diff --git a/src/backends/detection/detection.cpp b/src/backends/detection/detection.cpp
index 3fb20fa11c928f4ee86dc81589c538c04451f5fb..2be1c928bd6b6b6cfc4b724085277ab4bbf988d4 100644
--- a/src/backends/detection/detection.cpp
+++ b/src/backends/detection/detection.cpp
@@ -44,7 +44,7 @@ ThreadStore getThreadStoreOfCaller() {
 
 double getTimeStamp() {
 #ifdef WITH_MPI
-  return MPI_Wtime();
+  return PMPI_Wtime();
 #else
   using clock = std::chrono::system_clock;
   auto now = clock::now();
diff --git a/src/backends/dlb_talp_tree/dlb_talp_tree.cpp b/src/backends/dlb_talp_tree/dlb_talp_tree.cpp
index 19e7bcb50293fc35c1e946e17ba862fcc7833927..50802a7a27eff0ac652c56a4f4d1e525599784e0 100644
--- a/src/backends/dlb_talp_tree/dlb_talp_tree.cpp
+++ b/src/backends/dlb_talp_tree/dlb_talp_tree.cpp
@@ -5,6 +5,8 @@
 #include <iomanip>
 #include <iostream>
 #include <nlohmann/json.hpp>
+#include <numeric>
+#include <set>
 #include <sstream>
 #include <stack>
 
@@ -101,16 +103,72 @@ void DLBTalpTreeStrategy::Finalize() noexcept {
   // Stop the top region which was automatically started by TALP.
   talp_profiling_strategy_.RegionStop({.name = top_region_});
 
-  // Collect the POP metrics from all the regions
+  int num_regions = regions_.size();
+
+  std::vector<std::size_t> region_hashes;
+  region_hashes.reserve(num_regions);
+
   for (auto &region : regions_) {
-    std::string name = region_hash_to_string(region.first);
-    if (region.second.name.compare(top_region_) == 0) {
+    region_hashes.push_back(region.first);
+  }
+
+  std::set<std::size_t> unique_regions;
+  if (mpi_helper_.IsUsingMPI()) {
+#ifdef WITH_MPI
+    int mpi_size = mpi_helper_.getCommSize();
+    std::vector<int> counts;
+    std::vector<int> displs;
+    counts.resize(mpi_size);
+    displs.resize(mpi_size);
+
+    MPI_Allgather(&num_regions, 1, MPI_INT, counts.data(), 1, MPI_INT,
+                  MPI_COMM_WORLD);
+
+    int num_all_regions = 0;
+    for (int i = 0; i < counts.size(); i++) {
+      num_all_regions += counts[i];
+    }
+
+    int last = 0;
+    for (int i = 0; i < counts.size(); i++) {
+      counts[i] *= sizeof(std::size_t);
+      displs[i] = last;
+      last += counts[i];
+    }
+
+    std::vector<std::size_t> all_regions(num_all_regions);
+    PMPI_Allgatherv(
+        // What we send
+        region_hashes.data(), region_hashes.size() * sizeof(std::size_t),
+        MPI_BYTE,
+
+        // What we receive
+        all_regions.data(), counts.data(), displs.data(), MPI_BYTE,
+
+        // To whom
+        MPI_COMM_WORLD);
+
+    for (std::size_t region_hash : all_regions) {
+      unique_regions.insert(region_hash);
+    }
+#endif
+  } else {
+    for (std::size_t region_hash : region_hashes) {
+      unique_regions.insert(region_hash);
+    }
+  }
+
+  // Collect the POP metrics from all the regions
+  for (auto &region_hash : unique_regions) {
+    auto &region = regions_[region_hash];
+
+    std::string name = region_hash_to_string(region_hash);
+    if (region.name.compare(top_region_) == 0) {
       name = top_region_;
     }
 
     auto handle = DLB_MonitoringRegionRegister(name.c_str());
-    int dlb_error =
-        DLB_TALP_CollectPOPMetrics(handle, &region.second.pop_metrics);
+    int dlb_error = DLB_TALP_CollectPOPMetrics(handle, &region.pop_metrics);
 
     if (dlb_error != DLB_SUCCESS) {
       // Warn about the error trying to close region name with hash
diff --git a/src/utils/parallelism_helper.cpp b/src/utils/parallelism_helper.cpp
index 2955cc4108acd2643eae001bdecdbf7e53140134..7b63ba3329d8c41b6828c2943f442315e9fbf046 100644
--- a/src/utils/parallelism_helper.cpp
+++ b/src/utils/parallelism_helper.cpp
@@ -24,14 +24,16 @@ bool MPIHelper::IsRankNumber(int asked_rank) const {
 
 int MPIHelper::getRankNumber() const { return mpi_comm_rank; }
 
+int MPIHelper::getCommSize() const { return mpi_comm_size; }
+
 MPIHelper::MPIHelper() {
   if (!mpi_information.valid) {
     int is_initialized;
-    MPI_Initialized(&is_initialized);
+    PMPI_Initialized(&is_initialized);
     mpi_information.mpi_is_initialized = static_cast<bool>(is_initialized);
     if (mpi_information.mpi_is_initialized) {
-      MPI_Comm_size(MPI_COMM_WORLD, &mpi_information.comm_size);
-      MPI_Comm_rank(MPI_COMM_WORLD, &mpi_information.rank);
+      PMPI_Comm_size(MPI_COMM_WORLD, &mpi_information.comm_size);
+      PMPI_Comm_rank(MPI_COMM_WORLD, &mpi_information.rank);
     }
     mpi_information.valid = true;
   }
@@ -52,6 +54,8 @@ bool MPIHelper::IsRankNumber(int asked_rank) const { return asked_rank == 0; }
 
 int MPIHelper::getRankNumber() const { return mpi_comm_rank; }
 
+int MPIHelper::getCommSize() const { return mpi_comm_size; }
+
 MPIHelper::MPIHelper() {}
 
 #endif
diff --git a/src/utils/parallelism_helper.hpp b/src/utils/parallelism_helper.hpp
index 7561d026345079f38b3035538cf2753391f296e3..e92bf2117e42ded0590f70db5a6ff7c641a84153 100644
--- a/src/utils/parallelism_helper.hpp
+++ b/src/utils/parallelism_helper.hpp
@@ -23,6 +23,8 @@ class MPIHelper {
   bool IsRankNumber(int asked_rank) const;
 
   int getRankNumber() const;
+
+  int getCommSize() const;
 };
 
 #endif  // NESMIK_PARALLELISM_HELPER_HPP
diff --git a/src/utils/terminator.cpp b/src/utils/terminator.cpp
index b9b3ca0653db87e9e8a35874ea57823bcee8b4ee..527198ffe0f76be6c06e59627fc0ba72bac77572 100644
--- a/src/utils/terminator.cpp
+++ b/src/utils/terminator.cpp
@@ -13,7 +13,7 @@ void Terminator::exit(int exitcode) {
   }
 #ifdef WITH_MPI
   if (mpi_helper.IsUsingMPI()) {
-    MPI_Abort(MPI_COMM_WORLD, exitcode);
+    PMPI_Abort(MPI_COMM_WORLD, exitcode);
   } else
 #endif
   {
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 384053429942e9a7ca5d6aae9b6b43be418d5862..9223038c8cfd3f81239e8e7805592ea0f053ff4e 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -43,7 +43,15 @@ add_test(NAME TestCppTalpTree
                   PRIVATE ${NESMIK_PUBLIC_HEADERS})
 
             add_test(NAME TestCppTalpTreeMPI
-                  COMMAND mpirun -np 2 TestCppTalpTree)
+                  COMMAND TestCppTalpTree)
+
+            add_executable(TestCppTalpTreeMPI_MPMD cpp/TestCppTalpTreeMPI_MPMD.cpp)
+            target_link_libraries(TestCppTalpTreeMPI_MPMD nesmik::nesmik MPI::MPI_CXX)
+            target_include_directories(TestCppTalpTreeMPI_MPMD
+                  PRIVATE ${NESMIK_PUBLIC_HEADERS})
+
+            add_test(NAME TestCppTalpTreeMPI_MPMD
+                  COMMAND TestCppTalpTreeMPI_MPMD)
 
       endif()
 
diff --git a/tests/cpp/TestCppTalpTreeMPI_MPMD.cpp b/tests/cpp/TestCppTalpTreeMPI_MPMD.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..600f2d3b49e80b7e762143a0642378a7e6b0fd4d
--- /dev/null
+++ b/tests/cpp/TestCppTalpTreeMPI_MPMD.cpp
@@ -0,0 +1,15 @@
+#include <mpi.h>
+
+#include "nesmik/nesmik.hpp"
+int main() {
+  MPI_Init(NULL, NULL);
+  int rank;
+  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+  nesmik::nesmik_init();
+  nesmik::region_start("Top");
+  nesmik::region_start(std::to_string(rank));
+  nesmik::region_stop(std::to_string(rank));
+  nesmik::region_stop("Top");
+  nesmik::nesmik_finalize();
+  MPI_Finalize();
+}