dtrace discuss - Dec 2006 - Access hardware counters with DTrace

Could find a thread dated  jun 2005 where this was raised by Brendan Greg. 
Basic question still is if it is possible to get access to the CPU hardware
counters using DTrace.

--Paul
 
 
This message posted from opensolaris.org

Hi Paul,
>Could find a thread dated  jun 2005 where this was raised by Brendan Greg. 
>Basic question still is if it is possible to get access to the CPU hardware
counters using DTrace.
>
At the minute the answer is, no. However, I created a prototype
of a performance counter provider a while ago and I''ve (literally!)
just started another push on it. There is a placeholder RFE:

    6486156: DTrace cpc provider

If you are interested in getting some bits to play around
with, contact me off-alias.

Cheers.

Jon.

Frank Hoffman replied to Paul''s initial question but the cc:
got "magically" altered to ''opensolaris-discuss at
opensolaris.org''.
For those that aren''t on that alias, Frank''s reply was:
> Can you provide a reference URL for that thread ?
>
> The "normal" method to access the PMCs is via cpc(3CPC), or 
> cputrack(1M). Probably mentioned in the orig thread, hence the 
> question for the reference. Also curious as to what exactly it is you 
> need them for in DTrace ? Can you give us a "mockup D" of what
you''d
> like to have ?
>
> Thanks,
> FrankH. 

and my reply to Frank was:
>> The "normal" method to access the PMCs is via cpc(3CPC), or 
>> cputrack(1M). Probably mentioned in the orig thread, hence the 
>> question for the reference. Also curious as to what exactly it is you 
>> need them for in DTrace ? 
>
>
> The *huge* advantage that you get with DTrace is that of context.
> Being able to tie user-land or kernel context on a global basis
> to events is a massive advantage over what is there at the minute.
>
> Imagine just being able to get a global view of exactly where
> your DTLB misses are coming from. That''s powerful stuff.
>
> Jon.

So lets continue on this thread:

the URL requested:
http://www.opensolaris.org/jive/thread.jspa?messageID=3756&#3756

As Jon explained it is indeed the context that the events are happening that
might be extremely helpfull here. I know I have L2 misses. Not a clue
what''s causing this. Same holds for all other samplers available.
System wide statistisc do not enable easy drill down to the application(s) that
are causing it. It is even a bigger challenge to determine why they are causing
this.
Since most apps do not have appropriate instrumentation build in --and most who
do, don''t have docs helping you to understand which knobs, if any,
there are in order to fiddle with it. So the
next best thing is a util like DTrace. 

--Paul
 
 
This message posted from opensolaris.org

Hi Jon,

I am also looking something very similar. I want to use hardware performance
counters inside dtrace. Could you provide me some pointers if this could be
done.

Thanks,
Neelam


--
This message posted from opensolaris.org

Hi Neelam,
>I am also looking something very similar. I want to use hardware performance
counters inside dtrace. Could you provide me some pointers if this could be
done.
>  
>
The CPC provider I''m been working on is taking shape, albeit slowly.
I hope to have some archives ready for around the Christmas period
for interested parties to take for a test drive. I''ve actually had it 
working
for a long time but it''s undergoing some re-structuring and
augmentation
at the minute so it''s not quite ready for an outing in its current
form.

When it''s ready I''ll send something out to the list.

Jon.