httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From sameer <sam...@c2.net>
Subject Re: potential Apache performance problem: size of the TCP send buffer
Date Tue, 05 Nov 1996 06:36:58 GMT
> 
> I believe the OS default is fine, unless you are transferring large files
> across a high-bandwidth, low-latency link.  I.e. between universities on an ATM
> network.  For general Internet purposes, the default is probably fine.  The
> patch submitter, johnh@isi.edu, is not on the list, so you might want to
> confirm this with him, but this was at least my understanding.
> 

	Actually the problem is high bandwidth high latency links. If
you have two empty T1s connected to an unsaturated T3 backbone on
eithre sides of the US you will get 100ms latency but max throughput
under ideal conditions is still 1.5Mbps. Because of the 100ms latency
and 4k buffer size, you will only get 40kbps. You need to increase the
buffer size to get throughput which reaches the max available bandwidth.

	Stevens' Volume I has a good chapter on LFNs.

	I think The OS default isn't good for websites serving pages
worldwide to people on fast links. For websites serving pages to
people locally over fast links or worldwide over slow links, it is
fine. I am in favor of John's 16k buffer suggestion.

> 	Brian
> 
> On Tue, 5 Nov 1996, Rob Hartill wrote:
> > So is there a universally useful number we should be using to make sure
> > the OS isn't setting it too small ?
> > 
> > Brian Behlendorf wrote:
> > >On Mon, 4 Nov 1996, Rob Hartill wrote:
> > >> Ben Laurie wrote:
> > >> 
> > >> >Brian Behlendorf wrote:
> > >> >> This patch looks good to me - a patch against the current cvs
tree is appended.
> > >> >> Any other +1's?
> > >> >
> > >> >No doubt there'll be some portability issues, but +1 anyway.
> > >> 
> > >> Is the default set optimally now ?  This lets us change a value, but is
> > >> it set to a reasonable value by default ?
> > >
> > >The default doesn't change anything, the setsockopt which would effect the
> > >change doesn't get called.
> > >
> > >	Brian
> > 
> > 
> > 
> 
> --=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=--
> brian@organic.com  www.apache.org  hyperreal.com  http://www.organic.com/JOBS
> 


-- 
Sameer Parekh					Voice:   510-986-8770
C2Net						FAX:     510-986-8777
The Internet Privacy Provider
http://www.c2.net/				sameer@c2.net

Mime
View raw message