Use Barriers in profiling

Author: lukasdreyer

src/t8_forest/t8_forest.cxx

@@ -3032,6 +3032,9 @@ t8_forest_commit (t8_forest_t forest)
   T8_ASSERT (!forest->committed);
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of commit */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->commit_runtime = sc_MPI_Wtime ();
   }
 
@@ -3247,6 +3250,9 @@ t8_forest_commit (t8_forest_t forest)
 
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of commit */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->commit_runtime = sc_MPI_Wtime () - forest->profile->commit_runtime;
   }
 

src/t8_forest/t8_forest_balance.cxx

@@ -164,6 +164,9 @@ t8_forest_balance (t8_forest_t forest, int repartition)
   adap_stats = ghost_stats = partition_stats = NULL;
 
   if (forest->profile != NULL) {
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     /* Profiling is enable, so we measure the runtime of balance */
     forest->profile->balance_runtime = -sc_MPI_Wtime ();
     /* We store the individual adapt, ghost, and partition runtimes */
@@ -285,6 +288,9 @@ t8_forest_balance (t8_forest_t forest, int repartition)
 
   if (forest->profile != NULL) {
     /* Profiling is enabled, so we measure the runtime of balance. */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->balance_runtime += sc_MPI_Wtime ();
     forest->profile->balance_rounds = count_rounds;
     /* Print the runtime of adapt/ghost/partition */

src/t8_forest/t8_forest_ghost.cxx

@@ -1551,6 +1551,9 @@ t8_forest_ghost_create_ext (t8_forest_t forest, int unbalanced_version)
 
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of ghost_create */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->ghost_runtime = -sc_MPI_Wtime ();
     /* DO NOT DELETE THE FOLLOWING line.
      * even if you do not want this output. It fixes a bug that occurred on JUQUEEN, where the
@@ -1608,6 +1611,9 @@ t8_forest_ghost_create_ext (t8_forest_t forest, int unbalanced_version)
 
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of ghost_create */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->ghost_runtime += sc_MPI_Wtime ();
     /* We also store the number of ghosts and remotes */
     if (ghost != NULL) {

src/t8_forest/t8_forest_partition.cxx

@@ -92,6 +92,9 @@ t8_forest_partition_create_offsets (t8_forest_t forest)
 {
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->forest_offsets_runtime = -sc_MPI_Wtime ();
   }
 
@@ -120,6 +123,9 @@ t8_forest_partition_create_offsets (t8_forest_t forest)
   t8_shmem_array_end_writing (forest->element_offsets);
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->forest_offsets_runtime += sc_MPI_Wtime ();
   }
 }
@@ -266,6 +272,9 @@ t8_forest_partition_create_first_desc (t8_forest_t forest)
 {
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->first_descendant_runtime = -sc_MPI_Wtime ();
   }
   sc_MPI_Comm comm;
@@ -348,6 +357,9 @@ t8_forest_partition_create_first_desc (t8_forest_t forest)
   #endif
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->first_descendant_runtime += sc_MPI_Wtime ();
   }
 }
@@ -358,6 +370,9 @@ t8_forest_partition_create_tree_offsets (t8_forest_t forest)
 
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->cmesh_offsets_runtime = -sc_MPI_Wtime ();
   }
 
@@ -426,6 +441,9 @@ t8_forest_partition_create_tree_offsets (t8_forest_t forest)
   }
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->cmesh_offsets_runtime += sc_MPI_Wtime ();
   }
 }
@@ -1209,6 +1227,9 @@ t8_forest_partition (t8_forest_t forest)
 
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->partition_runtime = sc_MPI_Wtime ();
 
     /* DO NOT DELETE THE FOLLOWING line.
@@ -1239,6 +1260,9 @@ t8_forest_partition (t8_forest_t forest)
 
   if (forest->profile != NULL) {
     /* If profiling is enabled, we measure the runtime of partition */
+#if T8_ENABLE_PROFILE_BARRIER
+    sc_MPI_Barrier(forest->mpicomm);
+#endif
     forest->profile->partition_runtime = sc_MPI_Wtime () - forest->profile->partition_runtime;
 
     /* DO NOT DELETE THE FOLLOWING line.