W3C home > Mailing lists > Public > ietf-http-wg-old@w3.org > May to August 1995

Re: Content-Transfer-Encoding "packet"

From: Roy Fielding <fielding@beach.w3.org>
Date: Mon, 24 Jul 1995 15:28:36 -0400
Message-Id: <199507241928.PAA00843@beach.w3.org>
To: Dave Kristol <dmk@allegra.att.com>
Cc: http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
>Could you elaborate?  I'd be curious to know the tradeoffs you examined
>before choosing this approach.  In particular, was the overhead of an
>ASCII packet length (i.e., human readable) so onorous?

An ASCII decimal format is, yes.  Human readable was not an issue
in this case.
 
The advantages that the one-byte packet-size has is that it takes up
a minimal amount of space in the stream of bits to the transfer and
is trivial for any system to produce or consume.  Allowing larger
packets means we have to use decimal (with the additional CRLF delimiters)
or hope that everyone remembers to read the number in network byte order.
 
*NOTE*  We need to keep in mind that the CTE packet size has little
        to do with the TCP packet size.  There is absolutely nothing
        preventing an application from writing up to X CTE packets
        per network write, where
 
                      Usable Window
                  X = -------------
                          256
 
It does have an impact on the receiving application's read buffer
size, but that is usually not an issue these days (most systems are
capable of reading data from the TCP buffers much faster than it gets
filled, so the application's reads are not in the critical path).

The BIG question is what effect it will have on the number of
internal system calls and data copies required by the server or 
the client.  That I do not know, and Jim Gettis (the W3C person who
is about to start looking into HTTP/2.x) mentioned earlier today
that they may overwhelm any gain from simplicity.  Of course, this
is also dependent on the average size of dynamic data transfers
(i.e., if the vast majority of transfers is "small", than a small
packet size is preferable).

===================================================================
So, here is a task for the WG:

   If you have a simulator handy (or just some free time on your
   hands), please attempt to analyze the behavior of a representative
   sample of dynamic responses, with the idea being a comparison
   between the simple "256packet" implementation of my last message
   <199507230403.AAA10044@beach.w3.org> and the following
   "decimal packet" CTE:

   Entity-Body = *( packet ) 
                 "0" CRLF
                 footer
                 CRLF
 
   packet      = packet-size packet-data
 
   packet-size = ("1"|"2"|"3"|"4"|"5"|"6"|"7"|"8"|"9") *DIGIT
 
   packet-data = packet-size(OCTET) CRLF
 
   footer      = *( Entity-Header )

and report your results back to the mailing list.  Keep in mind
that the goal is consensus and submittal of a proposed standard RFC
on HTTP/1.1 by September 21.
===================================================================

>1) If I understand correctly, there is always at least a blank line
>following packetized data (for a null footer Entity-Header).  True?

That depends on where you consider the beginning of the line to be.
If you consider it to be after the zero (NUL), then yes, since the
256packet scheme would end in
   <last-packet> NUL CRLF
if there is no footer.

In the decimal packet scheme there would indeed always be an empty line,
since it would end in
   <last-packet> 0 CRLF CRLF
if there is no footer.

>2) Does the footer applies only for C-T-E "packet", or would it
>apply to other (to be specified) C-T-E's?

Only to CTE packet, unless we defined more new CTEs (unlikely,
except for the possibility of a packet64 CTE for 7bit transfer).

>3) What happens if (yes, it would be bizarre) one of the footers is
>Content-Length?  (It looks like it gets overwritten with the one that
>the packetizer computes.)

It gets overwritten (in fact, it *must* be overwritten).

>4) Will support for packetization be a required part of HTTP/1.1?

Yes.

>5) How will acceptable packet sizes be negotiated (or specified)?

Well, if we have small packets, they won't be.  If we have decimal
packets, then that is another matter for discussion.  Does it matter?


 ....Roy T. Fielding  Department of ICS, University of California, Irvine USA
                      Visiting Scholar, MIT/LCS + World-Wide Web Consortium
                      (fielding@w3.org)                (fielding@ics.uci.edu)
Received on Monday, 24 July 1995 12:30:40 EDT

This archive was generated by hypermail pre-2.1.9 : Wednesday, 24 September 2003 06:31:23 EDT