Sar understanding

Malaclypse the Elder dwc at homxc.UUCP
Wed Aug 10 06:23:29 AEST 1988


In article <446 at axis.fr>, loic at axis.fr (Loic Dachary) writes:
> 
> 
>      sar -d : . Is there any ceil values for these data ? Specially
>                 %busy, r+w/s, avwait. I know for example that runq-sz
>                 should not be greater than 2, and %rcache lower than 90%.
>               . Can blk/s be recalculated (approximately) with other values ?
>                 I tried to sum bread and bwrite without great success 
> 		(bread + bwrite <= blks/s / 2, could that mean bread is
> 		1024 byte blocks and blks/s 512 byte blocks ?)

you have to also add in the i/o from swapping and/or page stealing.
and a ceil for %busy is 100% :-)  the r+w/s is determined by the rate
at which the disk can do i/os.  avwait has no ceiling as anyone who
has worked on a busy system knows.

>      sar -q : . swpq-sz and swpocc have no values most of the time. But
>                 when the system is heavy loaded some values like 1.2 and 10%
>                 appear. I guess this is swapping ;-). But what hell does that
>  		mean in a paged system ?

in a paged system, you still try to do process swaps to reduce memory
contention between processes.  so this is still a measure of the amount
of memory contention.

>      sar -p : . The most stupid question due to my ignorance : what is a
>                 reclaimed page ? (rclm/s).

yeah, i have complaints about the name.  i can't remember if it means pages
reclaimed for the system (e.g. page steals) or whether it means pages
that are reclaimed by the page fault handler (e.g. finding it in the
page cache).  well, i just looked and it means the number of pages that
have been freed thru the kernel memfree() routine.  this can result from
page steals but can also result from process swaps, exits and other events
in which the system can reclaim memory.

>   Finally I'm interested in any idea or pointer about
>      
>      . Typical methods of diagnostics using sar.

try looking for a reference to 'TUNEX' in one of the USENIX proceedings
a couple of years ago.  its an expert system applied to the tuning of
unix systems for performance (pre-paging though...i think that behrohk
samadi, the TUNEX developer, has some plans to do a version for paging).

danny chen
att!homxc!dwc



More information about the Comp.unix.questions mailing list