Re: [dnsoverhttp] [Ext] Fwd: New Version Notification for draft-hoffman-dns-over-https-00.txt

[Ben Schwartz <[e-mail]>]

I spent a few minutes playing with Martin's old bhpack code and got the
following results for HPACK Huffman-code queries.

HPACK(base64(random bytes)): 1.19 bytes per byte
HPACK(url-encode(random bytes)): 1.82 bytes per byte

HPACK(base64(random lower-case letters)): 1.15 bytes per letter
HPACK(lower-case letters): 0.82 bytes per letter

I compute that percent-encoding is about 50% larger than base64 for random
bytes in HTTP/2.  (I think it was 86% larger in HTTP/1.1).  For a mix of
random bytes and lower-case letters, percent-encoding beats base64 in
HTTP/2 when the input is 66% letters or more.  DNS queries have about 20
bytes in binary, so that would be about 40 name characters.

So base64 still wins, but I'm OK with percent encoding.

On Thu, May 4, 2017 at 8:16 PM, Martin Thomson <[e-mail]>
wrote:

> On 5 May 2017 at 01:46, Paul Hoffman <[e-mail]> wrote:
> > If we're going to use the wire protocol (which I think is good,
> particularly for extensions), then it should actually be the wire protocol.
> Therefore, I would strongly prefer "MUST be zero" to "be close to, but not
> exactly, the wire protocol but not exactly in order to help caching".
>
>
> This suggests a potential solution to Patrick's parameter negotiation
> dilemma.  Use those two bytes to switch query format...
>
> Also, zero encodes terribly in hashes.  Pick something nice.
>
> (I'll get back in my box now, the bad ideas should abate momentarily.)
>