<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
  </head>
  <body>
    <p>Hei,</p>
    <p>that was the reason for increased run times. When removing
      #pragma omp parallel for, my loop took ~18 seconds. When changing
      it to #pragma omp parallel for num_threads(2) or #pragma omp
      parallel for num_threads(4) (on a i7-6700), the loop took ~16 s,
      but when increasing it to #pragma omp parallel for num_threads(8),
      the loop took 28 s.</p>
    <p>Regards,</p>
    <p>Roland<br>
    </p>
    <div class="moz-cite-prefix">Am 17.02.21 um 18:51 schrieb Matthew
      Knepley:<br>
    </div>
    <blockquote type="cite"
cite="mid:CAMYG4Gmej=YDgdPPb8P60-oViESJO32=VE6iAobG4qHuy7=L8g@mail.gmail.com">
      <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
      <div dir="ltr">Jed, is it possible that this is an
        oversubscription penalty from bad OpenMP settings? <said by a
        person who knows less about OpenMP than cuneiform>
        <div><br>
        </div>
        <div>  Thanks,</div>
        <div><br>
        </div>
        <div>     Matt</div>
      </div>
      <br>
      <div class="gmail_quote">
        <div dir="ltr" class="gmail_attr">On Wed, Feb 17, 2021 at 12:11
          PM Roland Richter <<a href="mailto:roland.richter@ntnu.no"
            moz-do-not-send="true">roland.richter@ntnu.no</a>> wrote:<br>
        </div>
        <blockquote class="gmail_quote" style="margin:0px 0px 0px
          0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
          <div>
            <div>My PetscScalar is complex double (i.e. even higher
              penalty), but my matrix has a size of 8kk elements, so
              that should not an issue.<br>
              Regards,<br>
              Roland
              <hr style="display:inline-block;width:98%">
              <div id="gmail-m_2709721564415767467x_divRplyFwdMsg"
                dir="ltr"><font style="font-size:11pt" face="Calibri,
                  sans-serif" color="#000000"><b>Von:</b> Jed Brown <<a
                    href="mailto:jed@jedbrown.org" target="_blank"
                    moz-do-not-send="true">jed@jedbrown.org</a>><br>
                  <b>Gesendet:</b> Mittwoch, 17. Februar 2021 17:49:49<br>
                  <b>An:</b> Roland Richter; PETSc<br>
                  <b>Betreff:</b> Re: [petsc-users] Explicit linking to
                  OpenMP results in performance drop and wrong results</font>
                <div> </div>
              </div>
            </div>
            <font size="2"><span style="font-size:10pt">
                <div>Roland Richter <<a
                    href="mailto:roland.richter@ntnu.no" target="_blank"
                    moz-do-not-send="true">roland.richter@ntnu.no</a>>
                  writes:<br>
                  <br>
                  > Hei,<br>
                  ><br>
                  > I replaced the linking line with<br>
                  ><br>
                  > //usr/lib64/mpi/gcc/openmpi3/bin/mpicxx 
                  -march=native -fopenmp-simd<br>
                  > -DMKL_LP64 -m64<br>
                  >
                  CMakeFiles/armadillo_with_PETSc.dir/Unity/unity_0_cxx.cxx.o
                  -o<br>
                  > bin/armadillo_with_PETSc <br>
                  >
                  -Wl,-rpath,/opt/boost/lib:/opt/fftw3/lib64:/opt/petsc_release/lib<br>
                  > /usr/lib64/libgsl.so /usr/lib64/libgslcblas.so
                  -lgfortran <br>
                  > -L${MKLROOT}/lib/intel64 -Wl,--no-as-needed
                  -lmkl_intel_lp64<br>
                  > -lmkl_gnu_thread -lmkl_core -lgomp -lpthread -lm
                  -ldl<br>
                  > /opt/boost/lib/libboost_filesystem.so.1.72.0<br>
                  > /opt/boost/lib/libboost_mpi.so.1.72.0<br>
                  > /opt/boost/lib/libboost_program_options.so.1.72.0<br>
                  > /opt/boost/lib/libboost_serialization.so.1.72.0<br>
                  > /opt/fftw3/lib64/libfftw3.so
                  /opt/fftw3/lib64/libfftw3_mpi.so<br>
                  > /opt/petsc_release/lib/libpetsc.so<br>
                  > /usr/lib64/gcc/x86_64-suse-linux/9/libgomp.so<br>
                  > /<br>
                  ><br>
                  > and now the results are correct. Nevertheless,
                  when comparing the loop<br>
                  > in line 26-28 in file test_scaling.cpp<br>
                  ><br>
                  > /#pragma omp parallel for//<br>
                  > //    for(int i = 0; i < r_0 * r_1; ++i)//<br>
                  > //        *(out_mat_ptr + i) = (*(in_mat_ptr + i)
                  * scaling_factor);/<br>
                  ><br>
                  > the version without /#pragma omp parallel/ for is
                  significantly faster<br>
                  > (i.e. 18 s vs 28 s) compared to the version with
                  /omp./ Why is there<br>
                  > still such a big difference?<br>
                  <br>
                  Sounds like you're using a profile to attribute time?
                  Each `omp parallel` region incurs a cost ranging from
                  about a microsecond to 10 or more microseconds
                  depending on architecture, number of threads, and
                  OpenMP implementation. Your loop (for double
                  precision) operates at around 8 entries per clock
                  cycle (depending on architecture) if the operands are
                  in cache so the loop size r_0 * r_1 should be at least
                  10000 just to pay off the cost of `omp parallel`.<br>
                </div>
              </span></font>
          </div>
        </blockquote>
      </div>
      <br clear="all">
      <div><br>
      </div>
      -- <br>
      <div dir="ltr" class="gmail_signature">
        <div dir="ltr">
          <div>
            <div dir="ltr">
              <div>
                <div dir="ltr">
                  <div>What most experimenters take for granted before
                    they begin their experiments is infinitely more
                    interesting than any results to which their
                    experiments lead.<br>
                    -- Norbert Wiener</div>
                  <div><br>
                  </div>
                  <div><a href="http://www.cse.buffalo.edu/~knepley/"
                      target="_blank" moz-do-not-send="true">https://www.cse.buffalo.edu/~knepley/</a><br>
                  </div>
                </div>
              </div>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
  </body>
</html>