Re: [quicwg/base-drafts] Handling of corrupt Retry packets (#3014)

[Jana Iyengar <[e-mail]>]

In PR #3120, @huitema wrote:

In Singapore, we were debating whether to just do checksum, versus encrypt and use an initial vector based on the CID. I wanted to compare the two variants using a benchmark designed for Picotls, but I ran into a little bit of trouble because of the structure of the picotls encryption API, which looks like:

1) Initialize an encryption context with the key context, the IV, and the authenticated data
2) Encrypt the message
3) Compute the AEAD checksum

With that API, there is no option to not use an IV,. Using a constant IV would have exactly the same cost as using a variable IV. I went on to compare the time to encrypt 64 bytes versus the time to just authenticate 64 bytes. Results are below for two implementations of AES128 GCM, one from OpenSSL crypto library and another from Windows Bcrypt library -- both measured using 64 bit code on my Windows laptop:


   | encrypt | decrypt | authenticate | verify
--  | -- | -- | -- | --
bcrypt  | 0.35751 | 0.22602 | 0.24687 | 0.20189
openssl 1.1.1c | 0.30211 | 0.23956 | 0.24296 | 0.23926

The times are measured here in microseconds -- these are the average on 100,000 operations. We see that the times are fractions of microseconds in both cases. The encryption operation is more expensive than a simple authentication, 45% more with brcrypt, 24% more with openssl. The decrypt operation is 12% more expensive than the simple authenticate with brcypt, almost the same with openssl.


-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/quicwg/base-drafts/issues/3014#issuecomment-561955416