Hi guys,
I have just run into an issue with pmcstat(8) on a FreeBSD 12.1 AMD box. Basically, all the samples that happen in our userland application are counted as dubious-frames. We can see the samples in the kernel, but not the samples from our application.
Here is what we are doing for instance:
And we can directly see the issue in the output:
And if we run the application and run pmcstat in top mode (-T), we can see many unresolved samples that happen to be the samples from our application (they are growing quickly according to the load of the program).
We never saw that in our previous architectures.
Does anybody has a hint at what we could be looking?
Thanks
I have just run into an issue with pmcstat(8) on a FreeBSD 12.1 AMD box. Basically, all the samples that happen in our userland application are counted as dubious-frames. We can see the samples in the kernel, but not the samples from our application.
Here is what we are doing for instance:
Code:
pmcstat -S unhalted-core-cycles -O out.pmc /our_application
pmcstat -R out.c -z 32 -G out.stacks
Code:
CONVERSION STATISTICS:
#exec/elf 7
#samples/total 2321771
#callchain/dubious-frames 930247
We never saw that in our previous architectures.
Does anybody has a hint at what we could be looking?
Thanks