Re: Performance tuning

[Joe Pluta <joepluta@xxxxxxxxxxxxxxxxx>]

This is good stuff, Paul, I really appreciate it.  You're kind of 
reinforcing what Diego said about pool allocation, and also touching on 
my fear that maybe I'm looking to do that because I "think" it helps, 
but that maybe it doesn't anymore.  And of course the only way to really 
know is to do the analysis, make changes and see if it helps.  And if 
you do something egregious, dealing with the inevitable (and justified) 
negative feedback from the user community.

Faulting and paging.  That's what led me to segregate Java workload back 
in the day... ah, good times.


On 11/30/2018 1:49 PM, Musselman, Paul wrote:
> We've been using AutoTune from Help Systems for decades.  As we use it, it's pretty much set-it-and-forget-it.  Of course, at this time we're blessed with an abundance of hardware horsepower, so tuning is pretty much academic.
>
> Our main requirement is to keep interactive (including web) response time complaint-free.
>
> In general terms, we try to give interactive/web-based tasks enough memory and CPU to keep them happy.  And they seem to use whatever you give them!
>
> We try to keep interactive / web / batch / spool applications separated, more from habit than from analysis.  But it works for us.  Spooling, as we learned early on, needs more memory than the early recommendations.  The goal is to let spool get in, print something, and get out quickly.  Normally, this is a relatively small hit on the CPU, but it does take memory to hold a bitmapped page with overlay!
>
> Don't forget system functions!  They need their own chunk of memory as well.  Don't shortchange the system!
>
> Faulting and paging rates.  The most mystical part of tuning!  Old guidance was that pages are not as bad as faults-- it just means the system requested something not in main memory (ie randomly accessing a record in a file (*sigh* Row in a table)).  Faults mean the system's look-ahead paging algorithm missed, and something needed (like the next lines of the program, or the next record in a file being read sequentially) weren't in memory.  Faulting should be kept to a minimum; Pages low enough that the users don't complain about response time!
>
> Paul E Musselman
> PaulMmn@xxxxxxxxxxxxx