W3C home > Mailing lists > Public > ietf-http-wg@w3.org > July to September 2012

Re: FYI... Binary Optimized Header Encoding for SPDY

From: Adrien W. de Croy <adrien@qbik.com>
Date: Thu, 02 Aug 2012 05:21:53 +0000
To: "Yoav Nir" <ynir@checkpoint.com>, "Mike Belshe" <mike@belshe.com>
Cc: "James M Snell" <jasnell@gmail.com>, "ietf-http-wg@w3.org" <ietf-http-wg@w3.org>
Message-Id: <emc1f2400c-3d0f-48eb-8eea-cb91edab99b3@bombed>

need to be able to handle / transport private/unknown/extension headers.

I'd propose just flagging the header as such, and making the payload 
contain the header name followed by a single 00 byte then the value.  
Since header names have to be ASCII, this should be no problem.

Also, since this gives you an option to transport ANY header.

Also, length of size fields could be a frame attribute, rather than 
just 32 bits per-header.  So, if the frame wants to, it could say 
lengths are in 16 bit words, rather than 32, or whatever.  If you have 
a header that won't fit, you'd mark it in the frame.  This would give 
an upgrade path for lengths without the cost now.

I don't really like using 32 bit for TV headers though.  That won't 
work for dates.  I'd prefer a byte for value TYPE, e.g. 

00 = 32 bit integer
01 = 64 bit integer
02 = datetime (possibly just 64 bit ns since epoch)
03 = text
04 = binary/opaque
05 - 7F extension types (length prefixed value)

You'd potentially use this instead of differentiating between TV and 
TLV headers, since anything with variable length (text or binary) would 
have a length field in there as well, fixed-length types wouldn't need 
it.

New types could be overloaded within binary, and the length would take 
care of it.

Or could just use the bit for whether there's a length field before 
payload.  But I think we need to cater for more than just 32bit integer 
when it comes to fixed-length value types (if only date/time).

Actually date time would benefit from future-proofing as well, since ns 
may be great resolution now, but hopeless in 10yrs.  Which would 
require either an extension type, or variable length type for datetime 
encoding (e.g. maybe just for fractional second part).


Adrien



------ Original Message ------
From: "Yoav Nir" <ynir@checkpoint.com>
To: "Mike Belshe" <mike@belshe.com>
Cc: "James M Snell" <jasnell@gmail.com>;"ietf-http-wg@w3.org" 
<ietf-http-wg@w3.org>
Sent: 2/08/2012 4:25:01 p.m.
Subject: Re: FYI... Binary Optimized Header Encoding for SPDY
>On Aug 1, 2012, at 8:30 PM, Mike Belshe wrote:
>
>
>>
>>A couple of thoughts:
>>
>>* Thanks for writing up!
>>
>>* I don't think we need utf-8 encoded headers.  Not sure how you'd pass them off to HTTP anyway?
>>
>>* The codepages seem like complexity, but I'm not sure the benefit.  I would remove them.
>>
>>* I would remove the flags too - per header flags - do we really need it?  I'd remove it without a very clear use case.
>>
>>* I know that 32bits seems like a lot.  Defining length fields has two routes:  fixed length or variable length.  I like the fixed length because I believe they are simpler.  However, the price of that simplicity is that you've got limits.  Everyone hates limits :-)  In your proposal you whacked the number of headers to 8 bits, or 256 headers.   While I agree this is an edge, I don't see a reason why it should be against the rules to have more.  Same for the length of a header value - you've used 16 bits (64KB).  While this seems massive by today's standards, in 10 years maybe 1MB cookies are the norm.  I don't know, but I'd hate to have the limit.  So.... this leaves us thinking that maybe we should use variable length encoding.  Personally, I think the fixed length simplicity is worth it.  But this is subjective, of course.   Just use 32bits everywhere - it works well and you won't notice the perf difference at all (I measured :-)
>>
>
>
>Hi
>
>I agree. I haven't written a draft like James, but I would like to suggest an alternative encoding, that I think has the advantage of being easier to parse:
>
>Every header begins with a 48-bit descriptor (I think 32-bit would be enough with a 16-bit length/value, but this is safer)
>
> L0cRiiiiiiiiiiiiVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVVV
>or
> L1cRiiiiiiiiiiiillllllllllllllllllllllllllllllll
>
>The first bit is the "Last" bit - 1 for the last header and 0 for all the previous ones
>The second bit is 0 for TV headers (type-value) and 1 for TLV headers (type-length-value)
>The third bit is 1 for critical (do not parse the request if you don't understand this) or zero for not
>The fourth bit is reserved.
>
>The next 12 bits are the type of header. There are actually 2^13 possible headers, because the same type can be used for both a TV and a TLV header.
>
>For TV, the next 32 bits hold the actual value - a number from 0 to 4,294,967,295 - should be enough for lots of headers, including those denoting time
>For TLV, the next 32 bits hold the length, and are followed by the value.
>
>Using your registries from section 1.4 as an example, we get this:
>
>20 01 00 02 00 00 (version=2.0)
>20 02 00 00 00 01 (method=GET)
>60 04 00 00 00 01 2F (path="/")
>C0 03 00 00 00 0F 77 77 77 2E 65 78 61 6d 70 6C 65 2E 63 6F 6D (host="www.example.com")
>
>That's 40 bytes. The HTTP/1.1 version is 39 bytes:
>
>GET / HTTP/1.1
>Host: www.example.com
>
>Of course, you save more on headers that have longer names, like "strict-transport-security" or "Original-Encoded-Information-Types" or "Downgraded-Disposition-Notification-To"
>You also save a lot on those that can be TV, and the value doesn't have to be numeric. It could be a bit string of whatever, as long as it fits in 32 bits.
>
>Yoav
>
>
>
>
Received on Thursday, 2 August 2012 05:22:20 GMT

This archive was generated by hypermail 2.2.0+W3C-0.50 : Thursday, 2 August 2012 05:22:31 GMT