[Starpu-devel] Performance metric

[ASD D <ajitsdeshpande@gmail.com>]

Hello,
I have few queries related to measuring performance of a application udner Starpu. (I read Chapter 6 of the Starpu Handbook - Performance Feedback but need some mroe data)

1] Trying to get some performance metrics from running code under Starpu. Although I want to do it for my own example, but let us for simplicity assume the basic_examples/vector_scal.c example

   How do I get performance metrics (time spent/cycles consumed) in each worker(CPU,CUDA,OPENCL)?

I have tried enabling the fxt trace library in the ./configure , as a result execution of my code, of it I do get a trace file  /tmp/
prof_file_XXX_YYY, then I was able to obtain paje.trace. which i viewed using Vite.
But part from that I am looking at some numbers (time in seconds spent by worker).

2] Then, I see following files(performance model files) -  .starpu/sampling/codelets/vector_scale.mymachine and vector_scale_power.mymachine, generated in my folder after execution of the app under starpu - 
  having content as below. What is it the data in it mean?  Is it any performance data? 

##################
# CPUs
# number of CPU architectures

0
##################
# CUDAs

# number of CUDA architectures
1
###########

# CUDA_0
# number of implementations
1

# Model for cuda_0_impl_0
# number of entries
1

# sumlnx        sumlnx2         sumlny          sumlnxlny       alpha           beta            n
0.000000e+00    0.000000e+00    0.000000e+00    0.000000e+00    nan             nan             0

# a             b               c
nan             nan             nan

# hash          size            mean            dev             sum             sum2            n
870a30aa        8192            4.338900e+01    0.000000e+00    4.338900e+01    1.882605e+03    1

##################
##################

# OPENCLs
# number of OPENCL architectures
0

##################
# GORDONs
# number of GORDON architectures

0

thank you.
-Ajit