[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Swamped! How to tune OBSD for massive web traffic?
* Robert Collins <firstname.lastname@example.org> [010120 13:58]:
> From: "Henning Brauer" <email@example.com>
> > On Fri, Jan 19, 2001 at 06:44:46PM -0800, Seth Arnold wrote:
> > > It might also help (depending on how your users use your site) to set
> > > the KeepAliveTimeout number higher as well; one minute seems like the
> > > highest that number should be pushed.
> > KeepAlive INCREASES Serverload as more httpds are running. Without KeepAlive
> > you can shoot nearly every router down with heavy used severs due to the
> > pure mass of tcp connections, so you'll have to find a good compromise. 15
> > seconds is normally a good value.
> I thought kept alive connections were less load on the os not more?
> Certainly that's what the http/1.1 performance tests indicate (they
> show up to 3 times the perfomance on a lan with http/1.1 over 1.0
> mainly due to pipelined requests on persistent connections.
There are differences between performance as a web browser will see it,
and performance as the server will see it, and there are differences
between the 'load average' and 'how much work is being done'. :)
I suggested increasing the KeepAliveTimeout to help avoid the three-way
handshake when clients connect for additional graphics. I figured that
if he was having trouble serving the content, that it would take more
than fifteen seconds or so for all the images to travel, and some images
might not load at all, etc.
If the end user doesn't do anything with the connection for a little
while, the server process doesn't do anything either, and just sits
there, taking up some space in the load average. If the end user does
use the connection, then increasing the limit has saved a (fairly
time-consuming) three-way handshake.
I suppose what should be learned from this thread is that
experimentation is the only way to find out for certain if tweaking a
particular setting will help -- and the note that load average is *not*
the end-all indicator of how well a server is doing should be heeded as
``Oh Lord; Ooh you are so big; So absolutely huge; Gosh we're all
really impressed down here, I can tell you.''