On Sun, Feb 8, 2009 at 4:11 AM, Ben Rockwood <benr at cuddletech.com> wrote:

> I'm generally happy with the zone level kstat data available today.  I
> think that maintaining the kind of granular data your talking about via
> kstats is excessive... but I'd be interested in other opinions.
>

I agree with Ben here, yes all the information in kstat is nice, but we do
have a problem with growing kstat, I'm not an experpert machine language
developer but here is my guess to what we will run into, yes the CPU and
compiler can hide away the increment of a memory location using spare
cycles,  but thanks to Solaris wider  CPU support gone are the days where
all huge  CPU caches, the typical x86 cpu has 512KB of cache, and every time
we increment a kstat counter we muddy the cache lines, so eventually we
could end up with 1000's of counters each using at lest 4bytes in l1 and l2
cache doing nothing more than counting things that most people won' t even
look at slowing down the systems. I think new counters shouldn't be
implemented unless the OS it self can use them to better tune it self
gaining performance. Otherwise the user can just  watch what ever they need
to using dtrace without the permanent cost of built in kstat growth.

>
> Using the existing kstats in conjunction with a process that "rolls up"
> daily extended accounting data would allow me to see not just how much CPU
> and memory a zone is using, but what processes they are most actively
> running... a historical top 10 workloads by day, etc.  That could be handy.
>

yes that would be nice.

James Dickens
uadmin.blogspot.com


>
> One thing that does bug me is that quantifying how much performance impact
> exacct has is very hard to determine.  I'd like more research there.  (The
> same goes for BSM actually.)
>
>
> As for improving the visability into sys%, I totally agree.  As the Solaris
> kernel gets more and more complex I think we're all seeing a lot more system
> time on our production systems and quantifying what that is, even with
> DTrace, is very complex.
>
> benr.
> --
> This message posted from opensolaris.org
> _______________________________________________
> observability-discuss mailing list
> observability-discuss at opensolaris.org
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: 
<http://mail.opensolaris.org/pipermail/observability-discuss/attachments/20090208/760c212f/attachment.html>

Reply via email to