On Tue, Jun 20, 2006 at 10:48:46PM -0400, Dennis Clarke wrote:
>
> I have a system here that runs snv_38 and I have a bit of code that cranks a
> large pile of numbers about. All of this is integer operations with the
> exception of maybe two lines that require floating point. That would be two
> lines out of about 200 lines of addition, subtractions, integer comparisons
> etc etc and all in all the code bit runs for about 19 secs.
>
> Now mpstat reports that one proc is totally tied up with the process thus :
>
> $ mpstat 2
...
> CPU minf mjf xcal intr ithr csw icsw migr smtx srw syscl usr sys wt idl
> 0 0 0 35 5414 210 103 0 2 0 0 162 0 4 0 96
> 1 0 0 5103 81 3 2 3 0 0 0 4 0 100 0 0
^^^^^^^^^^^^^
...
This indicates that almost no time is being spent in userland, and one CPU
is busy only doing kernel work. You probably have either a kernel thread
or heavy interrupt traffic eating up CPU 1.
If you want to investigate, something like:
dtrace -n '[EMAIL PROTECTED]()] = count()}' -n 'END{trunc(@a,20)}' \
-c './r'
will give you the most common kernel stacks while your process is running.
There might be a clue from that.
...
> PID USERNAME SIZE RSS STATE PRI NICE TIME CPU PROCESS/NLWP
> 1996 dclarke 1020K 704K cpu1 0 0 0:00:17 29% r/1
> 1994 dclarke 2772K 2476K cpu0 59 0 0:00:00 0.2% prstat/1
> 1995 dclarke 2084K 1336K sleep 59 0 0:00:00 0.1% mpstat/1
> 1978 dclarke 11M 3388K sleep 59 0 0:00:00 0.0% sshd/1
> 1921 dclarke 11M 3388K sleep 59 0 0:00:00 0.0% sshd/1
...
> So either mpstat is telling me that one processor is indeed totally busy or
> prstat is telling me that something called "r" is just taking 29% of the
> system CPU resources at its peak.
The CPU measurement ramps up slowly. The "PRI" of 0 indicates that you are
heavily CPU bound.
You should look at the prstat -m information for your process; it will
let you know precisely where its time is being spent. (the results are not
decayed)
Cheers,
- jonathan
--
Jonathan Adams, Solaris Kernel Development
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
