[Freeipa-users] nsslapd-dbcachesize and database size

[Andrew E. Bruno]

On Wed, Oct 14, 2015 at 07:59:23AM -0600, Rich Megginson wrote:
> On 10/14/2015 07:09 AM, Andrew E. Bruno wrote:
> >The load average on our freeipa replicas started to spike over the
> >last few days and we narrowed it down to a dbcache issue. Following the
> >guidelines here: https://github.com/richm/scripts/wiki/dbmon.sh
> >
> >We saw that the dbcachefree was 2.0% which indicates a lot of page
> >churn. Sure enough our nsslapd-dbcachesize was set to 2G and the size of
> >our database and index files was 3.1G:
> >
> >$ du -sh /var/lib/dirsrv/slapd-[domain]/db/
> >3.1G
> >
> >Once we increased nsslapd-dbcachesize to 6G load average went back to
> >normal and query response times improved. Interestingly, when we
> >restarted the dirsrv process the database size went down to 1.7G
> >
> >$ du -sh /var/lib/dirsrv/slapd-[domain]/db/
> >1.7G
> >
> >When we initially deployed freeipa, the size of our database and indexes
> >was about 400M which is why we set nsslapd-dbcachesize to 2G.
> 
> What about your cachememsize?

We have nsslapd-cachememsize set at 2G for the cn=userRoot. According to
dbmon.sh it looked OK and we didn't think it needed to be increased:

dbcachefree 6123847680 free% 95.055 roevicts 0 hit% 99 pagein 34260 pageout 308661

       dbname      count          free  free%    size
changelog:ent         84        158342   30.9  4210.2
changelog:dn       29716            58    0.0    74.0
 userroot:ent       8446    2091021433   97.4  6685.1
 userroot:dn        8446     523272268   99.8   120.3
    ipaca:ent        100        673516   47.9  7338.6
    ipaca:dn         100       1399359   99.4    80.1


The changelog:dn seems to vary so much not sure how to tune that one. Any
suggestions? It's almost always 0% free.

> 
> >
> >A few questions:
> >
> >1. What causes the increase in size of
> >/var/lib/dirsrv/slapd-[domain]/db/*  and should we periodically clean up?
> 
> Replication metadata accounts for some of this.  Fragmentation accounts for
> some of this.  You can periodically clean up, but you shouldn't have to.
> The growth should eventually hit a plateau.
> 
> >
> >2. How do you tune nsslapd-dbcachesize to account for this growth? The
> >dbmon.sh wiki suggests a 12% overhead but our db files and indexes seem
> >to grow much larger?
> 
> 12% is sort of a starting point.  There isn't a good way to tell how to
> account for replication metadata, fragmentation, etc.  Just monitor
> periodically and adjust as needed.

Great, Thanks Rich!