-
> completion and pipeline difference here. And indeed merchant (particularly
> Jericho) does have flexibility now Cisco and Juniper equivalent pipelines
do
> not have, simply because they get requirements from a wider audience.
> 
Hey Saku, 
I see where you're coming from and I do agree with the underlying premise, 
I guess it's hard to argue the case though as majority of products marked
"vendor silicon inside" are run to completion NPUs vs majority of products
labelled/marketed as "merchant silicon inside" are pipeline NPUs, so people
tend to generalize on the vendor vs merchant rather than run to completion
vs pipeline. 

For all the engineers, 
When in doubt whether you're dealing with pipeline or run to completion NPU,
Stuff the NPU with pps all the way till it starts dropping packets then dial
back a little so you can get packets through without any loss,
Now start adding features like QOS, Filters, IPFIX, etc... 
If the NPU won't drop any packets that means it's pipeline, if it starts to
drop more and more the more features you enable it's run to completion.

And obviously the list of features or combinations of features that you can
enable will be lot less on pipeline NPU -but each packet is guaranteed to
leave the pipeline precisely at its clocking speed, compared to run to
completion NPU where you can enable just about any combination of features
you can think of, but the NPU will take its time to process each packet
-meaning it might not have time for some newly arriving packets.


adam