IceLake TMA group gives NaN

[Nichols Romero]

Greetings LIKWID Developers,

I think this is bug or possibly something is not set-up correctly on my IceLake machine?

Here is what I see with the TMA group, they are all NaN

 (my apologies for the difficult to read copy-paste, it was the best I could do ):

+-----------------------+---------+---------------+---------------+---------------+---------------+---------------+---------------+---------------+---------------+
| Event | Counter | sal-dev04:0:0 | sal-dev04:1:1 | sal-dev04:2:2 | sal-dev04:3:3 | sal-dev04:4:4 | sal-dev04:5:5 | sal-dev04:6:6 | sal-dev04:7:7 |
+-----------------------+---------+---------------+---------------+---------------+---------------+---------------+---------------+---------------+---------------+
| Region calls | CTR | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
| INSTR_RETIRED_ANY | FIXC0 | 181638700000 | 209807700000 | 221863600000 | 221415400000 | 212531000000 | 227552600000 | 225201200000 | 175213400000 |
| CPU_CLK_UNHALTED_CORE | FIXC1 | 316011600000 | 316176100000 | 316032800000 | 316270200000 | 318650500000 | 316399000000 | 316249200000 | 316138200000 |
| CPU_CLK_UNHALTED_REF | FIXC2 | 235995600000 | 236032200000 | 236034400000 | 236118200000 | 236898400000 | 236157400000 | 236117800000 | 236033800000 |
| TOPDOWN_SLOTS | FIXC3 | 1579801000000 | 1580869000000 | 1580155000000 | 1581296000000 | 1592434000000 | 1581842000000 | 1581018000000 | 1578843000000 |
| RETIRING | TMA0 | nan | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| BAD_SPECULATION | TMA1 | nan | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| FRONTEND_BOUND | TMA2 | nan | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| BACKEND_BOUND | TMA3 | nan | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+-----------------------+---------+---------------+---------------+---------------+---------------+---------------+---------------+-_-------------+---------------+

[Nichols Romero]

I know that Thomas Gruber is on vacation.

Let me ask the question to the LIKWID user community different. Has anyone had success with the TMA group on IceLake? If so, can you share which version of Linux, the Linux kernel version, and the output of `perf -v`?

Thanks,

[Thomas Gruber]

Hi,

while I don't assume that the values are all correct, it works on our ICX nodes. Or do you actually mean ICL (desktop class chip)?

It seems you are using likwid-mpirun which might have a problem with parsing the results of the individual processes (float and not integer), so maybe try with likwid-perfctr first. If it works, the problem is only in likwid-mpirun and thus probably is parsing error.

Best,
Thomas

[Nichols Romero]

Hi Thomas,

So likwid-perfctr give all zeros for the TMA group. 

It is ICX:

n.a.romero@sal-dev04:~$ likwid-perfctr -i
--------------------------------------------------------------------------------
CPU name:       Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz
CPU type:       Intel Icelake SP processor
CPU clock:      2.60 GHz
CPU family:     6
CPU model:      106
CPU short:      ICX
CPU stepping:   6
CPU features:   FP ACPI MMX SSE SSE2 HTT TM RDTSCP MONITOR VMX EIST TM2 SSSE FMA SSE4.1 SSE4.2 AES AVX RDRAND AVX2 AVX512 RDSEED SSE3
CPU arch:       x86_64

Could it be a Linux kernel issue?

I have also been playing around with this tool https://github.com/aayasin/perf-tools

and it seems to work. But when I installed the tool, it auto-installed the latest version of perf-tools while LIKWID is use using the system installation of perftools.

The other performance groups seem to be OK, just the TMA doesn't seem to be working.

[Thomas Gruber]

With perf_event, I don't know. I just checked and at least the perf_event backend contains logic for getting the TMA data.