Perf tests for hw events
Folks,
In the process of writing perf support for Intel's Cache QoS Monitoring feature [0] I've had to write my own userland tests to drive tools/perf and indirectly the kernel internals. I'm now getting requests for these tests from various people and it occurs to me that they should probably live in the kernel tree.
The tests I've got do a couple of things like setting up a perf_event cgroup and assigning enough tasks to trigger the RMID recycling code in the CQM driver, ensuring that we can run multiple events simultaneously (that the event scheduling/rotation code works), etc.
Does anything like this already exist for hw events? I couldn't find anything specific to hw events from snooping around in tools/perf/tests.
I propose we add some hw event tests to the kernel tree. These will provide,
- regression tests
- a source of documentation for how to use the events
We should only test those hw events that are present on a user's machine; there's no sense in emulating things.
Thoughts?
[0] - https://lkml.kernel.org/r/[email protected]
Signature · 1 line
Matt Fleming, Intel Open Source Technology Center
Quoted 8 lines · Matt Fleming
Folks,
In the process of writing perf support for Intel's Cache QoS Monitoring feature [0] I've had to write my own userland tests to drive
tools/perfand indirectly the kernel internals. I'm now getting requests for these tests from various people and it occurs to me that they should probably live in the kernel tree.
Quoted 7 lines
The tests I've got do a couple of things like setting up a
perf_eventcgroup and assigning enough tasks to trigger the RMID recycling code in the CQM driver, ensuring that we can run multiple events simultaneously (that the event scheduling/rotation code works), etc.Does anything like this already exist for hw events? I couldn't find anything specific to hw events from snooping around in
tools/perf/tests.
that's the only place for this kind of stuff.. and AFAIK we use just some basic HW events like cycles, or software events
Quoted 6 lines
I propose we add some hw event tests to the kernel tree. These will provide,
- regression tests
- a source of documentation for how to use the events
yes, yes and yes ;-)
Quoted 2 lines
We should only test those hw events that are present on a user's machine; there's no sense in emulating things.
agreed.. I remember we were discussing with Arnaldo the strengthening of the test framework.. so it would detect what tests are runable under current machine and priviledge
IIRC we already have some CPU detection support:
int __attribute__ ((weak)) get_cpuid(char *buffer, size_t sz);
jirka
Quoted 7 lines
Thoughts?
[0] - https://lkml.kernel.org/r/[email protected]
Signature · 1 line
Matt Fleming, Intel Open Source Technology Center
Em Thu, Jan 22, 2015 at 12:40:14PM +0000, Matt Fleming escreveu:
Quoted 13 lines
Folks,
In the process of writing perf support for Intel's Cache QoS Monitoring feature [0] I've had to write my own userland tests to drive
tools/perfand indirectly the kernel internals. I'm now getting requests for these tests from various people and it occurs to me that they should probably live in the kernel tree.The tests I've got do a couple of things like setting up a
perf_eventcgroup and assigning enough tasks to trigger the RMID recycling code in the CQM driver, ensuring that we can run multiple events simultaneously (that the event scheduling/rotation code works), etc.
Quoted 8 lines
Does anything like this already exist for hw events? I couldn't find anything specific to hw events from snooping around in
tools/perf/tests.I propose we add some hw event tests to the kernel tree. These will provide,
- regression tests
- a source of documentation for how to use the events
That is the description for tools/perf/tests/ please send your patches for adding new entries there.
Quoted 2 lines
We should only test those hw events that are present on a user's machine; there's no sense in emulating things.
At some point 'perf test' should grow infrastructure to specify what is required for a test so that it auto-skips those, possibly not even bothering the user telling something can't be tested. Right now for things like tests that requires finding a vmlinux, if it doesn't find it, it will just print "Skipped", etc.
Quoted 3 lines
Thoughts?
- Arnaldo
Quoted 14 lines · Jiri Olsa
Quoted 13 lines · Matt Fleming
Folks,
In the process of writing perf support for Intel's Cache QoS Monitoring feature [0] I've had to write my own userland tests to drive
tools/perfand indirectly the kernel internals. I'm now getting requests for these tests from various people and it occurs to me that they should probably live in the kernel tree.The tests I've got do a couple of things like setting up a
perf_eventcgroup and assigning enough tasks to trigger the RMID recycling code in the CQM driver, ensuring that we can run multiple events simultaneously (that the event scheduling/rotation code works), etc.
Quoted 5 lines
Quoted 2 lines
Does anything like this already exist for hw events? I couldn't find anything specific to hw events from snooping around in
tools/perf/tests.that's the only place for this kind of stuff.. and AFAIK we use just some basic HW events like cycles, or software events
OK cool. I just wanted confirmation that tools/perf/tests was the best place for these tests.
Quoted 9 lines
Quoted 2 lines
We should only test those hw events that are present on a user's machine; there's no sense in emulating things.
agreed.. I remember we were discussing with Arnaldo the strengthening of the test framework.. so it would detect what tests are runable under current machine and priviledge
IIRC we already have some CPU detection support:
int__attribute__((weak))get_cpuid(char *buffer, size_t sz);
Aha, excellent. I hadn't noticed this. It should be possible to build something on top of this.
Signature · 1 line
Matt Fleming, Intel Open Source Technology Center
Quoted 6 lines · Arnaldo Carvalho de Melo
At some point 'perf test' should grow infrastructure to specify what is required for a test so that it auto-skips those, possibly not even bothering the user telling something can't be tested. Right now for things like tests that requires finding a vmlinux, if it doesn't find it, it will just print "Skipped", etc.
Yeah, that makes sense. I'll take a look at that.
Signature · 1 line
Matt Fleming, Intel Open Source Technology Center