Re: AEAD and header encryption overhead in QUIC

[Matt Joras <[e-mail]>]

Wow, really interesting results, thanks for sharing!

I was curious, since it's not mentioned on the PRs that I saw, were these
tests done with any adjustments to the ACK frequency on the client? Or is
this with the default ACK policy of ACKing every other?

Matt Joras

On Thu, Jun 18, 2020 at 1:24 AM Kazuho Oku <[e-mail]> wrote:

> Hello folks,
>
> TL;DR: When existing crypto libraries are used, the AEAD cost of QUIC can
> become somewhat greater than that of TLS over TCP, due to those crypto
> libraries not optimized for handling short AEAD blocks. Crypto libraries
> optimized for short AEAD blocks can tighten the gap. Also, crypto libraries
> optimized for QUIC can combine AEAD operation and header protection vector
> generation, effectively eliminating the cost of header protection.
>
> A bit longer story:
>
> Recently, when testing quicly [1] using a NIC with *hardware support* for
> UDP GSO offload, we noticed that it uses +30% CPU cycles compared to TLS
> over TCP. This was due to our underlying crypto not being optimized for
> short AEAD blocks, and due to the call overhead of doing header protection
> one packet at a time. [2]
>
> We went on to implement our own AES-GCM library optimized for short AEAD
> blocks (i.e. packet-sized), that also generates the header protection
> vector while doing AEAD computation. By switching to the new crypto
> library, the performance drawback of the crypto itself compared to large
> AEAD blocks when down from 30% to 10% [3].
>
> The overhead of header protection (relative to all the crypto cost) on the
> send side is now between 0.3% to 1% when using packets sized around 1400
> bytes to 1500 bytes. The ratio against the total transport cost is going to
> be less than that, and in practice, I think that they are negligible. The
> story might be a bit different on the receive-side due to the dependency
> between header protection and AEAD, though I am sure there would also be
> ways to optimize.
>
> In terms of CPU cycles spent by the QUIC stack, the overhead compared to
> TLS-over-TCP came down to 5%. quicly can now sustain the send rate of 5Gbps
> while spending about 30% of CPU cycles of one modern Intel CPU core [2].
>
> These numbers are observations from one TLS / QUIC stack doing
> micro-benchmarks, though I tend to think that sharing performance numbers
> lead to people having better expectations regarding the CPU cost of QUIC.
> Some of us were also concerned about the cost of header protection (a.k.a.
> packet number encryption) [4]. Hence posting here.
>
> [1] https://github.com/h2o/quicly
> [2] https://github.com/h2o/quicly/pull/359
> [3] https://github.com/h2o/picotls/pull/310
> [4]
> https://mailarchive.ietf.org/arch/msg/quic/Wy8JXwKyeYgWgRWizWDbyOLg0xc/
>
> --
> Kazuho Oku
>