On Mon, Apr 27, 2009 at 5:16 AM, Matt Ingenthron <[email protected]> wrote:
> Aubrey Li wrote:
>>
>> On Sun, Apr 26, 2009 at 9:44 PM, Mark Haywood <[email protected]>
>> wrote:
>>
>>>>
>>>> Should've been more specific in my reply, so:
>>>>
>>>> % vi /etc/power.conf  # Change to poll-mode
>>>> % pfexec pmconfig
>>>>
>>>> % pfexec powertop # shows proc. always at max speed % pfexec reboot #
>>>> After the reboot, powertop shows proc. at lower freq.
>>>>
>>>>
>>>
>>> It is either a bug or your system isn't idle enough for poll-mode to
>>> lower
>>> the freq. either. I have successfully switched event-mode and poll-mode
>>> many
>>> times without a reboot.
>>>
>>>
>>
>> I checked it just now on my system(OpenSolaris200906-snv111),
>> it works properly. When change to poll-mode, you'd better to set the
>> "cpu-threshold" to "1s". The default scanning cycle is 15s, which means
>> you need to wait for a while(>15s) to see the frequency transition.
>> ============================
>> cpupm                   enable poll-mode
>> cpu-threshold           1s
>> ============================
>> After issue "pmconfig", you should see the lowest frequency residency
>> in powertop if
>> your system is idle.
>>
>
> Some observations after some experimentation on my systems.  All were done
> with the system effectively 'idle'.  X was up with a GNOME session, and
> compiz was on, but most other apps that I usually run were down.  I was
> using gnome-terminal.
>
> If I switch to event-mode, powertop does occasionally, for a refresh, show
> the pstate changing from something other than the highest frequency.
>  Observing it for a minute or so shows that it only happens maybe 10-20% of
> the time.  via mpstat 5, both cores show a median of about 97% idle (just
> eyeball observations)
>
> In poll-mode with cpu-threshold set to 3s, I see the pstate changing much
> more frequently.  I also see pstates other than just the highest frequency
> and the lowest frequency.
> Like others have observed, powertop only ever shows me the 100% in any one
> pstate.  I believe that's where someone mentioned a bug.

Yes, As Mark pointed, it's a known bug:
http://bugs.opensolaris.org/bugdatabase/view_bug.do?bug_id=6829404.

> So I guess my question is, how should I get data to know if my usage
> pattern/system is best in event-mode or poll-mode?

Two big differences between event-mode and poll-mode:
1. response latency.
event-mode response much more quick than poll-mode. frequency transition in
the poll-mode is periodical and the time is specified by
"cpu-threshold", while in
the event-mode, the transition occurs in the thread switching, much
more sensitive
than the poll-mode.

2. transition condition.
poll-mode is using cpu utilization while event-mode is using cmt utilization.
event-mode also installs two governors to avoid transient load and
transient idle.
That means in event-mode: 1) very light load could generate frequency
transition.
2) persistent load will not generate frequency transtion. So, if the
load prefers the
 % cpu utilization, poll-mode is more reliable.

Thanks,
-Aubrey
_______________________________________________
indiana-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/indiana-discuss

Reply via email to