W3C home > Mailing lists > Public > ietf-http-wg-old@w3.org > January to April 2001

Re: Proposal (I-D) for extending HTTP to support out-of-order responses

From: Patrick McManus <mcmanus@appliedtheory.com>
Date: Thu, 12 Apr 2001 16:03:22 -0400
To: Jeffrey Mogul <mogul@pa.dec.com>
Cc: http-wg@hplb.hpl.hp.com
Message-ID: <20010412160322.A12494@AppliedTheory.com>

> I'm just looking for comments on whether the protocol makes sense
> (and, in particular, whether it might lead to screwups by caching
> proxies that don't comply with the HTTP specifications).

2 thoughts, the first is just a throw-away:

1 - HTTP is stateless. indeed 2616 in its intro describes itself this
    way.. this is a pretty fundamental change, even as an optional
    extension.. so the intro should probably discuss it. that's all.

2 - This isn't a safety/correctness thing, but it seems like its a bit of
  a problem to effective usage..

given: 1] non-safe methods like POST or PUT are more likely  to be
          barrier-requests than GETs
 
       2] POSTs and PUTs can have large request bodies with
       significant upload latencies

       3] large request bodies benefit from the expect/100-continue
       mechanism

       4] real networks often have symmetric up and down capactiy that
       its a big win to leverage by overlapping.

I'm thinking about the expect/continue mechanism and how the execution
of a resource by a big POST or PUT is likely to be a sequence point
(and thus a barrier request). the time spent uploading the data would
do well to be overlapped with any outstanding responses even if it
can't be executed until outstanding requests have been processed.. but
if we mark it a sequence point we can't do that because the server
can't say 100-continue until all of the outstanding requests have been
satisfied.. it'd be nice to be able to separate the out-of-order
properties of the 100 with that of the final response so that a "start
sending" signal could be given OOO but the resource itself not
executed OOO. Obviously this requires some server side buffering, and
if it's not willing to do that it could simply do the 100 in-order.

the obvious way to do this seems to be to associate an attribute with
100-continue.. say ERID that works similarly to RID, but applies only
to the 100 response, not the final one.

hopefully the following case played out 3 ways will help illustrate:

assume we've got a POST(d) that takes 6 units to send, and 5 responses
that take 4(a), 5(b.cgi), 2(c), and 2(d), 3(e) to send. throw in the
kicker that b.cgi takes 5 units of time to calculate its response and
d takes 2 units to process the big post.

first, the we'll do the easy case where d is not a request-barrier.


Time    Upstream			    Downstream
----	---------			    -----------
1        GET /a, RID: 1       
2        GET /b.cgi, RID: 2                 200 - RID : 1 (1/4)
3        GET /c, RID: 3			    200 - RID : 1 (2/4)
4        POST /d, RID:4, Expect:..	    200 - RID : 1 (3/4)
5					    200 - RID : 1 (4/4)
6					    100 - RID : 4 (1/1)
7	POST/d, RID: 4 body (1/6)	    200 - RID : 3 (1/2)
8	POST/d, RID: 4 body (2/6)	    200 - RID : 3 (2/2)
9	POST/d, RID: 4 body (3/6)	    200 - RID : 2 (1/5)
10	POST/d, RID: 4 body (4/6)	    200 - RID : 2 (2/5)
11	POST/d, RID: 4 body (5/6)	    200 - RID : 2 (3/5)
12	POST/d, RID: 4 body (6/6)	    200 - RID : 2 (4/5)
13	GET /e, RID: 5			    200 - RID : 2 (5/5)
14					    200 - RID : 5 (1/3)
15					    200 - RID : 5 (2/3)
16					    200 - RID : 5 (3/3)
17					    200 - RID : 4 (1/2)
18					    200 - RID : 4 (2/2)

18 units.. pretty cool.. 

but now, what if /d (the post) is a request barrier under mogul-00?

Time    Upstream			    Downstream
----	---------			    -----------
1        GET /a, RID: 1       
2        GET /b.cgi, RID: 2                 200 - RID : 1 (1/4)
3        GET /c, RID: 3			    200 - RID : 1 (2/4)
4        POST /d,  Expect:..		    200 - RID : 1 (3/4)
5					    200 - RID : 1 (4/4)
6					    200 - RID : 3 (1/2)
7					    200 - RID : 3 (2/2)
8					    200 - RID : 2 (1/5)
9					    200 - RID : 2 (2/5)
10					    200 - RID : 2 (3/5)
11					    200 - RID : 2 (4/5)
12					    200 - RID : 2 (5/5)
13					    100 continue
14       POST/d, body (1/6)
15       POST/d, body (2/6)
16       POST/d, body (3/6)
17       POST/d, body (4/6)
18       POST/d, body (5/6)
19       POST/d, body (6/6)
20	 GET /e, RID: 5			    [wait - processing d]
21		 			    [wait - processing d]
22					    200   (1/2)	 
23					    200   (2/2)	 
24					    200 - RID : 5 (1/3)
25					    200 - RID : 5 (1/3)
26					    200 - RID : 5 (1/3)

26 units.. not as cool.. but what if the 100-continue could be out of
order, without making the final response out of order?

Time    Upstream			    Downstream
----	---------			    -----------
1        GET /a RID: 1       
2        GET /b.cgi, RID: 2                 200 - RID : 1 (1/4)
3        GET /c, RID: 3			    200 - RID : 1 (2/4)
4        POST /d, Expect: 100-c..; ERID=E1  200 - RID : 1 (3/4)
5					    200 - RID : 1 (4/4)
6					    100 - ERID : E1 (1/1)
7	POST/d, RID: 4 body (1/6)	    200 - RID : 3 (1/2)
8	POST/d, RID: 4 body (2/6)	    200 - RID : 3 (2/2)
9	POST/d, RID: 4 body (3/6)	    200 - RID : 2 (1/5)
10	POST/d, RID: 4 body (4/6)	    200 - RID : 2 (2/5)
11	POST/d, RID: 4 body (5/6)	    200 - RID : 2 (3/5)
12	POST/d, RID: 4 body (6/6)	    200 - RID : 2 (4/5)
13	GET /e, RID: 5			    200 - RID : 2 (5/5)
14					    [ wait - processing d]
15					    200 (1/2)
16					    200 (1/2)
17					    200 - RID : 5 (1/3)
18					    200 - RID : 5 (2/3)
19					    200 - RID : 5 (3/3)

19 units!

-Patrick
Received on Thursday, 12 April 2001 21:07:24 EDT

This archive was generated by hypermail pre-2.1.9 : Wednesday, 24 September 2003 06:33:42 EDT