apr-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From r..@apache.org
Subject Re: [PATCH] speed up network timeout processing
Date Thu, 04 Jul 2002 17:25:56 GMT
On Wed, 3 Jul 2002, Ian Holsman wrote:

> Bill Stoddard wrote:
> >>>From: trawick@rdu88-250-182.nc.rr.com [mailto:trawick@rdu88-250-
> 
> <snip>
> 
> > I'd rather pay the maintenance overhead on this code. BTW, testing on AIX
> > yields up to a 9% increase in throughput with this patch (serving 500 byte
> > file out of mod_mem_cache). Huge win for a small change!
> > 
> > Bill
> > 
> 
> 
> fantastic speed improvement that rocks.
> 
> like Brian said the APR is to provide a consistent usable interface to 
> the outside world. if we choose to implement the timeout using the 
> poll() function instead of a select() function than that is perfectly
> valid.

If that was all we were doing, I would agree with you.  But `Jeff's patch
implements BOTH select and poll with an #ifdef, because not every platform
has poll().  This is exactly the reason for having apr_poll(), and not
using it is stupid.  If the argument is performance, then back it up with
numbers.  If the performance is as bad as everybody seems to think it is,
then fix the functions.  If the problem is the API itself, then fix the
API.

> another case in which we do the SAME THING is how we implement 
> sendfile() on machines which don't have it. we don't use apr_file_writev 
> we use another standard system call writev

It isn't the same thing.  First of all, according to my reading of the
code, we don't implement sendfile at all if the platform doesn't suppoort
it.  Apache itself does, but it uses sendv.  Second of all, as I said
above I wouldn't have a problem with the approach if it didn't require
#ifdef's to implement two options to do the same code, which is the exact
problem APR is supposed to solve.

Even worse, is that we are making decisions without ANY numbers to back up
what people aare saying.  I am planning to do some measurements on Friday
to look at the REAL numbers before we can make a decision.  If the numbers
are much worse for apr_poll, then I have some thoughts on how to improve
that API to fix the performance problems.

Ryan

_______________________________________________________________________________
Ryan Bloom                        	rbb@apache.org
550 Jean St
Oakland CA 94610
-------------------------------------------------------------------------------


Mime
View raw message