Event not supported

Event not supported

[Martin Ichilevici de Oliveira]

Hello,

I'm still learning perf, so I apologize in advance if my questions are
too simple and not suited here. I did, however, searched the web and
tried to find a solution by myself before posting to this mailing
list.

From my manufacturer's manual, I chose an event I'm interested in
monitoring. They also list unit masks, as below:

UnitMask    Description
7                  From Local node to Node 7
6                  From Local node to Node 6
5                  From Local node to Node 5
4                  From Local node to Node 4
3                  From Local node to Node 3
2                  From Local node to Node 2
1                  From Local node to Node 1
0                  From Local node to Node 0

The event is "CPU to DRAM Requests to Target Node" (NUMA machine) and
it's number is 0x1E0. As I understood, I should precede the event with
the unit mask. So, for example, if I'm interested in unit mask 2, I
should type:

perf stat -e r21e0 /bin/ls

However, this results in a <not supported> message. This happened for
all events that have a unit mask that I tried. Other events worked
fine.

I'm not sure whether I'm giving the right command or this is a problem
with perf/kernel version. As it turned out, I'm running Ubuntu with
kernel 3.11.10-03111002, but the latest available perf version from
the repositories is 3.8.0-35, which is the one I'm using.

Any help is appreciated.

Thank you,
Martin

Re: Event not supported

[Manuel Selva]

Hi,

What is your CPU ?

Regarding the perf version, on my Ubuntu 13.04 running linux 3.11.0-15 I 
successfully installed the associated perf version using:

sudo apt-get install linux-tools-3.11.0-15-generic


Manu

On 02/18/2014 05:51 PM, Martin Ichilevici de Oliveira wrote:
> Hello,
>
> I'm still learning perf, so I apologize in advance if my questions are
> too simple and not suited here. I did, however, searched the web and
> tried to find a solution by myself before posting to this mailing
> list.
>
>  From my manufacturer's manual, I chose an event I'm interested in
> monitoring. They also list unit masks, as below:
>
> UnitMask    Description
> 7                  From Local node to Node 7
> 6                  From Local node to Node 6
> 5                  From Local node to Node 5
> 4                  From Local node to Node 4
> 3                  From Local node to Node 3
> 2                  From Local node to Node 2
> 1                  From Local node to Node 1
> 0                  From Local node to Node 0
>
> The event is "CPU to DRAM Requests to Target Node" (NUMA machine) and
> it's number is 0x1E0. As I understood, I should precede the event with
> the unit mask. So, for example, if I'm interested in unit mask 2, I
> should type:
>
> perf stat -e r21e0 /bin/ls
>
> However, this results in a <not supported> message. This happened for
> all events that have a unit mask that I tried. Other events worked
> fine.
>
> I'm not sure whether I'm giving the right command or this is a problem
> with perf/kernel version. As it turned out, I'm running Ubuntu with
> kernel 3.11.10-03111002, but the latest available perf version from
> the repositories is 3.8.0-35, which is the one I'm using.
>
> Any help is appreciated.
>
> Thank you,
> Martin
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

Re: Event not supported

[Martin Ichilevici de Oliveira]

Hello Manuel,

My CPU is an AMD Interlagos, model 6272. All events that I'm unable to
profile are Northbridge events. I have just downgraded my kernel to
3.8.0-35 (matching my perf version), but the problem persists.

I was reading the proposed patch that added support for Northbridge
events [1] and the author mentions
"NB counters are enabled if the nb performance counter extensions
cpuid flag is set.

How can I check if such flag is set and, if it's disabled, set it?

Thank you,
Martin

[1]: http://www.spinics.net/lists/kernel/msg1359486.html

On Tue, Feb 18, 2014 at 3:22 PM, Manuel Selva <selva.manuel@gmail.com> wrote:
> Hi,
>
> What is your CPU ?
>
> Regarding the perf version, on my Ubuntu 13.04 running linux 3.11.0-15 I
> successfully installed the associated perf version using:
>
> sudo apt-get install linux-tools-3.11.0-15-generic
>
>
> Manu
>
>
> On 02/18/2014 05:51 PM, Martin Ichilevici de Oliveira wrote:
>>
>> Hello,
>>
>> I'm still learning perf, so I apologize in advance if my questions are
>> too simple and not suited here. I did, however, searched the web and
>> tried to find a solution by myself before posting to this mailing
>> list.
>>
>>  From my manufacturer's manual, I chose an event I'm interested in
>> monitoring. They also list unit masks, as below:
>>
>> UnitMask    Description
>> 7                  From Local node to Node 7
>> 6                  From Local node to Node 6
>> 5                  From Local node to Node 5
>> 4                  From Local node to Node 4
>> 3                  From Local node to Node 3
>> 2                  From Local node to Node 2
>> 1                  From Local node to Node 1
>> 0                  From Local node to Node 0
>>
>> The event is "CPU to DRAM Requests to Target Node" (NUMA machine) and
>> it's number is 0x1E0. As I understood, I should precede the event with
>> the unit mask. So, for example, if I'm interested in unit mask 2, I
>> should type:
>>
>> perf stat -e r21e0 /bin/ls
>>
>> However, this results in a <not supported> message. This happened for
>> all events that have a unit mask that I tried. Other events worked
>> fine.
>>
>> I'm not sure whether I'm giving the right command or this is a problem
>> with perf/kernel version. As it turned out, I'm running Ubuntu with
>> kernel 3.11.10-03111002, but the latest available perf version from
>> the repositories is 3.8.0-35, which is the one I'm using.
>>
>> Any help is appreciated.
>>
>> Thank you,
>> Martin
>> --
>> To unsubscribe from this list: send the line "unsubscribe
>> linux-perf-users" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>
>

Re: Event not supported

[Manuel Selva]

CPUID is an X86 instruction: http://en.wikipedia.org/wiki/CPUID used to 
get information about the CPU. The Core performance counter extensions 
is mentioned on the wikipedia page.

You'll need to write assembly code to use this instruction or use 
cpuid.h. Here is an example of using this instruction from a previous 
answer on this list:

#include <cpuid.h>
#include <stdio.h>
int main()
{
	unsigned a, b, c, d;
	/* check __get_cpuid_max here */
	__cpuid(10, a, b, c, d);
	printf("eax: %x ebx %x ecx %x edx %x\n", a, b, c, d);
	int i;
	for (i = 0; i < 10; i++)
		if (b & (1 << i))
			printf("event %d not supported\n", i);
	return 0;
}


Manu

On 02/18/2014 07:42 PM, Martin Ichilevici de Oliveira wrote:
> performance counter extensions
> cpuid

Re: Event not supported

[Martin Ichilevici de Oliveira]

Manu,

Thank you very much for the reference and code. I checked my cpu and
the Northbridge performance counter flag is set. So my processor does
support profiling NB events. The question now is why it's saying it's
not.

I realized that perf support for NB events was in fact only added on
linux 3.9 [1], so I upgraded both my kernel and perf to 3.11. So
that's out.

The only thing I can think of is the the simple case that I'm just
using the wrong command. As I said, I'm not entirely sure how to use
unit masks. Can you please confirm that my usage is correct?

Do you have any other ideas on what might be the issue?

Thank you,
Martin

[1]: https://github.com/torvalds/linux/commit/e259514eef764a5286873618e34c560ecb6cff13

On Tue, Feb 18, 2014 at 4:50 PM, Manuel Selva <selva.manuel@gmail.com> wrote:
> CPUID is an X86 instruction: http://en.wikipedia.org/wiki/CPUID used to get
> information about the CPU. The Core performance counter extensions is
> mentioned on the wikipedia page.
>
> You'll need to write assembly code to use this instruction or use cpuid.h.
> Here is an example of using this instruction from a previous answer on this
> list:
>
> #include <cpuid.h>
> #include <stdio.h>
> int main()
> {
>         unsigned a, b, c, d;
>         /* check __get_cpuid_max here */
>         __cpuid(10, a, b, c, d);
>         printf("eax: %x ebx %x ecx %x edx %x\n", a, b, c, d);
>         int i;
>         for (i = 0; i < 10; i++)
>                 if (b & (1 << i))
>                         printf("event %d not supported\n", i);
>         return 0;
> }
>
>
> Manu

Re: Event not supported

[Vince Weaver]

On Wed, 19 Feb 2014, Martin Ichilevici de Oliveira wrote:

> Manu,
> 
> Thank you very much for the reference and code. I checked my cpu and
> the Northbridge performance counter flag is set. So my processor does
> support profiling NB events. The question now is why it's saying it's
> not.
> 
> I realized that perf support for NB events was in fact only added on
> linux 3.9 [1], so I upgraded both my kernel and perf to 3.11. So
> that's out.

The support for AMD fam15h northbridge events was introduced in 3.9, but
the perf interface for it was changed in an incompatible way in 3.10.

To access the counters starting in 3.10 they are in a separate PMU.
You need to be root and you need to set the type field to the proper PMU.

You might be better off using something like libpfm4 to find out what aw 
values to use, trying to set them by hand is a hassle.

Vince

Re: Event not supported

[Martin Ichilevici de Oliveira]

On Wed, Feb 19, 2014 at 9:59 AM, Vince Weaver <vincent.weaver@maine.edu> wrote:
>
> The support for AMD fam15h northbridge events was introduced in 3.9, but
> the perf interface for it was changed in an incompatible way in 3.10.
>
> To access the counters starting in 3.10 they are in a separate PMU.
> You need to be root and you need to set the type field to the proper PMU.
>
> You might be better off using something like libpfm4 to find out what aw
> values to use, trying to set them by hand is a hassle.
>
> Vince

Vince,

as I could understand, any tool that I use to profile AMD fam15h
northbridge events will depend on Linux 3.9 or newer, am I right? If
that's the case, would you say that currently Linux 3.9 is the
best/easiest way to go?

I've been playing with kernel versions trying to figure this out and a
few times the machine broke. Since that's a remote machine, that's a
hassle, as I have to ask people to fix it for me, so I'd like to
minimize the chances of problems. I'm currently on Linux 3.8.

Thanks,
-Martin

Re: Event not supported

[Vince Weaver]

On Wed, 26 Feb 2014, Martin Ichilevici de Oliveira wrote:

> as I could understand, any tool that I use to profile AMD fam15h
> northbridge events will depend on Linux 3.9 or newer, am I right? If
> that's the case, would you say that currently Linux 3.9 is the
> best/easiest way to go?

It depends where your kernel comes from.

If you're using a stock Linux kernel compiled yourself, then yes you 
probably need at least 3.9 for it to work.

Sometimes if you're using a vendor supplied kernel they backport support 
for features, so it is not always clear what support there is in a kernel.

Linux 3.9 and Linux 3.10 have very different amd15h northbridge support.  
If you're looking for long-term usability of your project/tool/research 
you probably want to jump straight to 3.10 so you don't have to make 
everything work on 3.9 and then change it all again when you move to newer 
kernels.

Vince

Re: Event not supported

[Martin Ichilevici de Oliveira]

On Wed, Feb 26, 2014 at 6:38 PM, Vince Weaver <vincent.weaver@maine.edu> wrote:
> On Wed, 26 Feb 2014, Martin Ichilevici de Oliveira wrote:
>
>> as I could understand, any tool that I use to profile AMD fam15h
>> northbridge events will depend on Linux 3.9 or newer, am I right? If
>> that's the case, would you say that currently Linux 3.9 is the
>> best/easiest way to go?
>
> It depends where your kernel comes from.

I got it from the Ubuntu kernel ppa.

> If you're using a stock Linux kernel compiled yourself, then yes you
> probably need at least 3.9 for it to work.
>
> Sometimes if you're using a vendor supplied kernel they backport support
> for features, so it is not always clear what support there is in a kernel.
>
> Linux 3.9 and Linux 3.10 have very different amd15h northbridge support.
> If you're looking for long-term usability of your project/tool/research
> you probably want to jump straight to 3.10 so you don't have to make
> everything work on 3.9 and then change it all again when you move to newer
> kernels.
>
> Vince

I installed Linux 3.10 but it's still not clear to me how to profile
amd15h northbridge events, as I keep getting unsupported event
messages. Also, I tried using papi_component_avail and it says that
perf_events_uncore is disabled (no uncore PMUs or or events found).

In a previous message you mentioned

> To access the counters starting in 3.10 they are in a separate PMU.
> You need to be root and you need to set the type field to the proper PMU.

I tried with root but I wasn't sure how to set the type field to the
proper PMU. Could you please clarify? This has been a lot of trial and
(mostly) error. In the meantime, I went on to try your other
suggestion:

> You might be better off using something like libpfm4 to find out what aw
> values to use, trying to set them by hand is a hassle.

I tried using pfm_get_event_os_event_encoding() from libpfm4, but with
no success. The manpage for this function is from 2011 and most likely
lots has changed since. I've been reading the source to try to figure
something out, but I couldn't figure it out.

Any help is appreciated.

Thank you,
-Martin

Re: Event not supported

[Martin Ichilevici de Oliveira]

On Fri, Feb 28, 2014 at 6:35 AM, Martin Ichilevici de Oliveira
<martin.i.oliveira@gmail.com> wrote:
> On Wed, Feb 26, 2014 at 6:38 PM, Vince Weaver <vincent.weaver@maine.edu> wrote:
>> On Wed, 26 Feb 2014, Martin Ichilevici de Oliveira wrote:
>>
>>
>> Linux 3.9 and Linux 3.10 have very different amd15h northbridge support.
>> If you're looking for long-term usability of your project/tool/research
>> you probably want to jump straight to 3.10 so you don't have to make
>> everything work on 3.9 and then change it all again when you move to newer
>> kernels.
>>
>> Vince
>
> I installed Linux 3.10 but it's still not clear to me how to profile
> amd15h northbridge events, as I keep getting unsupported event
> messages. Also, I tried using papi_component_avail and it says that
> perf_events_uncore is disabled (no uncore PMUs or or events found).

Do I need to compile my kernel (3.10) with explicit support? I don't
understand why papi_component_avail keeps saying that
perf_events_uncore is disabled.

Re: Event not supported

[Vince Weaver]

On Tue, 11 Mar 2014, Martin Ichilevici de Oliveira wrote:

> On Fri, Feb 28, 2014 at 6:35 AM, Martin Ichilevici de Oliveira
> <martin.i.oliveira@gmail.com> wrote:
> > On Wed, Feb 26, 2014 at 6:38 PM, Vince Weaver <vincent.weaver@maine.edu> wrote:
> >> On Wed, 26 Feb 2014, Martin Ichilevici de Oliveira wrote:
> >>
> >>
> >> Linux 3.9 and Linux 3.10 have very different amd15h northbridge support.
> >> If you're looking for long-term usability of your project/tool/research
> >> you probably want to jump straight to 3.10 so you don't have to make
> >> everything work on 3.9 and then change it all again when you move to newer
> >> kernels.
> >>
> >> Vince
> >
> > I installed Linux 3.10 but it's still not clear to me how to profile
> > amd15h northbridge events, as I keep getting unsupported event
> > messages. Also, I tried using papi_component_avail and it says that
> > perf_events_uncore is disabled (no uncore PMUs or or events found).
> 
> Do I need to compile my kernel (3.10) with explicit support? I don't
> understand why papi_component_avail keeps saying that
> perf_events_uncore is disabled.

If you're having issues with PAPI it might be better to bring your 
questions to one of the PAPI mailing lists.

If you have a 3.10 x86_64 kernel it's not possible to disable perf_event 
or uncore support, so support is there.

You will need to run as root (or set the perf_event paranoid setting away 
from the default) to access the NorthBridge counters though.

PAPI at least is not well tested on amd fam15h hardware as we don't have 
acess to any test systems with the hardware.

Vince

Re: Event not supported

[Baptiste Lepers]

Hi,

I think that the format you are using is not correct.

To profile event 0xXXX with unit mask YY, the format is rX0000YYXX, so in your case: r1000002e0.

Also, beware that the unitmask in the documentation tells you the _bit(s)_ that need to be set.
So :
  From Local Node to Node 0 = r1000001e0
  From Local Node to Node 1 = r1000002e0
  From Local Node to Node 2 = r1000004e0
  From Local Node to Node 3 = r1000008e0
  From Local Node to Node 4 = r1000010e0
  From Local Node to Node 0 & 1 & 2 = r1000007e0
  etc.


Cheers,
Baptiste.

----- Mail original -----
> De: "Martin Ichilevici de Oliveira" <martin.i.oliveira@gmail.com>
> À: linux-perf-users@vger.kernel.org
> Envoyé: Mardi 18 Février 2014 17:51:46
> Objet: Event not supported
> 
> Hello,
> 
> I'm still learning perf, so I apologize in advance if my questions are
> too simple and not suited here. I did, however, searched the web and
> tried to find a solution by myself before posting to this mailing
> list.
> 
> From my manufacturer's manual, I chose an event I'm interested in
> monitoring. They also list unit masks, as below:
> 
> UnitMask    Description
> 7                  From Local node to Node 7
> 6                  From Local node to Node 6
> 5                  From Local node to Node 5
> 4                  From Local node to Node 4
> 3                  From Local node to Node 3
> 2                  From Local node to Node 2
> 1                  From Local node to Node 1
> 0                  From Local node to Node 0
> 
> The event is "CPU to DRAM Requests to Target Node" (NUMA machine) and
> it's number is 0x1E0. As I understood, I should precede the event with
> the unit mask. So, for example, if I'm interested in unit mask 2, I
> should type:
> 
> perf stat -e r21e0 /bin/ls
> 
> However, this results in a <not supported> message. This happened for
> all events that have a unit mask that I tried. Other events worked
> fine.
> 
> I'm not sure whether I'm giving the right command or this is a problem
> with perf/kernel version. As it turned out, I'm running Ubuntu with
> kernel 3.11.10-03111002, but the latest available perf version from
> the repositories is 3.8.0-35, which is the one I'm using.
> 
> Any help is appreciated.
> 
> Thank you,
> Martin
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users"
> in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

Re: Event not supported

[Jiri Olsa]

On Wed, Feb 19, 2014 at 03:08:19PM +0100, Baptiste Lepers wrote:
> Hi,
> 
> I think that the format you are using is not correct.
> 
> To profile event 0xXXX with unit mask YY, the format is rX0000YYXX, so in your case: r1000002e0.
> 
> Also, beware that the unitmask in the documentation tells you the _bit(s)_ that need to be set.
> So :
>   From Local Node to Node 0 = r1000001e0
>   From Local Node to Node 1 = r1000002e0
>   From Local Node to Node 2 = r1000004e0
>   From Local Node to Node 3 = r1000008e0
>   From Local Node to Node 4 = r1000010e0
>   From Local Node to Node 0 & 1 & 2 = r1000007e0
>   etc.

hi,
never tried on AMD, but AFAIK it exports the event format
terms in sysfs like Intel does:

[root@krava format]# pwd
/sys/bus/event_source/devices/cpu/format
[root@krava format]# ls
any  cmask  edge  event  in_tx  in_tx_cp  inv  ldlat  offcore_rsp  pc  umask

and you can use those terms to specify your event more user
friendly that using the r* style, like:

  perf record -e 'cpu/event=0x3c,umask=.../uk' ..


jirka


> 
> 
> Cheers,
> Baptiste.
> 
> ----- Mail original -----
> > De: "Martin Ichilevici de Oliveira" <martin.i.oliveira@gmail.com>
> > À: linux-perf-users@vger.kernel.org
> > Envoyé: Mardi 18 Février 2014 17:51:46
> > Objet: Event not supported
> > 
> > Hello,
> > 
> > I'm still learning perf, so I apologize in advance if my questions are
> > too simple and not suited here. I did, however, searched the web and
> > tried to find a solution by myself before posting to this mailing
> > list.
> > 
> > From my manufacturer's manual, I chose an event I'm interested in
> > monitoring. They also list unit masks, as below:
> > 
> > UnitMask    Description
> > 7                  From Local node to Node 7
> > 6                  From Local node to Node 6
> > 5                  From Local node to Node 5
> > 4                  From Local node to Node 4
> > 3                  From Local node to Node 3
> > 2                  From Local node to Node 2
> > 1                  From Local node to Node 1
> > 0                  From Local node to Node 0
> > 
> > The event is "CPU to DRAM Requests to Target Node" (NUMA machine) and
> > it's number is 0x1E0. As I understood, I should precede the event with
> > the unit mask. So, for example, if I'm interested in unit mask 2, I
> > should type:
> > 
> > perf stat -e r21e0 /bin/ls
> > 
> > However, this results in a <not supported> message. This happened for
> > all events that have a unit mask that I tried. Other events worked
> > fine.
> > 
> > I'm not sure whether I'm giving the right command or this is a problem
> > with perf/kernel version. As it turned out, I'm running Ubuntu with
> > kernel 3.11.10-03111002, but the latest available perf version from
> > the repositories is 3.8.0-35, which is the one I'm using.
> > 
> > Any help is appreciated.
> > 
> > Thank you,
> > Martin
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-perf-users"
> > in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html