beta of pfmon-3.2 available

beta of pfmon-3.2 available

[Stephane Eranian]

Hello,

I have finally released the first beta version of pfmon-3.2.
This version is a major update of version 3.0 released
2 years ago. They are many new features which I hope you'll find
useful.

Here are some of the new features:

	- support for perfmon v2.0 AND perfmon v2.2
	- support for all Itanium processors including Montecito
	- support for AMD X86-64 processors
	- support for Intel Pentium M/P6 processors
	- support for event sets and multiplexing (with perfmon v2.2)
	- instruction-based histrogram sampling module format
	- Data-EAR sampling module formats (Itanium only)
	- ability to exclude idle task for system-wide measurements
	- include or exclude interrupted triggered kernel execution for system
	  wide measurements (Itanium and perfmon v2.2 only )
	- support for rum/sum in monitored binaries (Itanium and perfmon v2.2 only)
	- support for cpu sets
	- man page for generic options

For non Itanium processors, the perfmon v2.2 patch must be applied to the kernel.
You need 2.6.17-rc1 or higher.

The Montecito support is currently missing some of the multi-threading features
as well as the cache MESI options for selected events.

You need libpfm-3.2-060421 or newer to compile pfmon. Similarly you may
want to use libunwind and libiberty to enable certain features. Make sure
you look at the README.

Full documentation including processors specific features is online at our
project's web site.

The tarball can be downloaded from our web site at:
	
		http://perfmon2.sf.net

Please test pfmon in your configuration and report any problems to the
perfmon mailing list.

Enjoy,

--
-Stephane

Re: beta of pfmon-3.2 available

[Andrew Morton]

On Wed, 26 Apr 2006 07:56:36 -0700
Stephane Eranian <eranian@hpl.hp.com> wrote:

> I have finally released the first beta version of pfmon-3.2.

perfctr was almost-ready-for-merge.  Then it was decided that perfmon was
the way ahead, and perfctr died.  And now perfmon isn't making progress in
the kernelwards direction (worse, perfmon is getting bigger, thus making a
merge harder and harder).  So we now have the worst of all worlds.

This is a problem.  I'd suggest that at this time we should be
concentrating on getting perfmon merged up rather than adding more stuff to
the out-of-tree version.

IOW: please send patches ;)

And keep sending them.  People want this.

Thanks.

Re: [perfmon] Re: beta of pfmon-3.2 available

[Stephane Eranian]

Andrew,

On Wed, May 03, 2006 at 05:47:11AM -0700, Andrew Morton wrote:
> 
> perfctr was almost-ready-for-merge.  Then it was decided that perfmon was
> the way ahead, and perfctr died.  And now perfmon isn't making progress in
> the kernelwards direction (worse, perfmon is getting bigger, thus making a
> merge harder and harder).  So we now have the worst of all worlds.
> 

I have been very busy cleaning up the code with some help of IBM. In the
last few months the code and was split into multiple C and header files.
The arch-specific modules for X86-64 and i386 have been merged.

I have also recently added a few features based upon requests from users.
For instance, the ability to automatically load PMU description modules. That
simplifies greatly the job of system administrators. I have also been busy talking
to Don Zickus about cleaning up the NMI watchdog timer and its use of performance
counters. He has produced a patch. I don't know if it is in your tree yet. Once that
happens, I will be able to simplify the perfmon code even more. So yes, the code is
changing but I think all of this is going into the direction of the kernel merge.

I have also received new feature requests from people of the SystemTap project.
They want to be able to call the interface from inside the kernel. This will add some
code but I don't anticipate it breaking the user level API. I think we all agree this
is something very useful to have. 

With Will Cohen from Redhat we are also looking at migrating Oprofile entirely over
to perfmon on x86 platforms. This can be accomplished very easily with perfmon. We did
that a long time ago on Itanium. In fact the exact same "glue" code can be shared between
all architectures. We are talking with John Levon about this. It is important that perfmon
provides smooth transition to Oprofile users.

> This is a problem.  I'd suggest that at this time we should be
> concentrating on getting perfmon merged up rather than adding more stuff to
> the out-of-tree version.
> 
I agree with you on this. Except for the recent Systemtap request, I do not think any *major* new
features were added in the last few months. By *major* I mean that changed the user level API
in an incompatbile way.

I may not have posted the full patches on lkml last time I released a patch but I will next time.

> IOW: please send patches ;)
> 
I was on vacation up until today. So I will likely release a new patch within a week.

> And keep sending them.  People want this.
> 
But please keep sending me constructive feeback as well.

-- 
-Stephane