Rafael Vanoni wrote:

> Hi Aubrey
> 
> Sorry for the late reply, I got caught up into another project.
> I'll work on your suggestions and have something for you next week.
> I'm going away for an OSS conference this week.
> 

No problem, ;-)

Thanks for your work!
-Aubrey

> 
> Aubrey Li wrote:
>> 2008/4/8 Rafael Vanoni <Rafael.Vanoni at sun.com>:
>>>  - added the -i option, which implements what we've been talking
>>> about correlating cyclic events with idle state transitions
>>> 
>> That's not what I expect.
>> 1) the bug is not fixed in the default output.
>> 2) output with -i option on my side, clock/cyclic wakeup is missing.
>> and -i option is useless by the common user.
>> ========
>> Wakeups-from-idle per second: 337.0     interval: 8.0s
>> 
>> Top causes for wakeups:
>> 44.6% (150.4)               <kernel> :
>> uhci`uhci_handle_root_hub_status_change
>> 19.8% ( 66.6)               <kernel> :
>> ehci`ehci_handle_root_hub_status_change
>> 10.0% ( 33.8)                   java :  <scheduled timeout
>> expiration> 
>>  3.0% ( 10.0)               <kernel> :  ata`ghd_timeout
>>  1.8% (  6.0)               <kernel> :  uhci`uhci_cmd_timeout_hdlr
>>  1.2% (  4.0)               <kernel> :  genunix`schedpaging
>>  0.9% (  3.0)               <kernel> :  genunix`lwp_timer_timeout
>>  0.8% (  2.7)            <interrupt> :  e1000g#1
>>  0.8% (  2.6)                  sched :  <scheduled timeout
>> expiration> 
>>  0.3% (  1.0)               <kernel> :  TS`ts_update
>>  0.3% (  1.0)               <kernel> :  e1000g`e1000g_local_timer
>>  ----snip---- ========
>> 
>> as we discussed before, the root cause is, cyclics are being batch
>> processed. so one wakeup has 3 or more cyclics. that causes the
>> count of "software waking up" doesn't match the counts of the "cpu
>> wakeups", and because these cyclics number is too large, always be
>> equal to the clock frequency, it looks very remarkable and
>> confusing.  
>> 
>> Maybe we should deal with this issue by this way:
>> 
>> (1) from the system admin's point of view, cyclics reports is
>> meaningless, even if they found inefficient cyclic subsystem
>> consumers, they can't do anything but just feel frustrated. However,
>> back to back applications is different, they can be killed by admin. 
>> 
>> (2) from the developer's point of view, cyclics report is necessary,
>> tracking down the inefficient consumers will improve the solaris
>> kernel. 
>> 
>> So I suggest we make (1) as default and add an option to enable (2)
>> or more events. 
>> 
>> Does this make sense?
>> 
>> Thanks,
>> -Aubrey
> 
> _______________________________________________
> tesla-dev mailing list
> tesla-dev at opensolaris.org
> http://mail.opensolaris.org/mailman/listinfo/tesla-dev


Reply via email to