[rescue] "svc.startd" question - 32% system utilization?

[Michael Parson]

On Fri, Sep 08, 2006 at 03:33:11PM -0500, Lionel Peterson wrote:
>> From: Bryan Gurney <arb_npx42 at comcast.net>
>> Date: 2006/09/08 Fri PM 01:34:09 CDT
>> To: The Rescue List <rescue at sunhelp.org>
>> Subject: Re: [rescue] "svc.startd" question - 32% system utilization?
>
>> On Fri, 08 Sep 2006 10:38:56 -0400, Lionel Peterson  
>> <lionel4287 at verizon.net>wrote:
>>
>>> Hello all,
>>>
>>> I have an Ultra 10 (440 MHz, 1024 Meg, IDE-based) acting as an
>>> install server (not Jumpstart, just serving up files for network
>>> installs), and while trying to get a Netra T1 setup, I noticed the
>>> "svc.startd" system process was consuming about 32% of the available
>>> system resources (as reported by 'top' - so apply a grain of salt to
>>> that number) - that seemed odd.
>>
>> Whenever I ran prstat or top on Solaris 9 or 10, the CPU usage
>> percentage figures seemed like a 1 minute average instead of a
>> near-instantaneous figure.  Can anybody confirm this somehow?  It
>> sort of bugs me to see a pegged process slowly crawl up in CPU usage
>> when top on Linux would show it as 99% instantaneously.
>
> The difference would be the time slice involved (obviously) - If it
> really was "instantaneous" it would only indicate "top" or "prstat"
> and peg it at 100% ;^)
>
> The thing I always heard from the performance engineers at $work-4 was
> that by imposing a load on the processor, top "pollutes" the results,
> and you need to consider that when reviewing results. 'vmstat' is
> similar, and imposes a much greater load on the CPU to reflect process
> start-up, you just give it a few intervals to let it "normalize".

It's a pretty basic tenant of science, you can't observe something
without changing the behavior.  You have to take into account the fact
that you are looking at it and what changes the tool used to make the
observation might have on the system.  There is a reason top defaults to
2 seconds and not every second.  I tend to prefer 3-5 second sampling.

-- 
Michael Parson
mparson at bl.org