httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Marc Slemko <ma...@worldgate.com>
Subject Re: process model again...
Date Wed, 08 Apr 1998 07:32:28 GMT
On Wed, 8 Apr 1998, Dean Gaudet wrote:

> > Even
> > if you remove minspareservers, you will still be yoyoing unless
> > you crank max high.  No matter how you do the pool, Max-Min has to
> > be > your average medium term connection change or you get killing
> > and starting.  I think this is just your server being very
> > freaky.  I don't see this to the same extent. 
> 
> There has to be a better algorithm.  I don't like people having to tweak
> these numbers.

Weighted averages over time, so when it starts up it has a poor idea, but
after running for a bit and seeing the load swings, it figures things out
with a goal of not having to start servers x%, not having to kill x%, etc.

> 
> I'm thinking of something that actually is based on the rate of
> connections, and the average service time.  Anyone study queuing theory? 
> I avoided that unfortunately. 

Yea, you could somewhat model this with queuing theory (and in some rough
calculations I did last year on a server here, the arrivals did fit into
something that could be modelled almost perfectly using a simle process) 
but all you get out of it is that you will need to have x running to be
able to serve requests without delay x% of the time.  I think that for
practical purposes you can get something better with a algorithm in the
server that takes a best guess based on the history; yea, that is stats
but for that stuff I find it normally is just common sense coding as well. 

Whatever you do, you will not get rid of issues when you have jumps from
30 to 2 in use.  There is no way around having too many servers hanging
around or having to kill servers when you have a... 1500% load change.
Once you get to a few hundred connections, it smooths out a lot better so
you probably only have 50% changes or less.


Mime
View raw message