Re: concurrent requests?

Henrik Nordstrom wrote:

> Adrian Chadd wrote:
>
> > Right, what clients actually do this? I guess if its there it should
> > be implemented properly. :-P .. Is there a chance of someone actually
> > tidying this up or can I just disable it totally for the time being ?
>
> Internet Explorer sometimes does pipelining I think.
>
> Other than this it is mostly batch tools.
>
> pipelining certainly deserves to be implemented properly. It is very
> important for HTTP performance over high latency links. What is properly
> in this case I don't know. Note that blindly doing pipelining from a
> proxy is not the proper way here. Actually, what Squid is doing is quite
> close to being the proper way but it is or at least have been quite
> error prone internally and somewhat sub-optimal..

I agree that pipelining is probably A Good Thing(tm), but in order to reach
it's peak performance you really also want to do chunked-encoding and chunk
interleaving. That more or less means having a connection mux/demux layer,
between the core app and the wire.

> One quite related thing which is certainly not done in the proper way is
> Squid's retrial of failed requests. Squid is way to eager to retry
> failed requests while much of this responsibility should be put on the
> originating user agent. Also, it is a bit careless when selecting a
> connection to send a request on in the first place (over optimized
> selection)..

Concur.

D
Received on Wed May 31 2000 - 19:11:38 MDT