Merge pull request #1327 from LLNL/bugfix/whitlock/point_query_tags

Fixes for tags that affected PointQuery on some systems.

CHANGELOG.md

@@ -20,6 +20,14 @@ and this project aspires to adhere to [Semantic Versioning](https://semver.org/s
 #### Conduit
 - Changed the MPI CMake target used by conduit from `MPI:MPI_CXX` to `MPI:MPI_C` to provide better compatibility with downstream tools.
 
+#### Blueprint
+- Certain algorithms that use MPI tags had their tag values lowered since some MPI implementations do not support large values.
+
+#### Relay
+- User-supplied warning and error handlers are suspended during `conduit::relay::communicate_using_schema::execute()` so exceptions will be thrown properly when there is an MPI error. The handlers are restored before the execute method returns.
+- `conduit::relay::communicate_using_schema::execute()` flushes logs as they are generated, in case of error. This is mostly to facilitate internal debugging.
+- Changes were made to how Relay queries the upper limit for MPI tags to work around problems on some systems.
+
 ## [0.9.2] - Released 2024-05-21
 
 ### Added

src/libs/blueprint/conduit_blueprint_mpi_mesh_parmetis.cpp

@@ -375,7 +375,7 @@ void generate_global_element_and_vertex_ids(conduit::Node &mesh,
 
     if (adjset_name != "")
     {
-        const int TAG_SHARED_NODE_SYNC = 175000000;
+        const int TAG_SHARED_NODE_SYNC = 175;
         // map of groups -> global vtx ids
         std::map<std::set<uint64>, std::vector<uint64>> groups_2_vids;
         // map of rank -> sends to/recvs from that rank of global vtx ids for
@@ -464,7 +464,7 @@ void generate_global_element_and_vertex_ids(conduit::Node &mesh,
             for (const std::set<uint64>& group : recv_groups)
             {
                 index_t domid = *(group.begin());
-                const int tag = conduit::relay::mpi::safe_tag(TAG_SHARED_NODE_SYNC + domid * 100 + group_idx);
+                const int tag = conduit::relay::mpi::safe_tag(TAG_SHARED_NODE_SYNC + domid * 100 + group_idx, comm);
                 async_recvs.push_back(MPI_Request{});
                 group_idx++;
                 std::vector<uint64>& recvbuf = groups_2_vids[group];
@@ -484,7 +484,7 @@ void generate_global_element_and_vertex_ids(conduit::Node &mesh,
             for (const std::set<uint64>& group : send_groups)
             {
                 index_t domid = *(group.begin());
-                const int tag = conduit::relay::mpi::safe_tag(TAG_SHARED_NODE_SYNC + domid * 100 + group_idx);
+                const int tag = conduit::relay::mpi::safe_tag(TAG_SHARED_NODE_SYNC + domid * 100 + group_idx, comm);
                 async_sends.push_back(MPI_Request{});
                 group_idx++;
                 std::vector<uint64>& sendbuf = groups_2_vids[group];

src/libs/blueprint/conduit_blueprint_mpi_mesh_utils.cpp

@@ -173,8 +173,8 @@ PointQuery::execute(const std::string &coordsetName)
     // the results. The results get stored in m_domResults.
     std::map<std::pair<int,int>, conduit::Node *> input_sends, result_sends,
                                                   input_recvs, result_recvs;
-    int inputs_tag = 55000000;
-    int results_tag = 66000000;
+    const int inputs_tag = 550;
+    const int results_tag = 660;
     for(int pass = 0; pass < 2; pass++)
     {
         conduit::relay::mpi::communicate_using_schema C(m_comm);
@@ -187,10 +187,10 @@ PointQuery::execute(const std::string &coordsetName)
 #endif
         for(size_t i = 0; i < allqueries.size(); i += 3)
         {
-            int asker  = allqueries[i];
-            int domain = allqueries[i+1];
-            int npts   = allqueries[i+2];
-            int owner = domain_to_rank[domain];
+            const int asker  = allqueries[i];
+            const int domain = allqueries[i+1];
+            const int npts   = allqueries[i+2];
+            const int owner = domain_to_rank[domain];
 
             if(asker == rank)
             {
@@ -278,6 +278,7 @@ PointQuery::execute(const std::string &coordsetName)
     for(auto it = result_recvs.begin(); it != result_recvs.end(); it++)
     {
         int domain = it->first.second;
+
         const conduit::Node &r = it->second->fetch_existing("results");
         auto acc = r.as_int_accessor();
         std::vector<int> &result = m_domResults[domain];
@@ -435,12 +436,12 @@ MatchQuery::execute()
     C.set_logging_root("mpi_matchquery");
     C.set_logging(true);
 #endif
-    int query_tag = 77000000;    
+    const int query_tag = 770;
     for(size_t i = 0; i < allqueries.size(); i += ntuple_values)
     {
-        int owner = allqueries[i];
-        int domain = allqueries[i + 1];
-        int query_domain = allqueries[i + 2];
+        const int owner = allqueries[i];
+        const int domain = allqueries[i + 1];
+        const int query_domain = allqueries[i + 2];
 
         auto oppositeKey = std::make_pair(query_domain, domain);
 
@@ -667,7 +668,7 @@ compare_pointwise_impl(conduit::Node &mesh, const std::string &adjsetName,
 
     // Iterate over each of the possible adjset relationships. Not all of these
     // will have adjset groups.
-    const int tag = 12211221;
+    const int tag = 122;
     for(int d0 = 0; d0 < maxDomains; d0++)
     {
         for(int d1 = d0 + 1; d1 < maxDomains; d1++)