httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dean Gaudet <dgau...@arctic.org>
Subject Re: roy's l_c perf patch and spareservers
Date Sat, 15 Feb 1997 07:22:18 GMT
My machines regularly show 20Mb free RAM even during peak loads.
(I'm using IRIX's osview command to get that figure.)

I suspect that the extra number of children blows the L2 cache so that
you spin too many cycles causing the cache to repop.  In my case these
aren't super-high-end SGIs, they're just IP22 Challenge Ss, and I think
the L2 is only 1Mb (real ram is 128Mb).  I'm pulling that figure
out of my butt based on what the same generation of Sparcs have.

I'm not convinced that's the only reason.  As I mentioned, I saw a change
from 50% to 20% cpu just by losing 20 spare children on average.

I should go stick some child creation/death logging code in. 

Dean

P.S. IdleOptimizationThought: It's too bad we can't better control who
gets to go into accept() 'cause we'd get the best locality using LIFO.
Starvation is good in this case.

On Sat, 15 Feb 1997, Marc Slemko wrote:

> On Sat, 15 Feb 1997, Chuck Murcko wrote:
> 
> > excessive forking and lousy response on load spikes; making Max too high
> > results in wasted machine resources. It's just solving a limit problem
> 
> Did you notice making max too high hurting anything if you have enough RAM
> to keep all the child processes in memory?  ie. is it just the extra
> footprint of the extra child or is there actually some interaction going
> on between the children that gets worse as the numbers increase?
> 
> 


Mime
View raw message