[j-nsp] Cost of vMX

[adamv0025 at netconsultings.com]

> Mark Tinka
> Sent: Sunday, April 22, 2018 11:03 AM
> 
> I posit that a number of cloud environments are either DIY (meaning
unlikely
> to forward serious traffic), or were installed by a vendor (meaning they
were
> pricey, and as such, went in with limited forwarding capacity).
> 
> To do stuff in CPU and memory should not be an issue - but I doubt that
> many cloud environments have been setup to replicate what a router/switch
> forwarding 100Gbps or more can do.
> 
> Of course, I could be wrong. 

You're right, but for different reasons.

Regarding the cloud deployments,
That really depends on how may VMs you expect to run on a single compute
Host.
Then the number and nature of the VMs (BW hungry or PPS hungry) dictates how
much traffic you may expect at each particular host.
And there are also physical constrains in terms of the number and type of
Host ports (max 4? but usually 2, speed: 10, 25 or 40gbps). 
And all this drives the forwarding needs for the router VNF on the host. 

And yes the important part is -this is not a dedicated router VNF use case,
so there have to be enough cycles left for the VMs to actually
generate/process the egress/ingress traffic. 
This is why some try minimizing impact on host's resources by offloading the
CP onto the SDN controller and just program the forwarding plane sitting on
host "remotely".
While this may help with RAM utilization (depending on the scale), it won't
help much with CPU requirements as bulk of the router VNF work is done in
data-plane anyways.

I recon 99% of the hosts out there have 2x10GE NICs, which any modern host
with any vendor VNF can do without breaking any sweat.

How I see it is that router VNF (PE) on host gives me the best flexibility
-i.e. allows me to realize the BW hungry ports (DC fabric) on dumb devices
with *rudimentary switching capabilities while all the complexity is pushed
onto the low bandwidth edge (in other words because on host its low pps/BW
it's actually cheap to perform complex operations on each packet).     
* I consider MPLS as rudimentary switching capability.  

adam

netconsultings.com
::carrier-class solutions for the telecommunications industry::