Tuning Question

[Cy Schubert]

In message <44odnziewr.fsf at be-well.ilk.org>, Lowell Gilbert writes:
> Cy Schubert <Cy.Schubert at komquats.com> writes:
> 
> > I have a question about tuning FreeBSD systems, specifically in regard to 
> > memory. At one time on Solaris systems it was recommended to keep scan rate
>  
> > below 200 pages per second. (Today it's 300 pages per second, dependent on 
> > the amount of memory, class of system.) Are there any recommendations or 
> > rules of thumb a person can use to determine when a memory upgrade is 
> > required? My machines are only busy during port builds when the scan rate 
> > can vary greatly and the page out rate could reach as high as two pages per
>  
> > second during brief periods. What kind of memory and paging metrics should 
> > I use on FreeBSD systems?
> 
> I'm not convinced that such a simple algorithm makes sense these days.
> If the system is normally pretty quiet, then it is unlikely you'll see
> any difference from optimizing memory behaviour further.  
> 
> I believe that when top(1) gives memory sizes, they are in bytes
> rather than pages as the manual indicates.

Top's output, as is free memory on all O/S's these days, is bogus. It's the 
size of the free memory pool which is available for immediate allocation. 
Used memory is just as  useless. It doesn't matter how much is swapped out, 
what matters is how much I/O is being performed to support VM. I know at 
work, which is an Oracle ghetto, paging should be kept at a minimum, 
especially the SGA. Other apps can afford more. In the case of an average 
FreeBSD system it's been guesswork.

I should also mention that even though Sun recommended certain metrics for 
their systems. In a previous life as an MVS (IBM mainframe) systems 
programmer, IBM recommended that no more than 5% of system resources should 
be used for paging, otherwise you're paging too much. I suppose this might 
be a good paper for someone to write. (As I'm already writing a book, I 
probably don't have time to research and write yet another subject.)


-- 
Cheers,
Cy Schubert <Cy.Schubert at komquats.com>
FreeBSD UNIX:  <cy at FreeBSD.org>   Web:  http://www.FreeBSD.org

			e**(i*pi)+1=0