Here's a little test program that checks whether software counters
(specifically, the task clock counter) work correctly when they're in
a group with hardware counters.
What it does is to create several groups, each with one hardware
counter, counting instructions, plus a task clock counter. It needs
to know an upper bound N on the number of hardware counters you have
(N defaults to 8), and it creates N+4 groups to force them to be
multiplexed. It also creates an overall task clock counter.
Then it spins for a while, and then stops all the counters and reads
them. It takes the total of the task clock counters in the groups and
computes the ratio of that total to the overall execution time from
the overall task clock counter.
That ratio should be equal to the number of actual hardware counters
that can count instructions. If the task clock counters in the groups
don't stop when their group gets taken off the PMU, the ratio will
instead be close to N+4. The program will declare that the test fails
if the ratio is greater than N (actually, N + 0.0001 to allow for FP
rounding errors).
Could someone run this on x86 on the latest PCL tree and let me know
what happens? I don't have an x86 crash box easily to hand. On
powerpc, it passes, but I think that is because I am missing setting
counter->prev_count in arch/powerpc/kernel/perf_counter.c, and I think
that means that enabling/disabling a group with a task clock counter
in it won't work correctly (I'll do a test program for that next).
Usage is: swsched-test [-c num-hw-counters] [-v]
Use -c N if you have more than 8 hardware counters. The -v flag makes
it print out the values of each counter.
Paul.
On Wed, 2009-03-18 at 10:27 +1100, Paul Mackerras wrote:
> Here's a little test program that checks whether software counters
> (specifically, the task clock counter) work correctly when they're in
> a group with hardware counters.
>
> What it does is to create several groups, each with one hardware
> counter, counting instructions, plus a task clock counter. It needs
> to know an upper bound N on the number of hardware counters you have
> (N defaults to 8), and it creates N+4 groups to force them to be
> multiplexed. It also creates an overall task clock counter.
>
> Then it spins for a while, and then stops all the counters and reads
> them. It takes the total of the task clock counters in the groups and
> computes the ratio of that total to the overall execution time from
> the overall task clock counter.
>
> That ratio should be equal to the number of actual hardware counters
> that can count instructions. If the task clock counters in the groups
> don't stop when their group gets taken off the PMU, the ratio will
> instead be close to N+4. The program will declare that the test fails
> if the ratio is greater than N (actually, N + 0.0001 to allow for FP
> rounding errors).
>
> Could someone run this on x86 on the latest PCL tree and let me know
> what happens? I don't have an x86 crash box easily to hand. On
> powerpc, it passes, but I think that is because I am missing setting
> counter->prev_count in arch/powerpc/kernel/perf_counter.c, and I think
> that means that enabling/disabling a group with a task clock counter
> in it won't work correctly (I'll do a test program for that next).
>
> Usage is: swsched-test [-c num-hw-counters] [-v]
>
> Use -c N if you have more than 8 hardware counters. The -v flag makes
> it print out the values of each counter.
I'll give it a run when I'm awake again, with the patches I send out --
esp the last one, x86 might stand a chance of actually working.
On Wed, 2009-03-18 at 10:27 +1100, Paul Mackerras wrote:
> Here's a little test program that checks whether software counters
> (specifically, the task clock counter) work correctly when they're in
> a group with hardware counters.
>
> What it does is to create several groups, each with one hardware
> counter, counting instructions, plus a task clock counter. It needs
> to know an upper bound N on the number of hardware counters you have
> (N defaults to 8), and it creates N+4 groups to force them to be
> multiplexed. It also creates an overall task clock counter.
>
> Then it spins for a while, and then stops all the counters and reads
> them. It takes the total of the task clock counters in the groups and
> computes the ratio of that total to the overall execution time from
> the overall task clock counter.
>
> That ratio should be equal to the number of actual hardware counters
> that can count instructions. If the task clock counters in the groups
> don't stop when their group gets taken off the PMU, the ratio will
> instead be close to N+4. The program will declare that the test fails
> if the ratio is greater than N (actually, N + 0.0001 to allow for FP
> rounding errors).
>
> Could someone run this on x86 on the latest PCL tree and let me know
> what happens? I don't have an x86 crash box easily to hand. On
> powerpc, it passes, but I think that is because I am missing setting
> counter->prev_count in arch/powerpc/kernel/perf_counter.c, and I think
> that means that enabling/disabling a group with a task clock counter
> in it won't work correctly (I'll do a test program for that next).
>
> Usage is: swsched-test [-c num-hw-counters] [-v]
>
> Use -c N if you have more than 8 hardware counters. The -v flag makes
> it print out the values of each counter.
Adapted the thing a little for the latest ABI, but it seems to work:
[root@opteron ~]# ./swsched-test
overall task clock: 3211403572
hw sum: 12150279972, task clock sum: 12845614288
ratio: 4.00
test passed