Re: pipelining and chunking

[Randy Terbush <ra...@zyzzyva.com>]

>From what I am hearing of 1.2, I would support any performance improvement
if we can show that it does not break things.


> So Henrik and I have concluded that the problem is in the libwww library
> and not in Apache.  He's working on that and will re-do the benchmarks
> when he's got it fixed.
> 
> In the meanwhile, I hope to get a chance later this week to use the libwww
> robot to test chunking.  I'd like to implement the performance tweak I was
> talking about, is there any chance of it getting into 1.2?  Here's the
> outline:
> 
> - add char *chunk_header to the buff structure.  If non-null then this points
>   to the first byte of a chunk header in the buffer.  This is set by the
>   switch to chunking, and at that time enough space is reserved for the
>   header (you know how big it'll be at that time).  Since the chunk trailer
>   requires two bytes I'll also decrease bufsiz by two at this time (or
>   implement some other kludge).
> 
> - when flushing the buffer, twiddle chunk_header if required, tack on the
>   chunk trailer, flush away.  Start off a fresh buffer with a new
>   chunk_header, and outcnt set appropriately.
> 
> This essentially eliminates all the bflush()s that we have left.  It's
> not as elegant as layered i/o, but it should do the job.
> 
> Dean