setup a cleanup routine for mvapich collectives.

We allocate some tables at first use that we never released.

diff --git a/src/include/smpi/smpi_interface.h b/src/include/smpi/smpi_interface.h
index e2ea5c5..73dd68d 100644 (file)
--- a/src/include/smpi/smpi_interface.h
+++ b/src/include/smpi/smpi_interface.h
@@ -92,6 +92,8 @@ XBT_PUBLIC_DATA(int (*mpi_coll_barrier_fun)(MPI_Comm comm));
 XBT_PUBLIC(void) coll_help(const char *category, s_mpi_coll_description_t * table);
 XBT_PUBLIC(int) find_coll_description(s_mpi_coll_description_t * table, char *name, const char *desc);
 
+XBT_PUBLIC_DATA(void) (*smpi_coll_cleanup_callback)(void);
+XBT_PUBLIC(void) smpi_coll_cleanup_mvapich2(void);
 extern XBT_PRIVATE double smpi_wtime_sleep;
 extern XBT_PRIVATE double smpi_iprobe_sleep;
 extern XBT_PRIVATE double smpi_test_sleep;

diff --git a/src/smpi/colls/smpi_mvapich2_selector.c b/src/smpi/colls/smpi_mvapich2_selector.c
index b15b5ad..97a876b 100644 (file)
--- a/src/smpi/colls/smpi_mvapich2_selector.c
+++ b/src/smpi/colls/smpi_mvapich2_selector.c
@@ -1006,3 +1006,32 @@ int smpi_coll_tuned_scatter_mvapich2(void *sendbuf,
   return (mpi_errno);
 }
 
+
+void smpi_coll_cleanup_mvapich2(){
+int i=0;
+if(mv2_alltoall_thresholds_table)
+  xbt_free(mv2_alltoall_thresholds_table[i]);
+xbt_free(mv2_alltoall_thresholds_table);
+xbt_free(mv2_size_alltoall_tuning_table);
+xbt_free(mv2_alltoall_table_ppn_conf);
+
+xbt_free(mv2_gather_thresholds_table);
+  if(mv2_allgather_thresholds_table)
+xbt_free(mv2_allgather_thresholds_table[0]);
+xbt_free(mv2_size_allgather_tuning_table);
+xbt_free(mv2_allgather_table_ppn_conf);
+xbt_free(mv2_allgather_thresholds_table);
+
+
+xbt_free(mv2_allgatherv_thresholds_table);
+xbt_free(mv2_reduce_thresholds_table);
+xbt_free(mv2_red_scat_thresholds_table);
+xbt_free(mv2_allreduce_thresholds_table);
+xbt_free(mv2_bcast_thresholds_table);
+if(mv2_scatter_thresholds_table)
+  xbt_free(mv2_scatter_thresholds_table[0]);
+xbt_free(mv2_scatter_thresholds_table);
+xbt_free(mv2_size_scatter_tuning_table);
+xbt_free(mv2_scatter_table_ppn_conf);
+}
+

diff --git a/src/smpi/colls/smpi_mvapich2_selector_stampede.h b/src/smpi/colls/smpi_mvapich2_selector_stampede.h
index d982400..02cf7ac 100644 (file)
--- a/src/smpi/colls/smpi_mvapich2_selector_stampede.h
+++ b/src/smpi/colls/smpi_mvapich2_selector_stampede.h
@@ -49,6 +49,8 @@ static void init_mv2_alltoall_tables_stampede(){
   int agg_table_sum = 0;
   mv2_alltoall_tuning_table **table_ptrs = NULL;
   mv2_alltoall_num_ppn_conf = 3;
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_alltoall_thresholds_table = xbt_malloc(sizeof(mv2_alltoall_tuning_table *)
       * mv2_alltoall_num_ppn_conf);
   table_ptrs = xbt_malloc(sizeof(mv2_alltoall_tuning_table *)
@@ -593,6 +595,8 @@ MV2_Gather_function_ptr MV2_Gather_intra_node_function = NULL;
 
 static void init_mv2_gather_tables_stampede(){
 
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_gather_tuning_table=7;
   mv2_gather_thresholds_table = xbt_malloc(mv2_size_gather_tuning_table*
       sizeof (mv2_gather_tuning_table));
@@ -678,6 +682,8 @@ mv2_allgatherv_tuning_table *mv2_allgatherv_thresholds_table = NULL;
 
 
 static void init_mv2_allgatherv_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_allgatherv_tuning_table = 6;
   mv2_allgatherv_thresholds_table = xbt_malloc(mv2_size_allgatherv_tuning_table *
       sizeof (mv2_allgatherv_tuning_table));
@@ -824,6 +830,8 @@ static  int MPIR_Allreduce_reduce_shmem_MV2( void *sendbuf,
 
 
 static void init_mv2_allreduce_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_allreduce_tuning_table = 8;
   mv2_allreduce_thresholds_table = xbt_malloc(mv2_size_allreduce_tuning_table *
       sizeof (mv2_allreduce_tuning_table));
@@ -1022,6 +1030,8 @@ int mv2_intra_node_knomial_factor = 4;
 
 static void init_mv2_bcast_tables_stampede(){
  //Stampede,
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
         mv2_size_bcast_tuning_table=8;
         mv2_bcast_thresholds_table = xbt_malloc(mv2_size_bcast_tuning_table *
                                                  sizeof (mv2_bcast_tuning_table));
@@ -1288,6 +1298,8 @@ int (*MV2_Reduce_intra_function)( void *sendbuf,
 
 
 static void init_mv2_reduce_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   /*Stampede*/
   mv2_size_reduce_tuning_table = 8;
   mv2_reduce_thresholds_table = xbt_malloc(mv2_size_reduce_tuning_table *
@@ -1530,6 +1542,8 @@ static  int MPIR_Reduce_Scatter_Basic_MV2(void *sendbuf,
 
 
 static void init_mv2_reduce_scatter_tables_stampede(){
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
   mv2_size_red_scat_tuning_table = 6;
   mv2_red_scat_thresholds_table = xbt_malloc(mv2_size_red_scat_tuning_table *
       sizeof (mv2_red_scat_tuning_table));
@@ -1654,7 +1668,9 @@ int MPIR_Scatter_mcst_wrap_MV2(void *sendbuf,
 
 
 static void init_mv2_scatter_tables_stampede(){
-  {
+  if(smpi_coll_cleanup_callback==NULL)
+    smpi_coll_cleanup_callback=&smpi_coll_cleanup_mvapich2;
+
     int agg_table_sum = 0;
     int i;
     mv2_scatter_tuning_table **table_ptrs = NULL;
@@ -1961,6 +1977,6 @@ static void init_mv2_scatter_tables_stampede(){
                 * mv2_size_scatter_tuning_table[i]));
     }
     xbt_free(table_ptrs);
-  }
+  
 }
 

diff --git a/src/smpi/smpi_coll.cpp b/src/smpi/smpi_coll.cpp
index 9f6d9bf..dcf9240 100644 (file)
--- a/src/smpi/smpi_coll.cpp
+++ b/src/smpi/smpi_coll.cpp
@@ -167,6 +167,7 @@ int (*mpi_coll_reduce_fun)(void *buf, void *rbuf, int count, MPI_Datatype dataty
 int (*mpi_coll_reduce_scatter_fun)(void *sbuf, void *rbuf, int *rcounts,MPI_Datatype dtype,MPI_Op  op,MPI_Comm  comm);
 int (*mpi_coll_scatter_fun)(void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype,int root, MPI_Comm comm);
 int (*mpi_coll_barrier_fun)(MPI_Comm comm);
+void (*smpi_coll_cleanup_callback)(void);
 
 
 int smpi_coll_tuned_alltoall_ompi2(void *sendbuf, int sendcount,

diff --git a/src/smpi/smpi_global.cpp b/src/smpi/smpi_global.cpp
index 1da6308..6bbc180 100644 (file)
--- a/src/smpi/smpi_global.cpp
+++ b/src/smpi/smpi_global.cpp
@@ -50,7 +50,6 @@ int* index_to_process_data = NULL;
 extern double smpi_total_benched_time;
 xbt_os_timer_t global_timer;
 MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
-
 MPI_Errhandler *MPI_ERRORS_RETURN = NULL;
 MPI_Errhandler *MPI_ERRORS_ARE_FATAL = NULL;
 MPI_Errhandler *MPI_ERRHANDLER_NULL = NULL;
@@ -471,6 +470,8 @@ void smpi_global_destroy(void)
   if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
     smpi_comm_cleanup_smp(MPI_COMM_WORLD);
     smpi_comm_cleanup_attributes(MPI_COMM_WORLD);
+    if(smpi_coll_cleanup_callback!=NULL)
+      smpi_coll_cleanup_callback();
     xbt_free(MPI_COMM_WORLD);
   }
 
@@ -593,6 +594,7 @@ static void smpi_init_options(){
     int barrier_id = find_coll_description(mpi_coll_barrier_description, xbt_cfg_get_string("smpi/barrier"),"barrier");
     mpi_coll_barrier_fun = (int (*)(MPI_Comm comm)) mpi_coll_barrier_description[barrier_id].coll;
 
+    smpi_coll_cleanup_callback=NULL;
     smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold");
     smpi_running_power = xbt_cfg_get_double("smpi/running-power");
     smpi_privatize_global_variables = xbt_cfg_get_boolean("smpi/privatize-global-variables");