No, actually, I think what Adrian is saying as that you can make better
use of your
time than trying to reconcile CPU utilization across different utilities
that use different
methods for doing the math. There's nothing interesting or useful in
having per-process
utilization align with system-wide utilization, and the fact that the
utilities used to
collect and report this data use different methods for doing so. The
system tools,
vmstat, sar, mpstat, use the kstats framework and clock-interrupt driven
granularity
for the data (in Solaris 8 and Solaris 9).
Solaris 8 and 9 both have various degrees of inaccuracy when it comes to
reporting
CPU utilization. Some time spent handling interrupts is missed, etc.
It's not unusual
for customers that move from Solaris 8 or 9 to Solaris 10 to report
problems about
using more CPU in Solaris 10. The fact is that, for the most part,
they're not using
more CPU, Solaris 10 implements a new mechanism for tracking CPU utilization
(CPU microstates), and is more accurate. Solaris 8 and 9 were using the
same amount
of CPU, it just was not getting reported by vmstat, etc...
So....if you want to get a sense for how busy your CPUs are system-wide,
use vmstat and/or mpstat. If you want to determine which processes/threads
are consuming CPU, and which are the top consumers, use prstat(1), and
prstat -m (microstates). You will have answers to both questions ("how busy
are my CPUs", and "which processes are burning CPU cycles, and which
processes are my top CPU consumers") using these tools, and that's all
you need
to know for performance/capacity work.
And, keep in mind that CPU utilization, in-and-of-itself, is useless. You
MUST correlate CPU utilization with a real workload metric (transastions
per minute, latency, batch run times, whatever)....
I've looked at hundreds of production systems where the tools reported fully
utilized CPUs, and were capable of doing more work within performance
requirements. I've looked at hundreds of problem systems (performance
problems) that had tons of idle CPU.
HTH,
/jim
CedBer wrote:
> Thank's for responses...
>
> So, you're saying that I must divide the top result (for a process cpu usage)
> by 8 rather than 4 (top and prstat will have the same value). And the
> difference between prstat (or top) sum and vmstat is "normal" ?!?
>
> CedBer
>
>
> This message posted from opensolaris.org
> _______________________________________________
> perf-discuss mailing list
> perf-discuss@opensolaris.org
>
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
