refactor and repair the spack environment

Reviewed-on: portelli/lattice-benchmarks#2

Grid/Benchmark_Grid.cpp

@@ -73,6 +73,8 @@ class Benchmark
         {local[0] * mpi[0], local[1] * mpi[1], local[2] * mpi[2], local[3] * mpi[3]});
     GridCartesian *TmpGrid = SpaceTimeGrid::makeFourDimGrid(
         latt4, GridDefaultSimd(Nd, vComplex::Nsimd()), GridDefaultMpi());
+    Grid::Coordinate shm;
+    GlobalSharedMemory::GetShmDims(mpi, shm);
 
     uint64_t NP = TmpGrid->RankCount();
     uint64_t NN = TmpGrid->NodeCount();
@@ -85,7 +87,9 @@ class Benchmark
     std::cout << GridLogMessage << "* OpenMP threads : " << GridThread::GetThreads()
               << std::endl;
 
-    std::cout << GridLogMessage << "* MPI tasks      : " << GridCmdVectorIntToString(mpi)
+    std::cout << GridLogMessage << "* MPI layout     : " << GridCmdVectorIntToString(mpi)
+              << std::endl;
+    std::cout << GridLogMessage << "* Shm layout     : " << GridCmdVectorIntToString(shm)
               << std::endl;
 
     std::cout << GridLogMessage << "* vReal          : " << sizeof(vReal) * 8 << "bits ; "
@@ -118,6 +122,7 @@ class Benchmark
     for (unsigned int i = 0; i < mpi.size(); ++i)
     {
       tmp["mpi"].push_back(mpi[i]);
+      tmp["shm"].push_back(shm[i]);
     }
     tmp["ranks"] = NP;
     tmp["nodes"] = NN;
@@ -132,6 +137,8 @@ class Benchmark
 
     Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
     Coordinate mpi_layout = GridDefaultMpi();
+    Coordinate shm_layout;
+    GlobalSharedMemory::GetShmDims(mpi_layout, shm_layout);
 
     for (int mu = 0; mu < Nd; mu++)
       if (mpi_layout[mu] > 1)
@@ -143,8 +150,8 @@ class Benchmark
     std::cout << GridLogMessage << "Benchmarking threaded STENCIL halo exchange in "
               << nmu << " dimensions" << std::endl;
     grid_small_sep();
-    grid_printf("%5s %5s %15s %15s %15s %15s %15s\n", "L", "dir", "payload (B)",
-                "time (usec)", "rate (GB/s/node)", "std dev", "max");
+    grid_printf("%5s %5s %7s %15s %15s %15s %15s %15s\n", "L", "dir", "shm",
+                "payload (B)", "time (usec)", "rate (GB/s/node)", "std dev", "max");
 
     for (int lat = 16; lat <= maxlat; lat += 8)
     {
@@ -173,74 +180,80 @@ class Benchmark
       for (int dir = 0; dir < 8; dir++)
       {
         int mu = dir % 4;
-        if (mpi_layout[mu] > 1)
+        if (mpi_layout[mu] == 1) // skip directions that are not distributed
+          continue;
+        bool is_shm = mpi_layout[mu] == shm_layout[mu];
+        bool is_partial_shm = !is_shm && shm_layout[mu] != 1;
+
+        std::vector<double> times(Nloop);
+        for (int i = 0; i < NWARMUP; i++)
+        {
+          int xmit_to_rank;
+          int recv_from_rank;
+
+          if (dir == mu)
+          {
+            int comm_proc = 1;
+            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
+          }
+          else
+          {
+            int comm_proc = mpi_layout[mu] - 1;
+            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
+          }
+          Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank, (void *)&rbuf[dir][0],
+                              recv_from_rank, bytes);
+        }
+        for (int i = 0; i < Nloop; i++)
         {
 
-          std::vector<double> times(Nloop);
-          for (int i = 0; i < NWARMUP; i++)
+          dbytes = 0;
+          double start = usecond();
+          int xmit_to_rank;
+          int recv_from_rank;
+
+          if (dir == mu)
           {
-            int xmit_to_rank;
-            int recv_from_rank;
-
-            if (dir == mu)
-            {
-              int comm_proc = 1;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-            }
-            else
-            {
-              int comm_proc = mpi_layout[mu] - 1;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-            }
-            Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
-                                (void *)&rbuf[dir][0], recv_from_rank, bytes);
+            int comm_proc = 1;
+            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
           }
-          for (int i = 0; i < Nloop; i++)
+          else
           {
-
-            dbytes = 0;
-            double start = usecond();
-            int xmit_to_rank;
-            int recv_from_rank;
-
-            if (dir == mu)
-            {
-              int comm_proc = 1;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-            }
-            else
-            {
-              int comm_proc = mpi_layout[mu] - 1;
-              Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
-            }
-            Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
-                                (void *)&rbuf[dir][0], recv_from_rank, bytes);
-            dbytes += bytes;
-
-            double stop = usecond();
-            t_time[i] = stop - start; // microseconds
+            int comm_proc = mpi_layout[mu] - 1;
+            Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
           }
-          timestat.statistics(t_time);
+          Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank, (void *)&rbuf[dir][0],
+                              recv_from_rank, bytes);
+          dbytes += bytes;
 
-          dbytes = dbytes * ppn;
-          double bidibytes = 2. * dbytes;
-          double rate = bidibytes / (timestat.mean / 1.e6) / 1024. / 1024. / 1024.;
-          double rate_err = rate * timestat.err / timestat.mean;
-          double rate_max = rate * timestat.mean / timestat.min;
-          grid_printf("%5d %5d %15d %15.2f %15.2f %15.1f %15.2f\n", lat, dir, bytes,
-                      timestat.mean, rate, rate_err, rate_max);
-          nlohmann::json tmp;
-          nlohmann::json tmp_rate;
-          tmp["L"] = lat;
-          tmp["dir"] = dir;
-          tmp["bytes"] = bytes;
-          tmp["time_usec"] = timestat.mean;
-          tmp_rate["mean"] = rate;
-          tmp_rate["error"] = rate_err;
-          tmp_rate["max"] = rate_max;
-          tmp["rate_GBps"] = tmp_rate;
-          json_results["comms"].push_back(tmp);
+          double stop = usecond();
+          t_time[i] = stop - start; // microseconds
         }
+        timestat.statistics(t_time);
+
+        dbytes = dbytes * ppn;
+        double bidibytes = 2. * dbytes;
+        double rate = bidibytes / (timestat.mean / 1.e6) / 1024. / 1024. / 1024.;
+        double rate_err = rate * timestat.err / timestat.mean;
+        double rate_max = rate * timestat.mean / timestat.min;
+        grid_printf("%5d %5d %7s %15d %15.2f %15.2f %15.1f %15.2f\n", lat, dir,
+                    is_shm           ? "yes"
+                    : is_partial_shm ? "partial"
+                                     : "no",
+                    bytes, timestat.mean, rate, rate_err, rate_max);
+        nlohmann::json tmp;
+        nlohmann::json tmp_rate;
+        tmp["L"] = lat;
+        tmp["dir"] = dir;
+        tmp["shared_mem"] = is_shm;
+        tmp["partial_shared_mem"] = is_partial_shm;
+        tmp["bytes"] = bytes;
+        tmp["time_usec"] = timestat.mean;
+        tmp_rate["mean"] = rate;
+        tmp_rate["error"] = rate_err;
+        tmp_rate["max"] = rate_max;
+        tmp["rate_GBps"] = tmp_rate;
+        json_results["comms"].push_back(tmp);
       }
       for (int d = 0; d < 8; d++)
       {

Grid/systems/tursa/spack-bootstrap.sh

@@ -4,7 +4,13 @@ set -euo pipefail
 
 gcc_spec='gcc@9.4.0'
 cuda_spec='cuda@11.4.0'
-hdf5_spec='hdf5@1.10.7'
+
+# hdf5 and fftw depend on OpenMPI, which we install manually. To make sure this
+# dependency is picked by spack, we specify the compiler here explicitly. For
+# most other packages we dont really care about the compiler (i.e. system
+# compiler versus ${gcc_spec})
+hdf5_spec="hdf5@1.10.7+cxx+threadsafe%${gcc_spec}"
+fftw_spec="fftw%${gcc_spec}"
 
 if (( $# != 1 )); then
     echo "usage: $(basename "$0") <env dir>" 1>&2
@@ -18,7 +24,7 @@ cd "${cwd}"
 
 # General configuration ########################################################
 # build with 128 tasks
-echo 'config:                                  
+echo 'config:
   build_jobs: 128
   build_stage:
     - $spack/var/spack/stage
@@ -38,26 +44,23 @@ rm external.yaml
 
 # Base compilers ###############################################################
 # configure system base
+
+spack env create base
+spack env activate base
 spack compiler find --scope site
 
-# install GCC, CUDA & LLVM
-spack install ${gcc_spec} ${cuda_spec} llvm
-
-spack load llvm
+# install GCC, CUDA
+spack add ${gcc_spec} ${cuda_spec}
+spack concretize
+spack env depfile -o Makefile.tmp
+make -j128 -f Makefile.tmp
 spack compiler find --scope site
-spack unload llvm
-
-spack load ${gcc_spec}
-spack compiler find --scope site
-spack unload ${gcc_spec}
 
 # Manual compilation of OpenMPI & UCX ##########################################
 # set build directories
 mkdir -p "${dir}"/build
 cd "${dir}"/build
 
-spack load ${gcc_spec} ${cuda_spec}
-
 cuda_path=$(spack find --format "{prefix}" cuda)
 gdrcopy_path=/mnt/lustre/tursafs1/apps/gdrcopy/2.3.1
 
@@ -124,8 +127,8 @@ mkdir build_gpu; cd build_gpu
              --with-cuda="${cuda_path}" --disable-getpwuid        \
              --with-verbs --with-slurm --enable-mpi-fortran=all   \
              --with-pmix=internal --with-libevent=internal
-make -j 128 
-make install 
+make -j 128
+make install
 cd ..
 
 # openmpi cpu build
@@ -141,60 +144,62 @@ make -j 128
 make install
 cd "${dir}"
 
+ucx_spec_gpu="ucx@1.12.0.GPU%${gcc_spec}"
+ucx_spec_cpu="ucx@1.12.0.CPU%${gcc_spec}"
+openmpi_spec_gpu="openmpi@4.1.1.GPU%${gcc_spec}"
+openmpi_spec_cpu="openmpi@4.1.1.CPU%${gcc_spec}"
+
 # Add externals to spack
 echo "packages:
   ucx:
     externals:
-    - spec: \"ucx@1.12.0.GPU%gcc@9.4.0\"
+    - spec: \"${ucx_spec_gpu}\"
       prefix: ${dir}/prefix/ucx_gpu
-    - spec: \"ucx@1.12.0.CPU%gcc@9.4.0\"
+    - spec: \"${ucx_spec_cpu}\"
       prefix: ${dir}/prefix/ucx_cpu
     buildable: False
   openmpi:
     externals:
-    - spec: \"openmpi@4.1.1.GPU%gcc@9.4.0\"
+    - spec: \"${openmpi_spec_gpu}\"
       prefix: ${dir}/prefix/ompi_gpu
-    - spec: \"openmpi@4.1.1.CPU%gcc@9.4.0\"
+    - spec: \"${openmpi_spec_cpu}\"
       prefix: ${dir}/prefix/ompi_cpu
     buildable: False" > spack.yaml
 
 spack config --scope site add -f spack.yaml
 rm spack.yaml
-spack install ucx@1.12.0.GPU%gcc@9.4.0 openmpi@4.1.1.GPU%gcc@9.4.0
-spack install ucx@1.12.0.CPU%gcc@9.4.0 openmpi@4.1.1.CPU%gcc@9.4.0
+spack env deactivate
 
 cd "${cwd}"
 
 # environments #################################################################
 dev_tools=("autoconf" "automake" "libtool" "jq" "git")
-ompi_gpu_hash=$(spack find --format "{hash}" openmpi@4.1.1.GPU)
-ompi_cpu_hash=$(spack find --format "{hash}" openmpi@4.1.1.CPU)
 
 spack env create grid-gpu
 spack env activate grid-gpu
-spack add ${gcc_spec} ${cuda_spec} "${dev_tools[@]}" 
-spack add ucx@1.12.0.GPU%gcc@9.4.0 openmpi@4.1.1.GPU%gcc@9.4.0
-spack add ${hdf5_spec}+cxx+threadsafe ^/"${ompi_gpu_hash}"
-spack add fftw ^/"${ompi_gpu_hash}"
-spack add openssl gmp mpfr c-lime
-spack install
+spack compiler find --scope site
+spack add ${gcc_spec} ${cuda_spec} ${ucx_spec_gpu} ${openmpi_spec_gpu}
+spack add ${hdf5_spec} ${fftw_spec}
+spack add openssl gmp mpfr c-lime "${dev_tools[@]}"
+spack concretize
+spack env depfile -o Makefile.tmp
+make -j128 -f Makefile.tmp
 spack env deactivate
 
 spack env create grid-cpu
 spack env activate grid-cpu
-spack add llvm "${dev_tools[@]}" 
-spack add ucx@1.12.0.CPU%gcc@9.4.0 openmpi@4.1.1.CPU%gcc@9.4.0
-spack add ${hdf5_spec}+cxx+threadsafe ^/"${ompi_cpu_hash}"
-spack add fftw ^/"${ompi_cpu_hash}"
-spack add openssl gmp mpfr c-lime
-spack install
+spack compiler find --scope site
+spack add ${gcc_spec} ${ucx_spec_cpu} ${openmpi_spec_cpu}
+spack add ${hdf5_spec} ${fftw_spec}
+spack add openssl gmp mpfr c-lime "${dev_tools[@]}"
+spack concretize
+spack env depfile -o Makefile.tmp
+make -j128 -f Makefile.tmp
 spack env deactivate
 
-spack install jq git
-
 # Final setup ##################################################################
 spack clean
-spack gc -y
+#spack gc -y  # "spack gc" tends to get hung up for unknown reasons
 
 # add more environment variables in module loading
 spack config --scope site add 'modules:prefix_inspections:lib:[LD_LIBRARY_PATH,LIBRARY_PATH]'