Model code (C code) profiling with Score-P

[ikitayama]

I've collected NEURON performance profile with Score-P:

Estimated aggregate size of event trace:                   66 bytes
Estimated requirements for largest trace buffer (max_buf): 66 bytes
Estimated memory requirements (SCOREP_TOTAL_MEMORY):       4097kB
(hint: When tracing set SCOREP_TOTAL_MEMORY=4097kB to avoid intermediate flushes
 or reduce requirements using USR regions filters.)

flt     type max_buf[B] visits time[s] time[%] time/visit[us]  region
         ALL         65      2    0.97   100.0      487146.23  ALL
      SCOREP         41      1    0.97    99.8      971972.20  SCOREP
         USR         24      1    0.00     0.2        2320.25  USR

      SCOREP         41      1    0.97    99.8      971972.20  python3.8
         USR         24      1    0.00     0.2        2320.25  modl_reg

I was actually hoping I get more than modl_reg as I believe model C codes are instrumented with the compiler automation. Does this mean from the Python interpreter's view, during the simulation only modl_reg API does matter?

[ikitayama]

I understand there's LLNL/Caliper which is supported NEURON, but as a long time Score-P user, I'd like to carry out the performance evaluation with Score-P.

[pramodk]

@ikitayama : In general, would be helpful if you post detailed build commands and score-p options you used.

I haven't used score-p recently with NEURON but I would recommend to us sampling method. Could you try and see what you get?

[ikitayama]

If Score-P is loaded, in terms of Spack, then given that there's a wrapper like scorep-icc etc., one would just set CC=scorep-icc and CXX=scorep-icpc and type nrnivmodl mod

[ikitayama]

... and I was using Python bindings for Score-P: https://github.com/score-p/scorep_binding_python; so once that's Python module installed, we just issue python -m scorep script.py.

[ikitayama]

For my research, model needs to run 10 minutes of biological time or longer, model's performance is critical.

[pramodk]

If Score-P is loaded, in terms of Spack, then given that there's a wrapper like scorep-icc etc., one would just set CC=scorep-icc and CXX=scorep-icpc and type nrnivmodl mod

Yes, thats true. But, in the past I have seen that the compiler instrumentation disable vectorisation in coreneuron and result into very bad performance. I should re-look into this.

... and I was using Python bindings for Score-P: https://github.com/score-p/scorep_binding_python; so once that's Python module installed, we just issue python -m scorep script.py.

I wouldn't go for python bindings. When we use nrnivmodl then it createsspecial binary under x86_64 directory. And we can launch python script as:

mpiexec -n X ./x86_64/special -mpi -python script.py <args> 

So what we need to do is to make sure special is compiled correctly with score-p flags/instrumentation and use standard score-p approach to profile C/C++ application.