treematch hang

[ggouaillardet]

@bosilca the distgraph_test_4 test from the ibm test suite might hang depending on the machine topology.
with 4 mpi tasks, it works fine on a server with 2 sockets / 12 cores / 24 threads, but it hangs on my VM with 1 socket / 4 cores / 4 threads

the inlined topology can be used to evidence the issue

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE topology SYSTEM "hwloc.dtd">
<topology>
  <object type="Machine" os_index="0" cpuset="0x0000000f" complete_cpuset="0x0000000f" allowed_cpuset="0x0000000f" nodeset="0x00000001" complete_nodeset="0x00000001" allowed_nodeset="0x00000001">
    <object type="Package" os_index="0" cpuset="0x0000000f" complete_cpuset="0x0000000f" allowed_cpuset="0x0000000f" nodeset="0x00000001" complete_nodeset="0x00000001" allowed_nodeset="0x00000001">
      <object type="PU" os_index="0" cpuset="0x00000001" complete_cpuset="0x00000001" allowed_cpuset="0x00000001" nodeset="0x00000001" complete_nodeset="0x00000001" allowed_nodeset="0x00000001"/>
      <object type="PU" os_index="1" cpuset="0x00000002" complete_cpuset="0x00000002" allowed_cpuset="0x00000002" nodeset="0x00000001" complete_nodeset="0x00000001" allowed_nodeset="0x00000001"/>
      <object type="PU" os_index="2" cpuset="0x00000004" complete_cpuset="0x00000004" allowed_cpuset="0x00000004" nodeset="0x00000001" complete_nodeset="0x00000001" allowed_nodeset="0x00000001"/>
      <object type="PU" os_index="3" cpuset="0x00000008" complete_cpuset="0x00000008" allowed_cpuset="0x00000008" nodeset="0x00000001" complete_nodeset="0x00000001" allowed_nodeset="0x00000001"/>
    </object>
  </object>
</topology>

run as is

$ mpirun -np 4 ./distgraph_test_4
pass!!

run with a simple topology

$ mpirun --mca hwloc_base_topo_file smp.xml -np 4 ./distgraph_test_4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 
nb_constraints = 4, N= 4; nb_processing units = 4
========== Centralized Reordering ========= 

HANG

but this works fine without the treematch module

$ mpirun --mca hwloc_base_topo_file smp.xml --mca topo ^treematch -np 4 ./distgraph_test_4
pass!!

note i had to push a5440ad so the topology file is used by the treematch module (otherwise, hwloc use the topology file, but treematch use the topology of the node)

can you please have a look at this ?

[bosilca]

@ggouaillardet a quick look at the test seems to indicate that the test is broken. On my machine the test does not deadlock or segfault, but instead it triggers a nice MPI abort due to a wrong argument in MPI_Dist_graph_create.

If I look how this function is called, for cnt = 9 I have:

n = 4
sources[] = {0, 1, 2, 3}
degrees[] = {2, 2, 2, 3}
destinations[] = {1, 2, 0, 2, 0, 1, 0, 1, 10}

The last destination is clearly wrong (larger than the number of nodes), but should be valid as the total number of degrees is 11.

[ggouaillardet]

@bosilca goot catch !
when cnt is 9, destinations does not contains enough elements (!)
i will fix that and review it all from now
(and that is why you can get a nice error when i just get a hang)

[ggouaillardet]

@bosilca i pushed open-mpi/ompi-tests@0c77d1ba7b65896a84afba0405e2477dcbaa2875 in order to fix the issue you identified.

here is a main subroutine that can replace the one from distgraph_test_4.c in order to go straight to the hang

int main(int argc, char *argv[]){

    int rank, num_procs, rc, root;
    MPI_Comm newcomm;
    int cnt;

    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &num_procs);

    Reorder = 1;
    root = 0;

    const_graph(root);
    for(cnt=1; cnt <= MAX_TEST; cnt++){
        create_graph(cnt, root, &newcomm);
        if (MPI_SUCCESS != print_result(newcomm, cnt, root)){
            MPI_Abort(MPI_COMM_WORLD, -1);
        }
        if (MPI_SUCCESS != p2p_test(newcomm, cnt, root)){
            MPI_Abort(MPI_COMM_WORLD, -1);
        }
        if (MPI_SUCCESS != coll_test(newcomm, cnt, root)){
            MPI_Abort(MPI_COMM_WORLD, -1);
        }
        MPI_Comm_free(&newcomm);
    }
    dest_graph();

    MPI_Finalize();
    if(!rank){
        printf("pass!!\n");
        fflush(stdout);
    }
    return 0;
}

hang occurs with 4 MPI tasks on a 4 cores box, or on a 16 cores box if --bind-to core is used.
(i guess you have more than 4 cores on your machine, so you did not see the hang)

if you run with --bind-to core, then at topo_treematch_dist_graph_create.c:308, oversubscribing_objs is false
then you can jump to line 387
topo->weighted is true on rank 0, so rank 0 calls coll_gather
but topo->weighted is false on other ranks, so they do not call coll_gather and hence the hang

a possible fix is to coll_gather no matter what (since topo->weighted might not have the same value on all nodes.

here is a patch for that (it does include some code factorization)

diff --git a/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c b/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
index 6c31d1f..0b2b249 100644
--- a/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
+++ b/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
@@ -389,30 +389,28 @@ int mca_topo_treematch_dist_graph_create(mca_topo_base_module_t* topo_module,
             fprintf(stderr,"========== Centralized Reordering ========= \n");

             local_pattern = (double *)calloc(size*size,sizeof(double));
-            if( true == topo->weighted ) {
-                for(i = 0; i < topo->indegree ; i++)
-                    local_pattern[topo->in[i]] += topo->inw[i];
-                for(i = 0; i < topo->outdegree ; i++)
-                    local_pattern[topo->out[i]] += topo->outw[i];
-                if (OMPI_SUCCESS != (err = comm_old->c_coll.coll_gather(MPI_IN_PLACE, size, MPI_DOUBLE,
-                                                                        local_pattern, size, MPI_DOUBLE,
-                                                                        0, comm_old,
-                                                                        comm_old->c_coll.coll_gather_module)))
-                    return err;
-            }
         } else {
             local_pattern = (double *)calloc(size,sizeof(double));
-            if( true == topo->weighted ) {
-                for(i = 0; i < topo->indegree ; i++)
-                    local_pattern[topo->in[i]] += topo->inw[i];
-                for(i = 0; i < topo->outdegree ; i++)
-                    local_pattern[topo->out[i]] += topo->outw[i];
-                if (OMPI_SUCCESS != (err = comm_old->c_coll.coll_gather(local_pattern, size, MPI_DOUBLE,
-                                                                        NULL,0,0,
-                                                                        0, comm_old,
-                                                                        comm_old->c_coll.coll_gather_module)))
-                    return err;
-            }
+        }
+        if( true == topo->weighted ) {
+            for(i = 0; i < topo->indegree ; i++)
+                local_pattern[topo->in[i]] += topo->inw[i];
+            for(i = 0; i < topo->outdegree ; i++)
+                local_pattern[topo->out[i]] += topo->outw[i];
+        }
+        if(0 == rank) {
+            err = comm_old->c_coll.coll_gather(MPI_IN_PLACE, size, MPI_DOUBLE,
+                                               local_pattern, size, MPI_DOUBLE,
+                                               0, comm_old,
+                                               comm_old->c_coll.coll_gather_module);
+        } else {
+            err = comm_old->c_coll.coll_gather(local_pattern, size, MPI_DOUBLE,
+                                               NULL,0,0,
+                                               0, comm_old,
+                                               comm_old->c_coll.coll_gather_module);
+        }
+        if (OMPI_SUCCESS != err) {
+            return err;
         }

         if( rank == local_procs[0]) {

if the fix looks good to you, i guess a similar one should be applied around line 731 (Partially Distributed Reordering)

note that when running with --bind-to core, the treematch module issues a few printf
(no such thing without --bind-to core)

[ggouaillardet]

@bosilca
here is a trivial patch to silence the treematch outputs

diff --git a/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c b/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
index 6c31d1f..d97b017 100644
--- a/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
+++ b/ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
@@ -386,7 +386,9 @@ int mca_topo_treematch_dist_graph_create(mca_topo_base_module_t* topo_module,
          */
         if(0 == rank) {

+#ifdef __DEBUG__
             fprintf(stderr,"========== Centralized Reordering ========= \n");
+#endif

             local_pattern = (double *)calloc(size*size,sizeof(double));
             if( true == topo->weighted ) {
diff --git a/ompi/mca/topo/treematch/treematch/tm_kpartitioning.c b/ompi/mca/topo/treematch/treematch/tm_kpartitioning.c
index 3aaed6a..017f3ed 100644
--- a/ompi/mca/topo/treematch/treematch/tm_kpartitioning.c
+++ b/ompi/mca/topo/treematch/treematch/tm_kpartitioning.c
@@ -426,8 +426,7 @@ tree_t *kpartition_build_tree_from_topology(tm_topology_t *topology,double **com
   verbose_level = get_verbose_level();

   if(verbose_level>=INFO)
-    printf("Number of constraints: %d\n", nb_constraints);
-  printf("Number of constraints: %d, N=%d\n", nb_constraints, N);
+    printf("Number of constraints: %d, N=%d\n", nb_constraints, N);

   nb_cores=nb_processing_units(topology);

diff --git a/ompi/mca/topo/treematch/treematch/tm_tree.c b/ompi/mca/topo/treematch/treematch/tm_tree.c
index 0f41958..3305e2d 100644
--- a/ompi/mca/topo/treematch/treematch/tm_tree.c
+++ b/ompi/mca/topo/treematch/treematch/tm_tree.c
@@ -1611,7 +1611,8 @@ tree_t * build_tree_from_topology(tm_topology_t *topology, double **com_mat, int

   nb_constraints = check_constraints (topology, &constraints);

-  printf("nb_constraints = %d, N= %d; nb_processing units = %d\n",nb_constraints, N, nb_processing_units(topology));
+  if(verbose_level>=INFO)
+    printf("nb_constraints = %d, N= %d; nb_processing units = %d\n",nb_constraints, N, nb_processing_units(topology));

   if(N>nb_constraints){
     if(verbose_level >= CRITICAL){

btw, any reason why you use printf instead of the ompi logging facility ?

[bosilca]

The efforts to fix this issue unfold in https://github.com/bosilca/ompi/tree/topic/treematch