Re: Log level and performances

[Daniel van Eeden <daniel@snow.nl>]

Most syslog implementations allow logfiles to be written eighter in sync
or async mode. What does your's do?

Daniel van Eeden <daniel_e@dds.nl>

On Fri, 2004-04-16 at 10:52, BAILLEUX Benoît FTRD/DMI/CAE wrote:
> Two questiosn about OL performances when logging is activated.
> Someone asks me recently if logging has a high effect on OpenLDAP performances. As I have no information to answer him, I play with an OL instance already up, and I have a strange result.
> The platform :
> Dual processor PC, with Linux MDK 9.1, two HDD, 384 Mo RAM
> OpenLDAP 2.1.5, BDB 4.2.50 (data files and log file on separate disks)
> Directory loaded with two millions tiny entries, based on a very specific schema.
> Excerpt of the slapd.conf file :
> 
> # -----------
> allow bind_v2
> threads         4
> reverse-lookup  off
> loglevel        <various - see below>
> 
> database        bdb
> suffix          ...
> rootdn          ...
> rootpw          ...
> 
> cachesize       500000
> idlcachesize    20000
> searchstack     8
> # -----------
> 
> I use DirectoryMark (DM) to play just the same scenarios three times : one time with no log at all, one time with logelevel=256 (directed to syslog) and one time with loglevel=0 but the "-d 256" option on command line, with redirection to a file.
> The DM scenario is very simple : 95% exact search, 5% modify. 5 client process with 2 threads. Measures are made after a warming up.
> 
> The results are (tests repeated 3 times) :
>  - no log       : 2540 op/s (cool !)
>  - loglevel=256 : 29   op/s (uh !)
>  - slapd -d 256 : 2577 op/s
> 
> So I have two questions : 
>  - is there a problem with my configuration that makes slapd so slow when it logs to syslog ?
>  - why is the throughput slightly better (and not clearly worst) when OL logs to stdout ?