170594 - pipelining batching should be more flexible

Status: RESOLVED FIXED

(Core :: Networking: HTTP, defect, P2)

Description

[Judson Valeski]

I'm bugifying a conversation I just had w/ Darin... please correct summary or
description as necessary.

Our current pipelining implementation buffers up multiple requests into a single
stream, then hands that stream to the connection layer which shuttles it to the
server in a single request. That's all well and good, however, it's not granular
enough. If the first two (of four) requests are written, and two more (total of
six) have been created in the meantime, we won't get the new two into the
"batch" until the full initial "batch" has been sent.

it sounds like the concept of a "batch" needs to be broken apart such that as
requests get sent, new requests can be added to the "batch" in the pipeline.

Comment 1

[Darin Fisher]

difficult to say if this is something i can do in the 1.2 beta timeframe given
the likelihood of introducing regressions/crashes.  targeting 1.3 alpha for now.

Comment 2

[Michael Buckland]

Bulk adding topembed keyword.  Gecko/embedding needed.

Comment 3

[Chris Nalls]

Marking topembed+ as per topembed triage.

Comment 4

[benc]

-> http

Comment 5

[Darin Fisher]

my patch for bug 176919 includes the fix for this bug.  latest results indicate
that pipelining Tp to cowtools over a high-bandwidth LAN connection jumps from a
3% improvement to a 7% improvement.  what remains is to test it over DSL and
modem connections.

the algorithm i've implemented is as follows:

1- build up an initial pipeline (4 max requests) and send them in one chunk. 
this is usually written out as two TCP segments depending on the MSS for the
connection.

2- when the first response completes, send another request.

3- repeat (2).

if N responses are read from the socket at a time (read size is 4k), then we
will queue up at most N more requests before writing to the socket again.  as a
result we get a continuously overlapped stream of requests/responses, which i
think explains the increased performance.

Comment 6

[Darin Fisher]

FIXED with patch from bug 176919.