[Cryptography] Cycles overhead for TLS

Darkpassenger darkpassenger at unseen.is
Thu Sep 24 14:21:05 EDT 2015


On 2015-09-23 14:59, Ryan Carboni wrote:
> I'm not sure so these are rough guesses, but:
> 
> 5 cycles per byte for TCP overhead.
> 100,000 cycles for ECC key exchange.
> Several thousand cycles for PRF key generation.
> 10 to 1 cycles per byte for symmetric crypto
> Average webpage size is 1 megabyte, so maybe average TLS connection
> eventually transfers 15 megabytes.
> I think any performance improvements in asymmetric and symmetric
> cryptography would be minor compared to TCP overhead.

there are modern cpu instructions helping with crypto , to name one :
AES set of instructions in new intel chips . there are cheap crypto
accelerator modules used where performance matters and fast optical
networks dont care about tcp weight since mature lower level protocols
handle things properly for todays environment ( sonet , sdh on wire
and wcdma on air , for example ) . i havent seen engineers in
carrier-class operators i met ever worry about crypto weight
..comparing to the other common "issues" with performance in those
environment .


More information about the cryptography mailing list