Subject: Re: vm tuning/monitoring
To: Simon Burge <simonb@wasabisystems.com>
From: SODA Noriyuki <soda@sra.co.jp>
List: tech-kern
Date: 09/10/2004 17:45:01
>>>>> On Fri, 10 Sep 2004 12:25:36 +1000,
	Simon Burge <simonb@wasabisystems.com> said:

> I added some evcnt counters and after 25 hours on my 512MB 2.0 branch
> i386 (with yamt's recent anonpages accounting fix) I see:
> 
> 	thoreau 10> vmstat -ev | grep uvmscan
> 	uvmscan anonunder                                 0        0 misc
> 	uvmscan fileunder                                 0        0 misc
> 	uvmscan execunder                              1745        0 misc
> 	uvmscan anonover                                457        0 misc
> 	uvmscan fileover                               1079        0 misc
> 	uvmscan execover                                  0        0 misc
>
> Is it worth adding this to our sources (maybe #ifdef UVMSCAN_COUNTERS)?
> The "uvmscan" might be better as "uvmpd_scan" too...

I'm not sure whether just counting those numbers is really useful or not.

For example,

- fileover is just fine. actually, causing fileover condition is most
  important reason for existence of vm.{anon,exec,file}{min,max}.
- not causing fileover condition may be fine too, as far as anonover
  and execover don't happen at the same time.
- causing execover also may be fine, as far as fileover happens at
  the same time.
- even causing anonover may be fine, if fileover and execover happen
  at the same time.

According to your result, there isn't any execover although there
is anonover, this probably means your vm.execmax is too large.

Anyway, current defaults of {exec,file}{min,max} are too large,
IMHO.
--
soda