Re: Packet number encryption

[Mikkel Fahnøe Jørgensen <[e-mail]>]

Running the math with 0.1% PN encrypt overhead and an assumed 1200 byte
packet size

We get 66 cycles for PN encrypt (lower bound) and 66000 cycles for total
processing. AEAD is 1200 * 1,03 = 1236 cycles.
This means non-crypto processing is 66000 - 66 - 1236 = 64698 cycles.
1200/

On a 3GHz core this is 21566ns.
The Solarflare 98 ns/68 bytes result scaled to 1200 bytes is 1729 ns, which
suggests there is a theoretical 10x optimisation potential (which of course
is not realistic in a large scale production environment).

In packets per second this is 1e9/21566 packets/s = 46369pps
With 1200 bytes / packet this is 55,64 MiB/s.

This is roughly 1/2 Gbps which could be related to the network throughput
including loss and retransmission, context switches, and cache effects.

To compare with raw processing time, 55MiB/s is more than 10x slower than a
JSON parser that I have.


On 12 February 2018 at 08.19.07, Victor Vasiliev ([e-mail]) wrote:

 I looked at the profiles of various workloads we're running in production,
and I estimate that none of them would be impacted by PN encryption by
worse than 0.1%.