gmane.comp.linux.perfmon2.devel

Re: LIBPFM 4, POWER 8: Initial PMU event list

Stephane Eranian — 2013-06-17T20:19:17

Carl,

Patch applied.
It would be nice if you could provide some validation tests
in tests/validate_power.c. Just try to encode the key events
or events which are more difficult to handle than others.

I am surprised to see that with PPC8, there is no priv
level filter supported by the hardware. Can't I say, I
only want to monitor at kernel vs. user level?


Thanks.


On Mon, Jun 17, 2013 at 9:45 PM, Carl E. Love <cel< at >linux.vnet.ibm.com> wrote:

------------------------------------------------------------------------------
This SF.net email is sponsored by Windows:

Build for Windows Store.

http://p.sf.net/sfu/windows-dev2dev

Re: LIBPFM 4, POWER 8: Initial PMU event list

Carl E. Love — 2013-06-17T19:45:59

 <snip>


Yup, missed that, there is a comment about that at the end of the enum.
Fixed that.  This was the only comment I saw.  Here is the updated
patch.

          Carl Love

---------------------------------------------------------------------------------------

From: Carl Love <cel< at >us.ibm.com>
Date: Tue, 4 Jun 2013 12:50:00 -0500
Subject: [PATCH] Libpfm4: Add IBM Power 8 event support

This patch adds support for the Power 8 CPU events to the
libpfm4 library.

Signed-off-by: Carl Love <cel< at >us.ibm.com>
---
 README                     |    1 +
 include/perfmon/pfmlib.h   |    2 +
 lib/Makefile               |    3 +-
 lib/events/power8_events.h | 1009 ++++++++++++++++++++++++++++++++++++++++++++
 lib/pfmlib_common.c        |    1 +
 lib/pfmlib_power8.c        |   57 +++
 lib/pfmlib_power_priv.h    |    1 +
 lib/pfmlib_priv.h          |    1 +
 8 files changed, 1074 insertions(+), 1 deletion(-)
 create mode 100644 lib/events/power8_events.h
 create mode 100644 lib/pfmlib_power8.c

diff --git a/README b/REA

Re: LIBPFM 4, POWER 8: Initial PMU event list

Stephane Eranian — 2013-06-17T19:12:10

Must add at the end of the list, otherwise you're changing the ABI.


------------------------------------------------------------------------------
This SF.net email is sponsored by Windows:

Build for Windows Store.

http://p.sf.net/sfu/windows-dev2dev

LIBPFM 4, POWER 8: Initial PMU event list

Carl E. Love — 2013-06-17T18:18:40

Stephane:

The following patch adds the PMU event support for the IBM Power 8
processor to libfm4.  This is an initial list of events.  As additional
events are approved for use, we will add to this list.  

Please review and let me know if it is acceptable.  Thanks.

                       Carl Love

-----------------------------------------------------------------------------
Libpfm4: Add IBM Power 8 event support

This patch adds support for the Power 8 CPU events to the
libpfm4 library.

Signed-off-by: Carl Love <cel< at >us.ibm.com>
---
 README                     |    1 +
 include/perfmon/pfmlib.h   |    1 +
 lib/Makefile               |    3 +-
 lib/events/power8_events.h | 1009 ++++++++++++++++++++++++++++++++++++++++++++
 lib/pfmlib_common.c        |    1 +
 lib/pfmlib_power8.c        |   57 +++
 lib/pfmlib_power_priv.h    |    1 +
 lib/pfmlib_priv.h          |    1 +
 8 files changed, 1073 insertions(+), 1 deletion(-)
 create mode 100644 lib/events/power8_events.h
 create mode 100644 lib/pfmlib_power8.c

libpfm-4.4.0 finally released!

Stephane Eranian — 2013-06-14T21:30:38

Hi,

I am happy to announce that libpfm-4.4.0 has finally been released.
It was way overdue but I decided to delay to include Intel Haswell
desktop core PMU and Intel IvyBridge-EP core PMU support.

Highlights of this release:
- Intel Haswell desktop core PMU
- Intel IvyBridge desktop core and uncore PMU
- Intel IvyBridge-EP core PMU
- Intel SandyBridge uncore PMU
- Intel Knights Corner
- Intel Atom Penwell, Cloveriew core PMU support
- AMD Fam15h Northbridge support
- ldlat modified to support PEBS-LL
- Intel SandyBridge, Nehalem, Westmere event table updates
- IBM Power7 event updates


The new tarball is available at the usual place:
   http://perfmon2.sf.net

Thanks to all the contributors.

------------------------------------------------------------------------------
This SF.net email is sponsored by Windows:

Build for Windows Store.

http://p.sf.net/sfu/windows-dev2dev

Re: current libpfm4 git repository

William Cohen — 2013-06-03T20:44:25

Hi Stephane,

Yes, that is the latest commit on the following git repo:


 git://perfmon2.git.sourceforge.net/gitroot/perfmon2/libpfm4

There was a more recent checkin for papi's libpfm below which I was wondering where it came from:


commit eb0f46045775ca4af2f175b7108fa978d58815dc
Author: James Ralph <ralph< at >icl.utk.edu>
Date:   Wed May 15 15:24:30 2013 -0400

    Libpfm4 update add Intel IvyBridge EP core PMU support
    
    commit fd5e4b5384faf95986872a5a5b5574492352f26e
    Author: Stephane Eranian <eranian< at >gmail.com>
    Date:   Wed May 15 14:24:13 2013 +0200
    
    add Intel IvyBridge EP core PMU support
    
    Based on SDM Vol 3b Jan 2013.
    IvyBridge EP a.k.a IvyTown.
    
    Signed-off-by: Stephane Eranian <eranian< at >gmail.com>

This seems to be in the following git repo:

http://sourceforge.net/p/perfmon2/libpfm4/ci/master/tree/

-Will



------------------------------------------------------------------------------
How ServiceNow helps IT people transform IT departments:
1. A cloud service

Re: current libpfm4 git repository

Stephane Eranian — 2013-06-01T10:44:23

Will,

Top of tree should be:
6e150ca add Intel Haswell core PMU support



On Fri, May 31, 2013 at 6:18 PM, William Cohen <wcohen< at >redhat.com> wrote:

------------------------------------------------------------------------------
Get 100% visibility into Java/.NET code with AppDynamics Lite
It's a free troubleshooting tool designed for production
Get down to code-level detail for bottlenecks, with <2% overhead.
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap2

current libpfm4 git repository

William Cohen — 2013-05-31T16:18:35

Hi,

The papi git repository seems to have some more recent libpfm4 than the libpfm4 git repository listed on:

 http://perfmon2.sourceforge.net/

The perfmon2 web page lists:

  $ git clone git://perfmon2.git.sourceforge.net/gitroot/perfmon2/libpfm4

Is there a different libpfm4 git repo that has those newer libpfm4 checkins?

-Will

------------------------------------------------------------------------------
Get 100% visibility into Java/.NET code with AppDynamics Lite
It's a free troubleshooting tool designed for production
Get down to code-level detail for bottlenecks, with <2% overhead.
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap2

Re: pfmon for libpfm-4.3.0

Junjie Qian — 2013-05-24T05:17:48

Hi,

Seems pfmon-3.9 not supported after 3.6.?, and you may want to try perf (https://perf.wiki.kernel.org/index.php/Tutorial). 
 
Thanks!
Best
Junjie


________________________________
 From: Zheng Da <zhengda1936< at >gmail.com>
To: perfmon2-devel< at >lists.sourceforge.net 
Sent: Monday, May 20, 2013 2:40 PM
Subject: [perfmon2] pfmon for libpfm-4.3.0
 


Hello,

I'm new to perform2. I want to use it to monitor Xeon E5 processors. It seems only libpfm-4.3.0 supports it. However, pfmon-3.9 can't be compiled with libpfm-4.3.0.
I have compiled and installed libpfm-4.3.0, but when I compiled pfmon-3.9, I got the following error:


make[1]: Entering directory `/home/zhengda/pfmon-3.9/pfmon'

cc  -g -ggdb -Wall -Werror -D_REENTRANT -I/usr/local/include -DCONFIG_PFMON_X86_64 -DPFMON_DEBUG -DDATADIR=\"/usr/local/share/pfmon\" -I. -I/usr/include/libelf -D_GNU_SOURCE -DPFMON_DEBUG -g -c pfmon.c
In file included from pfmon_support.h:28:0,
                 from pfmon.c:31:
pfmon.h:38:29: fatal error: perfmon/perfmon

Re: pfmon for libpfm-4.3.0

Stephane Eranian — 2013-05-20T20:17:05

Hi,

The perfmon2 project has been stopped in 2008 and at the 2.6.30 kernel version.

Libpfm-4.3 is meant to be used with the official perf_event kernel
interface which
is similar to the perfmon2 interface.

Consequently, the development of pfmon has stopped.
You need to use the perf tool now.

On Mon, May 20, 2013 at 9:40 PM, Zheng Da <zhengda1936< at >gmail.com> wrote:

------------------------------------------------------------------------------
AlienVault Unified Security Management (USM) platform delivers complete
security visibility with the essential security capabilities. Easily and
efficiently configure, manage, and operate all of your security controls
from a single console and one unified framework. Download a free trial.
http://p.sf.net/sfu/alienvault_d2d

pfmon for libpfm-4.3.0

Zheng Da — 2013-05-20T19:40:07

Hello,

I'm new to perform2. I want to use it to monitor Xeon E5 processors. It
seems only libpfm-4.3.0 supports it. However, pfmon-3.9 can't be compiled
with libpfm-4.3.0.
I have compiled and installed libpfm-4.3.0, but when I compiled pfmon-3.9,
I got the following error:

make[1]: Entering directory `/home/zhengda/pfmon-3.9/pfmon'
cc  -g -ggdb -Wall -Werror -D_REENTRANT -I/usr/local/include
-DCONFIG_PFMON_X86_64 -DPFMON_DEBUG -DDATADIR=\"/usr/local/share/pfmon\"
-I. -I/usr/include/libelf -D_GNU_SOURCE -DPFMON_DEBUG -g -c pfmon.c
In file included from pfmon_support.h:28:0,
                 from pfmon.c:31:
pfmon.h:38:29: fatal error: perfmon/perfmon.h: No such file or directory
compilation terminated.
make[1]: *** [pfmon.o] Error 1
make[1]: Leaving directory `/home/zhengda/pfmon-3.9/pfmon'
make: *** [all] Error 2

I can't find perfmon/perfmon.h anywhere. It's not in libpfm-4.3.0 or any
other Ubuntu packages. I use Ubuntu 12.04 by the way.
So what is the fix?

Thanks,
Da
------------------------------------

Re: Ivy Bridge Uncore support

Vince Weaver — 2013-05-03T21:44:48

well, for the obvious reasons that Intel doesn't support them.

What I'm saying is that my hack only added a model ID for the uncore, 
which is a separate PMU for the regular events.  Hence I didn't switch
my machine to all SNB events.. it still has the default IVB enabled.
I just added the SNB uncore in addition.


When I first started gathering results there weren't any IVB or SNB 
machines, and I just kept collecting the same set of events.

With x87 there is often a "gather all x87 ops" event.  As we all know
there's no direct SSE equivelent, you end up needing to collect like 5 
different events and add them all up, and I don't think anyone has really 
tested that that gives you anything useful.


it's easy enough with perf to just use the raw event numbers.  Would it 
make sense to export that capability to PAPI somehow?

Then if you really wanted to force things you could just put something 
like "r50003c:u" or whatever as the event name in papi_events.csv

Vince

------------------------------------

Re: Ivy Bridge Uncore support

Vince Weaver — 2013-05-03T15:42:27


My "hack" was only for the uncore, the regular Ivybridge events have been 
supported properly for some time.

You can look at the results in Table VI
of my recent ISPASS paper
 
  http://web.eece.maine.edu/~vweaver/projects/deterministic/ispass2013_deterministic.pdf

and see that x87 fp results using the undocumented events on IVB match 
pretty well with SNB.  The SSE results do not though.  And as for
SIMD_FP_256 I don't think I have any AVX test code handy.

It's not easy to re-add those events to PAPI though because libpfm4 
doesn't support them.

Vince

------------------------------------------------------------------------------
Get 100% visibility into Java/.NET code with AppDynamics Lite
It's a free troubleshooting tool designed for production
Get down to code-level detail for bottlenecks, with <2% overhead.
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap2

Re: Ivy Bridge Uncore support

Vince Weaver — 2013-05-02T20:53:42

So I am using my hacked version of libpfm4 that is detecting the 
model 58 IVB uncore as the same as model 42 SNB.

I have a few questions about libpfm4 uncore support.

For one, it looks like 4 different CBOXes are being created

[76, snb_unc_cbo0, "Intel Sandy Bridge C-box0 uncore"]
[77, snb_unc_cbo1, "Intel Sandy Bridge C-box1 uncore"]
[78, snb_unc_cbo2, "Intel Sandy Bridge C-box2 uncore"]
[79, snb_unc_cbo3, "Intel Sandy Bridge C-box3 uncore"]

but except for c-box0 having UNC_CLOCKTICKS they all seem to have 
identical events?  Is there a reason we need to export 4 pmus if
the events are all the same?  It's especially confusing as my IVB system
only has two cboxes.

Also I notice that 
   events/intel_snb_unc_events.h
lists the ARB uncore events but doesn't create a PMU for them.
Though now that I check, the linux-kernel doesn't seem to support the ARB
uncore.  Is that why it's not enabled in libpfm4?

Thanks,

Vince

------------------------------------------------------------------------------
Get

Re: [PATCH] Power 7, change the code for the PM_RUN_INST_CMPL and PM_RUN_CYC events

Carl E. Love — 2013-05-01T19:31:07

Stephane:

I withdraw the patch.

                  Carl Love

On Wed, 2013-05-01 at 21:27 +0200, Stephane Eranian wrote:



------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1

Re: [PATCH] Power 7, change the code for the PM_RUN_INST_CMPL and PM_RUN_CYC events

Stephane Eranian — 2013-05-01T19:27:29

Ok, so that means your patch is not required, doesn't it?


------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1

Re: [PATCH] Power 7, change the code for the PM_RUN_INST_CMPL and PM_RUN_CYC events

Carl E. Love — 2013-05-01T19:24:22

Stephane:

Yes, on further investigation it does appear that the kernel code should
do the right thing and use the fixed counter to make room for another
event on the programmable counter.  

                 Carl Love

On Wed, 2013-05-01 at 02:27 +0200, Stephane Eranian wrote:



------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1

Re: [PATCH] Power 7, change the code for the PM_RUN_INST_CMPL and PM_RUN_CYC events

Stephane Eranian — 2013-05-01T00:27:44

Something not clear to me here.
perf_event in the kernel does the event scheduling. So is it the case that
on Power, it is not able to recognize the event as also being supported
by a fixed counter?


------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1

Re: Ivy Bridge Uncore support

Stephane Eranian — 2013-04-30T19:48:52

Yes, it is. But as I said before. I cannot release the IVB-EP event table
until Intel releases the specs. I will do that immediately when this happens.
So for now, the best we can do is pretend it is equivalent to IVB.


------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1

Re: [PATCH] Power 7, change the code for the PM_RUN_INST_CMPL and PM_RUN_CYC events

Carl E. Love — 2013-04-30T17:24:16

The events are identical for the two counters.  In general, I don't
think the users know that there are two types of counters.  So, if we
gave them both values and the user picked the value for the programmable
counter for use in perf along with say four additional events then perf
would have to multiplex the five events on four programmable counters.
But if they were to pick the event for the fixed counter, perf would not
multiplex the events.  Not sure how you get a tool like event2raw to
give both codes plus the explanation about programmable and fixed
counters so the user would know the ramifications of choosing one versus
the other.  

We don't see any use scenario where using the event code in the
programmable counter would be needed or preferable to using the fixed
counter as the results are identical.  So, it seems better to have one
event for the fixed counter so the user doesn't have to know the details
of fixed versus programmable counters or inadvertently have events
multiplexed when it is not n

Re: [PATCH] Power 7, change the code for the PM_RUN_INST_CMPL and PM_RUN_CYC events

Vince Weaver — 2013-04-30T16:27:39


Instead of replacing the definition, would it make sense to just add an 
additional definition so there is two events listed, one fixed, one not?
That way a user could pick which one they wanted?

I guess on POWER it might not matter, but at least on recent Intel chips 
it would be nice if we could specify the difference as it turns out the 
fixed vs programmable events have subtle differences and it's hard to test 
this as it's impossible to specify one vs the other.

Vince

------------------------------------------------------------------------------
Introducing AppDynamics Lite, a free troubleshooting tool for Java/.NET
Get 100% visibility into your production application - at no cost.
Code-level diagnostics for performance bottlenecks with <2% overhead
Download for free and get started troubleshooting in minutes.
http://p.sf.net/sfu/appdyn_d2d_ap1