log iso-timestamp instead of seconds-since-start

Reviewed-on: portelli/lattice-benchmarks#7
Reviewed-by: Antonin Portelli <antonin.portelli@me.com>

Grid/Benchmark_Grid.cpp

@@ -1,7 +1,7 @@
 /*
 Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
 Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
-Copyright © 2022 Simon Buerger <simon.buerger@rwth-aachen.de>
+Copyright © 2024 Simon Buerger <simon.buerger@rwth-aachen.de>
 
 This is a fork of Benchmark_ITT.cpp from Grid
 
@@ -29,6 +29,43 @@ int NN_global;
 
 nlohmann::json json_results;
 
+// NOTE: Grid::GridClock is just a typedef to
+// `std::chrono::high_resolution_clock`, but `Grid::usecond` rounds to
+// microseconds (no idea why, probably wasnt ever relevant before), so we need
+// our own wrapper here.
+double usecond_precise()
+{
+  using namespace std::chrono;
+  auto nsecs = duration_cast<nanoseconds>(GridClock::now() - Grid::theProgramStart);
+  return nsecs.count() * 1e-3;
+}
+
+std::vector<std::string> get_mpi_hostnames()
+{
+  int world_size;
+  MPI_Comm_size(MPI_COMM_WORLD, &world_size);
+
+  char hostname[MPI_MAX_PROCESSOR_NAME];
+  int name_len = 0;
+  MPI_Get_processor_name(hostname, &name_len);
+
+  // Allocate buffer to gather all hostnames
+  std::vector<char> all_hostnames(world_size * MPI_MAX_PROCESSOR_NAME);
+
+  // Use MPI_Allgather to gather all hostnames on all ranks
+  MPI_Allgather(hostname, MPI_MAX_PROCESSOR_NAME, MPI_CHAR, all_hostnames.data(),
+                MPI_MAX_PROCESSOR_NAME, MPI_CHAR, MPI_COMM_WORLD);
+
+  // Convert the gathered hostnames back into a vector of std::string
+  std::vector<std::string> hostname_list(world_size);
+  for (int i = 0; i < world_size; ++i)
+  {
+    hostname_list[i] = std::string(&all_hostnames[i * MPI_MAX_PROCESSOR_NAME]);
+  }
+
+  return hostname_list;
+}
+
 struct time_statistics
 {
   double mean;
@@ -73,6 +110,8 @@ 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(4, 1);
+    GlobalSharedMemory::GetShmDims(mpi, shm);
 
     uint64_t NP = TmpGrid->RankCount();
     uint64_t NN = TmpGrid->NodeCount();
@@ -85,7 +124,9 @@ 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 +159,7 @@ 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 +174,8 @@ class Benchmark
 
     Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
     Coordinate mpi_layout = GridDefaultMpi();
+    Coordinate shm_layout(Nd, 1);
+    GlobalSharedMemory::GetShmDims(mpi_layout, shm_layout);
 
     for (int mu = 0; mu < Nd; mu++)
       if (mpi_layout[mu] > 1)
@@ -143,8 +187,8 @@ 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 +217,80 @@ 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++)
       {
@@ -251,6 +301,170 @@ class Benchmark
     return;
   }
 
+  static void Latency(void)
+  {
+    int Nwarmup = 100;
+    int Nloop = 300;
+
+    std::cout << GridLogMessage << "Benchmarking point-to-point latency" << std::endl;
+    grid_small_sep();
+    grid_printf("from to      mean(usec)           err           max\n");
+
+    int ranks;
+    int me;
+    MPI_Comm_size(MPI_COMM_WORLD, &ranks);
+    MPI_Comm_rank(MPI_COMM_WORLD, &me);
+
+    int bytes = 8;
+    void *buf_from = acceleratorAllocDevice(bytes);
+    void *buf_to = acceleratorAllocDevice(bytes);
+    nlohmann::json json_latency;
+    for (int from = 0; from < ranks; ++from)
+      for (int to = 0; to < ranks; ++to)
+      {
+        if (from == to)
+          continue;
+
+        std::vector<double> t_time(Nloop);
+        time_statistics timestat;
+        MPI_Status status;
+
+        for (int i = -Nwarmup; i < Nloop; ++i)
+        {
+          double start = usecond_precise();
+          if (from == me)
+          {
+            auto err = MPI_Send(buf_from, bytes, MPI_CHAR, to, 0, MPI_COMM_WORLD);
+            assert(err == MPI_SUCCESS);
+          }
+          if (to == me)
+          {
+            auto err =
+                MPI_Recv(buf_to, bytes, MPI_CHAR, from, 0, MPI_COMM_WORLD, &status);
+            assert(err == MPI_SUCCESS);
+          }
+          double stop = usecond_precise();
+          if (i >= 0)
+            t_time[i] = stop - start;
+        }
+        // important: only 'from' and 'to' have meaningful timings. we use
+        // 'from's.
+        MPI_Bcast(t_time.data(), Nloop, MPI_DOUBLE, from, MPI_COMM_WORLD);
+
+        timestat.statistics(t_time);
+        grid_printf("%2d %2d %15.4f %15.3f %15.4f\n", from, to, timestat.mean,
+                    timestat.err, timestat.max);
+        nlohmann::json tmp;
+        tmp["from"] = from;
+        tmp["to"] = to;
+        tmp["time_usec"] = timestat.mean;
+        tmp["time_usec_error"] = timestat.err;
+        tmp["time_usec_min"] = timestat.min;
+        tmp["time_usec_max"] = timestat.max;
+        tmp["time_usec_full"] = t_time;
+        json_latency.push_back(tmp);
+      }
+    json_results["latency"] = json_latency;
+
+    acceleratorFreeDevice(buf_from);
+    acceleratorFreeDevice(buf_to);
+  }
+
+  static void P2P(void)
+  {
+    // IMPORTANT: The P2P benchmark uses "MPI_COMM_WORLD" communicator, which is
+    // not the quite the same as Grid.communicator. Practically speaking, the
+    // latter one contains the same MPI-ranks but in a different order. Grid
+    // does this make sure it can exploit ranks with shared memory (i.e.
+    // multiple ranks on the same node) as best as possible.
+
+    // buffer-size to benchmark. This number is the same as the largest one used
+    // in the "Comms()" benchmark. ( L=48, Ls=12, double-prec-complex,
+    // half-color-spin-vector. ). Mostly an arbitrary choice, but nice to match
+    // it here
+    size_t bytes = 127401984;
+
+    int Nwarmup = 20;
+    int Nloop = 100;
+
+    std::cout << GridLogMessage << "Benchmarking point-to-point bandwidth" << std::endl;
+    grid_small_sep();
+    grid_printf("from to      mean(usec)           err           min           "
+                "bytes    rate (GiB/s)\n");
+
+    int ranks;
+    int me;
+    MPI_Comm_size(MPI_COMM_WORLD, &ranks);
+    MPI_Comm_rank(MPI_COMM_WORLD, &me);
+
+    void *buf_from = acceleratorAllocDevice(bytes);
+    void *buf_to = acceleratorAllocDevice(bytes);
+    nlohmann::json json_p2p;
+    for (int from = 0; from < ranks; ++from)
+      for (int to = 0; to < ranks; ++to)
+      {
+        if (from == to)
+          continue;
+
+        std::vector<double> t_time(Nloop);
+        time_statistics timestat;
+        MPI_Status status;
+
+        for (int i = -Nwarmup; i < Nloop; ++i)
+        {
+          double start = usecond_precise();
+          if (from == me)
+          {
+            auto err = MPI_Send(buf_from, bytes, MPI_CHAR, to, 0, MPI_COMM_WORLD);
+            assert(err == MPI_SUCCESS);
+          }
+          if (to == me)
+          {
+            auto err =
+                MPI_Recv(buf_to, bytes, MPI_CHAR, from, 0, MPI_COMM_WORLD, &status);
+            assert(err == MPI_SUCCESS);
+          }
+          double stop = usecond_precise();
+          if (i >= 0)
+            t_time[i] = stop - start;
+        }
+        // important: only 'from' and 'to' have meaningful timings. we use
+        // 'from's.
+        MPI_Bcast(t_time.data(), Nloop, MPI_DOUBLE, from, MPI_COMM_WORLD);
+
+        timestat.statistics(t_time);
+        double rate = bytes / (timestat.mean / 1.e6) / 1024. / 1024. / 1024.;
+        double rate_err = rate * timestat.err / timestat.mean;
+        double rate_max = rate * timestat.mean / timestat.min;
+        double rate_min = rate * timestat.mean / timestat.max;
+
+        grid_printf("%2d %2d %15.4f %15.3f %15.4f %15d %15.2f\n", from, to, timestat.mean,
+                    timestat.err, timestat.min, bytes, rate);
+
+        nlohmann::json tmp;
+        tmp["from"] = from;
+        tmp["to"] = to;
+        tmp["bytes"] = bytes;
+        tmp["time_usec"] = timestat.mean;
+        tmp["time_usec_error"] = timestat.err;
+        tmp["time_usec_min"] = timestat.min;
+        tmp["time_usec_max"] = timestat.max;
+        tmp["time_usec_full"] = t_time;
+        nlohmann::json tmp_rate;
+        tmp_rate["mean"] = rate;
+        tmp_rate["error"] = rate_err;
+        tmp_rate["max"] = rate_max;
+        tmp_rate["min"] = rate_min;
+        tmp["rate_GBps"] = tmp_rate;
+
+        json_p2p.push_back(tmp);
+      }
+    json_results["p2p"] = json_p2p;
+
+    acceleratorFreeDevice(buf_from);
+    acceleratorFreeDevice(buf_to);
+  }
+
   static void Memory(void)
   {
     const int Nvec = 8;
@@ -512,8 +726,6 @@ class Benchmark
 
         FGrid->Broadcast(0, &ncall, sizeof(ncall));
 
-        Dw.ZeroCounters();
-
         time_statistics timestat;
         std::vector<double> t_time(ncall);
         for (uint64_t i = 0; i < ncall; i++)
@@ -708,7 +920,6 @@ class Benchmark
         uint64_t ncall = 500;
 
         FGrid->Broadcast(0, &ncall, sizeof(ncall));
-        Ds.ZeroCounters();
 
         time_statistics timestat;
         std::vector<double> t_time(ncall);
@@ -776,11 +987,47 @@ int main(int argc, char **argv)
 {
   Grid_init(&argc, &argv);
 
+  int Ls = 1;
+  bool do_su4 = true;
+  bool do_memory = true;
+  bool do_comms = true;
+  bool do_flops = true;
+
+  // NOTE: these two take O((number of ranks)^2) time, which might be a lot, so they are
+  // off by default
+  bool do_latency = false;
+  bool do_p2p = false;
+
   std::string json_filename = ""; // empty indicates no json output
   for (int i = 0; i < argc; i++)
   {
-    if (std::string(argv[i]) == "--json-out")
+    auto arg = std::string(argv[i]);
+    if (arg == "--json-out")
       json_filename = argv[i + 1];
+    if (arg == "--benchmark-su4")
+      do_su4 = true;
+    if (arg == "--benchmark-memory")
+      do_memory = true;
+    if (arg == "--benchmark-comms")
+      do_comms = true;
+    if (arg == "--benchmark-flops")
+      do_flops = true;
+    if (arg == "--benchmark-latency")
+      do_latency = true;
+    if (arg == "--benchmark-p2p")
+      do_p2p = true;
+    if (arg == "--no-benchmark-su4")
+      do_su4 = false;
+    if (arg == "--no-benchmark-memory")
+      do_memory = false;
+    if (arg == "--no-benchmark-comms")
+      do_comms = false;
+    if (arg == "--no-benchmark-flops")
+      do_flops = false;
+    if (arg == "--no-benchmark-latency")
+      do_latency = false;
+    if (arg == "--no-benchmark-p2p")
+      do_p2p = false;
   }
 
   CartesianCommunicator::SetCommunicatorPolicy(
@@ -792,12 +1039,6 @@ int main(int argc, char **argv)
 #endif
   Benchmark::Decomposition();
 
-  int do_su4 = 1;
-  int do_memory = 1;
-  int do_comms = 1;
-  int do_flops = 1;
-  int Ls = 1;
-
   int sel = 4;
   std::vector<int> L_list({8, 12, 16, 24, 32});
   int selm1 = sel - 1;
@@ -830,6 +1071,22 @@ int main(int argc, char **argv)
     Benchmark::Comms();
   }
 
+  if (do_latency)
+  {
+    grid_big_sep();
+    std::cout << GridLogMessage << " Latency benchmark " << std::endl;
+    grid_big_sep();
+    Benchmark::Latency();
+  }
+
+  if (do_p2p)
+  {
+    grid_big_sep();
+    std::cout << GridLogMessage << " Point-To-Point benchmark " << std::endl;
+    grid_big_sep();
+    Benchmark::P2P();
+  }
+
   if (do_flops)
   {
     Ls = 1;
@@ -889,6 +1146,8 @@ int main(int argc, char **argv)
     json_results["flops"] = tmp_flops;
   }
 
+  json_results["hostnames"] = get_mpi_hostnames();
+
   if (!json_filename.empty())
   {
     std::cout << GridLogMessage << "writing benchmark results to " << json_filename

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]'

Quda/Benchmark_Quda.cpp

@@ -2,23 +2,73 @@
 #include <array>
 #include <blas_quda.h>
 #include <cassert>
+#include <chrono>
 #include <color_spinor_field.h>
+#include <communicator_quda.h>
 #include <dirac_quda.h>
+#include <fstream>
 #include <gauge_tools.h>
 #include <memory>
 #include <mpi.h>
 #include <stdio.h>
 #include <stdlib.h>
 
-using namespace quda;
-
 // remove to use QUDA's own flop counting instead of Grid's convention
 #define FLOP_COUNTING_GRID
 
+#include "json.hpp"
+using nlohmann::json;
+json json_results;
+
+using namespace quda;
+
+// thanks chatGPT :)
+std::string get_timestamp()
+{
+  // Get the current time
+  auto now = std::chrono::system_clock::now();
+
+  // Convert the current time to a time_t object
+  std::time_t currentTime = std::chrono::system_clock::to_time_t(now);
+
+  // Format the time using std::put_time
+  std::stringstream ss;
+  ss << std::put_time(std::localtime(&currentTime), "%Y%m%d %H:%M:%S");
+
+  return ss.str();
+}
+
 // This is the MPI grid, i.e. the layout of ranks
 int nranks = -1;
 std::array<int, 4> mpi_grid = {1, 1, 1, 1};
 
+// run f() in a loop for roughly target_time seconds
+// returns seconds per iteration it took
+template <class F> double bench(F const &f, double target_time, int niter_warmup = 5)
+{
+  device_timer_t timer;
+  timer.start();
+  for (int iter = 0; iter < niter_warmup; ++iter)
+    f();
+  timer.stop();
+
+  double secs = timer.last() / niter_warmup;
+  int niter = std::max(1, int(target_time / secs));
+  // niter = std::min(1000, niter);
+  // printfQuda("during warmup took %f s/iter, deciding on %d iters\n", secs, niter);
+
+  // important: each rank has its own timer, so their measurements can slightly vary. But
+  // 'niter' needs to be consistent (bug took me a couple hours to track down)
+  comm_broadcast_global(&niter, sizeof(niter), 0);
+
+  timer.reset(__FUNCTION__, __FILE__, __LINE__);
+  timer.start();
+  for (int iter = 0; iter < niter; ++iter)
+    f();
+  timer.stop();
+  return timer.last() / niter;
+}
+
 void initComms(int argc, char **argv)
 {
   // init MPI communication
@@ -43,6 +93,9 @@ void initComms(int argc, char **argv)
   for (int d = 0; d < 4; d++)
     if (mpi_grid[d] > 1)
       commDimPartitionedSet(d);
+
+  json_results["geometry"]["ranks"] = nranks;
+  json_results["geometry"]["mpi"] = mpi_grid;
 }
 
 // creates a random gauge field. L = local(!) size
@@ -149,11 +202,8 @@ ColorSpinorField make_source(int L, int Ls = 1)
   return src;
 }
 
-void benchmark_wilson()
+void benchmark_wilson(std::vector<int> const &L_list, double target_time)
 {
-  int niter = 20;
-  int niter_warmup = 10;
-
   printfQuda("==================== wilson dirac operator ====================\n");
 #ifdef FLOP_COUNTING_GRID
   printfQuda("IMPORTANT: flop counting as in Benchmark_Grid\n");
@@ -163,8 +213,10 @@ void benchmark_wilson()
 #endif
   printfQuda("%5s %15s %15s\n", "L", "time (usec)", "Gflop/s/rank");
 
-  for (int L : {8, 12, 16, 24, 32, 48})
+  for (int L : L_list)
   {
+    // printfQuda("starting wilson L=%d\n", L);
+
     auto U = make_gauge_field(L);
     auto src = make_source(L);
 
@@ -179,44 +231,41 @@ void benchmark_wilson()
     // (the additional nullptr's are for smeared links and fancy preconditioners and such.
     // Not used for simple Wilson fermions)
     dirac.updateFields(&U, nullptr, nullptr, nullptr);
+    auto res = ColorSpinorField(ColorSpinorParam(src));
+    auto f = [&]() { dirac.Dslash(res, src, QUDA_EVEN_PARITY); };
 
-    auto tmp = ColorSpinorField(ColorSpinorParam(src));
-
-    // couple iterations without timing to warm up
-    for (int iter = 0; iter < niter_warmup; ++iter)
-      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
-
-    // actual benchmark with timings
+    // first run to get the quda tuning out of the way
     dirac.Flops(); // reset flops counter
-    device_timer_t device_timer;
-    device_timer.start();
-    for (int iter = 0; iter < niter; ++iter)
-      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
-    device_timer.stop();
+    f();
+    double flops = 1.0 * dirac.Flops();
 
-    double secs = device_timer.last() / niter;
+    // actual benchmarking
+    auto start_time = get_timestamp();
+    double secs = bench(f, target_time);
+    auto end_time = get_timestamp();
 
 #ifdef FLOP_COUNTING_GRID
     // this is the flop counting from Benchmark_Grid
     double Nc = 3;
     double Nd = 4;
     double Ns = 4;
-    double flops =
-        (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops = (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
     flops *= L * L * L * L / 2.0;
-#else
-    double flops = 1.0 * dirac.Flops() / niter;
 #endif
 
     printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+
+    json tmp;
+    tmp["L"] = L;
+    tmp["Gflops_wilson"] = flops / secs * 1e-9;
+    tmp["start_time"] = start_time;
+    tmp["end_time"] = end_time;
+    json_results["flops"]["results"].push_back(tmp);
   }
 }
 
-void benchmark_dwf()
+void benchmark_dwf(std::vector<int> const &L_list, double target_time)
 {
-  int niter = 20;
-  int niter_warmup = 10;
-
   printfQuda("==================== domain wall dirac operator ====================\n");
 #ifdef FLOP_COUNTING_GRID
   printfQuda("IMPORTANT: flop counting as in Benchmark_Grid\n");
@@ -226,8 +275,9 @@ void benchmark_dwf()
 #endif
   printfQuda("%5s %15s %15s\n", "L", "time (usec)", "Gflop/s/rank");
   int Ls = 12;
-  for (int L : {8, 12, 16, 24})
+  for (int L : L_list)
   {
+    // printfQuda("starting dwf L=%d\n", L);
     auto U = make_gauge_field(L);
     auto src = make_source(L, Ls);
 
@@ -243,45 +293,43 @@ void benchmark_dwf()
     // insert gauge field into the dirac operator
     // (the additional nullptr's are for smeared links and fancy preconditioners and such)
     dirac.updateFields(&U, nullptr, nullptr, nullptr);
+    auto res = ColorSpinorField(ColorSpinorParam(src));
+    auto f = [&]() { dirac.Dslash(res, src, QUDA_EVEN_PARITY); };
 
-    auto tmp = ColorSpinorField(ColorSpinorParam(src));
-
-    // couple iterations without timing to warm up
-    for (int iter = 0; iter < niter_warmup; ++iter)
-      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
-
-    // actual benchmark with timings
+    // first run to get the quda tuning out of the way
     dirac.Flops(); // reset flops counter
-    device_timer_t device_timer;
-    device_timer.start();
-    for (int iter = 0; iter < niter; ++iter)
-      dirac.Dslash(tmp, src, QUDA_EVEN_PARITY);
-    device_timer.stop();
+    f();
+    double flops = 1.0 * dirac.Flops();
 
-    double secs = device_timer.last() / niter;
+    // actual benchmarking
+    auto start_time = get_timestamp();
+    double secs = bench(f, target_time);
+    auto end_time = get_timestamp();
 
 #ifdef FLOP_COUNTING_GRID
     // this is the flop counting from Benchmark_Grid
     double Nc = 3;
     double Nd = 4;
     double Ns = 4;
-    double flops =
-        (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
+    flops = (Nc * (6 + (Nc - 1) * 8) * Ns * Nd + 2 * Nd * Nc * Ns + 2 * Nd * Nc * Ns * 2);
     flops *= L * L * L * L * Ls / 2.0;
-#else
-    double flops = 1.0 * dirac.Flops() / niter;
 #endif
 
     printfQuda("%5d %15.2f %15.2f\n", L, secs * 1e6, flops / secs * 1e-9);
+    json tmp;
+    tmp["L"] = L;
+    tmp["Gflops_dwf4"] = flops / secs * 1e-9;
+    tmp["start_time"] = start_time;
+    tmp["end_time"] = end_time;
+    json_results["flops"]["results"].push_back(tmp);
   }
 }
 
-void benchmark_axpy()
+void benchmark_axpy(std::vector<int> const &L_list, double target_time)
 {
   // number of iterations for warmup / measurement
   // (feel free to change for noise/time tradeoff)
-  constexpr int niter_warmup = 10;
-  constexpr int niter = 20;
+  constexpr int niter_warmup = 5;
 
   printfQuda("==================== axpy / memory ====================\n");
 
@@ -305,11 +353,11 @@ void benchmark_axpy()
 
   printfQuda("%5s %15s %15s %15s %15s\n", "L", "size (MiB/rank)", "time (usec)",
              "GiB/s/rank", "Gflop/s/rank");
-  std::vector L_list = {8, 12, 16, 24, 32, 48};
   for (int L : L_list)
   {
-    // IMPORTANT: all of `param.x`, `field_elements`, `field.Bytes()`
-    //            are LOCAL, i.e. per rank / per GPU
+    // printfQuda("starting axpy L=%d\n", L);
+    //  IMPORTANT: all of `param.x`, `field_elements`, `field.Bytes()`
+    //             are LOCAL, i.e. per rank / per GPU
 
     param.x[0] = L;
     param.x[1] = L;
@@ -336,26 +384,41 @@ void benchmark_axpy()
     double flops = 2 * field_elements;
     double memory = 3 * sizeof(float) * field_elements;
 
-    // do some iterations to to let QUDA do its internal tuning and also stabilize cache
-    // behaviour and such
-    for (int iter = 0; iter < niter_warmup; ++iter)
-      blas::axpy(1.234, fieldA, fieldB);
+    auto f = [&]() { blas::axpy(1.234, fieldA, fieldB); };
 
-    // running the actual benchmark
-    device_timer_t device_timer;
-    device_timer.start();
-    for (int iter = 0; iter < niter; ++iter)
-      blas::axpy(1.234, fieldA, fieldB);
-    device_timer.stop();
-    double secs = device_timer.last() / niter; // seconds per iteration
+    // first run to get the quda tuning out of the way
+    f();
 
-    printfQuda("%5d %15.2f %15.2f %15.2f %15.2f\n", L, memory / 1024. / 1024., secs * 1e6,
-               memory / secs / 1024. / 1024. / 1024., flops / secs * 1e-9);
+    // actual benchmarking
+    auto start_time = get_timestamp();
+    double secs = bench(f, target_time);
+    auto end_time = get_timestamp();
+
+    double mem_MiB = memory / 1024. / 1024.;
+    double GBps = mem_MiB / 1024 / secs;
+    printfQuda("%5d %15.2f %15.2f %15.2f %15.2f\n", L, mem_MiB, secs * 1e6, GBps,
+               flops / secs * 1e-9);
+
+    json tmp;
+    tmp["L"] = L;
+    tmp["size_MB"] = mem_MiB;
+    tmp["GBps"] = GBps;
+    tmp["GFlops"] = flops / secs * 1e-9;
+    tmp["start_time"] = start_time;
+    tmp["end_time"] = end_time;
+    json_results["axpy"].push_back(tmp);
   }
 }
 
 int main(int argc, char **argv)
 {
+  std::string json_filename = ""; // empty indicates no json output
+  for (int i = 0; i < argc; i++)
+  {
+    if (std::string(argv[i]) == "--json-out")
+      json_filename = argv[i + 1];
+  }
+
   initComms(argc, argv);
 
   initQuda(-1); // -1 for multi-gpu. otherwise this selects the device to be used
@@ -367,14 +430,28 @@ int main(int argc, char **argv)
   printfQuda("MPI layout = %d %d %d %d\n", mpi_grid[0], mpi_grid[1], mpi_grid[2],
              mpi_grid[3]);
 
-  benchmark_axpy();
+  benchmark_axpy({8, 12, 16, 24, 32, 48}, 1.0);
 
   setVerbosity(QUDA_SILENT);
-  benchmark_wilson();
-  benchmark_dwf();
+  benchmark_wilson({8, 12, 16, 24, 32, 48}, 1.0);
+  benchmark_dwf({8, 12, 16, 24, 32}, 1.0);
   setVerbosity(QUDA_SUMMARIZE);
 
   printfQuda("==================== done with all benchmarks ====================\n");
+
+  if (!json_filename.empty())
+  {
+    printfQuda("writing benchmark results to %s\n", json_filename.c_str());
+
+    int me = 0;
+    MPI_Comm_rank(MPI_COMM_WORLD, &me);
+    if (me == 0)
+    {
+      std::ofstream json_file(json_filename);
+      json_file << std::setw(2) << json_results;
+    }
+  }
+
   endQuda();
   quda::comm_finalize();
   MPI_Finalize();

Quda/build-benchmark.sh

@@ -28,5 +28,5 @@ mkdir -p "${PREFIX_DIR}"
 
 LINK_FLAGS="-Wl,-rpath,$QUDA_DIR/lib: $QUDA_DIR/lib/libquda.so $EXTRA_LIBS -lpthread -lmpi"
 
-g++ $BUILD_FLAGS  -I$QUDA_DIR/include   -c -o $BUILD_DIR/Benchmark_Quda.o  $script_dir/Benchmark_Quda.cpp
+g++ $BUILD_FLAGS -I$QUDA_DIR/include/targets/cuda  -I$QUDA_DIR/include   -c -o $BUILD_DIR/Benchmark_Quda.o  $script_dir/Benchmark_Quda.cpp
 g++ -g -O3 $BUILD_DIR/Benchmark_Quda.o -o $PREFIX_DIR/Benchmark_Quda $LINK_FLAGS -lmpi