Merge pull request #5230 from hjelmn/v3.1.x_dev_shm

Preferentially use /dev/shm on Linux

Author: bwbarrett

ompi/mca/osc/rdma/osc_rdma.h

@@ -106,6 +106,9 @@ struct ompi_osc_rdma_component_t {
 
     /** aggregation free list */
     opal_free_list_t aggregate;
+
+    /** directory where to place backing files */
+    char *backing_directory;
 };
 typedef struct ompi_osc_rdma_component_t ompi_osc_rdma_component_t;
 

ompi/mca/osc/rdma/osc_rdma_component.c

@@ -268,6 +268,18 @@ static int ompi_osc_rdma_component_register (void)
                                             MCA_BASE_VAR_SCOPE_GROUP, &ompi_osc_rdma_mtl_names);
     free(description_str);
 
+    if (0 == access ("/dev/shm", W_OK)) {
+        mca_osc_rdma_component.backing_directory = "/dev/shm";
+    } else {
+        mca_osc_rdma_component.backing_directory = ompi_process_info.proc_session_dir;
+    }
+
+    (void) mca_base_component_var_register (&mca_osc_rdma_component.super.osc_version, "backing_directory",
+                                            "Directory to place backing files for memory windows. "
+                                            "This directory should be on a local filesystem such as /tmp or "
+                                            "/dev/shm (default: (linux) /dev/shm, (others) session directory)",
+                                            MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0, OPAL_INFO_LVL_3,
+                                            MCA_BASE_VAR_SCOPE_READONLY, &mca_osc_rdma_component.backing_directory);
 
     /* register performance variables */
 
@@ -602,9 +614,9 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
         }
 
         /* allocate the shared memory segment */
-        ret = asprintf (&data_file, "%s"OPAL_PATH_SEP"window_%d.%s",
-                        ompi_process_info.job_session_dir, ompi_comm_get_cid (module->comm),
-                        ompi_process_info.nodename);
+        ret = asprintf (&data_file, "%s" OPAL_PATH_SEP "osc_rdma.%s.%x.%d",
+                        mca_osc_rdma_component.backing_directory, ompi_process_info.nodename,
+                        OMPI_PROC_MY_NAME->jobid, ompi_comm_get_cid(module->comm));
         if (0 > ret) {
             ret = OMPI_ERR_OUT_OF_RESOURCE;
             break;

ompi/mca/osc/sm/osc_sm.h

@@ -61,6 +61,8 @@ typedef struct ompi_osc_sm_node_state_t ompi_osc_sm_node_state_t;
 
 struct ompi_osc_sm_component_t {
     ompi_osc_base_component_t super;
+
+    char *backing_directory;
 };
 typedef struct ompi_osc_sm_component_t ompi_osc_sm_component_t;
 OMPI_DECLSPEC extern ompi_osc_sm_component_t mca_osc_sm_component;

ompi/mca/osc/sm/osc_sm_component.c

@@ -36,6 +36,7 @@ static int component_finalize(void);
 static int component_query(struct ompi_win_t *win, void **base, size_t size, int disp_unit,
                            struct ompi_communicator_t *comm, struct opal_info_t *info,
                            int flavor);
+static int component_register (void);
 static int component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit,
                             struct ompi_communicator_t *comm, struct opal_info_t *info,
                             int flavor, int *model);
@@ -51,6 +52,7 @@ ompi_osc_sm_component_t mca_osc_sm_component = {
             MCA_BASE_MAKE_VERSION(component, OMPI_MAJOR_VERSION, OMPI_MINOR_VERSION,
                                   OMPI_RELEASE_VERSION),
             .mca_open_component = component_open,
+            .mca_register_component_params = component_register,
         },
         .osc_data = { /* mca_base_component_data */
             /* The component is not checkpoint ready */
@@ -105,6 +107,23 @@ ompi_osc_sm_module_t ompi_osc_sm_module_template = {
     }
 };
 
+static int component_register (void)
+{
+    if (0 == access ("/dev/shm", W_OK)) {
+        mca_osc_sm_component.backing_directory = "/dev/shm";
+    } else {
+        mca_osc_sm_component.backing_directory = ompi_process_info.proc_session_dir;
+    }
+
+    (void) mca_base_component_var_register (&mca_osc_sm_component.super.osc_version, "backing_directory",
+                                            "Directory to place backing files for shared memory windows. "
+                                            "This directory should be on a local filesystem such as /tmp or "
+                                            "/dev/shm (default: (linux) /dev/shm, (others) session directory)",
+                                            MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0, OPAL_INFO_LVL_3,
+                                            MCA_BASE_VAR_SCOPE_READONLY, &mca_osc_sm_component.backing_directory);
+
+    return OPAL_SUCCESS;
+}
 
 static int
 component_open(void)
@@ -169,6 +188,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
 {
     ompi_osc_sm_module_t *module = NULL;
     int comm_size = ompi_comm_size (comm);
+    bool unlink_needed = false;
     int ret = OMPI_ERROR;
 
     if (OMPI_SUCCESS != (ret = check_win_ok(comm, flavor))) {
@@ -262,10 +282,10 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
         posts_size += OPAL_ALIGN_PAD_AMOUNT(posts_size, 64);
         if (0 == ompi_comm_rank (module->comm)) {
             char *data_file;
-            if (asprintf(&data_file, "%s"OPAL_PATH_SEP"shared_window_%d.%s",
-                         ompi_process_info.proc_session_dir,
-                         ompi_comm_get_cid(module->comm),
-                         ompi_process_info.nodename) < 0) {
+            ret = asprintf (&data_file, "%s" OPAL_PATH_SEP "osc_sm.%s.%x.%d",
+                           mca_osc_sm_component.backing_directory, ompi_process_info.nodename,
+                           OMPI_PROC_MY_NAME->jobid, ompi_comm_get_cid(module->comm));
+            if (ret < 0) {
                 return OMPI_ERR_OUT_OF_RESOURCE;
             }
 
@@ -274,6 +294,8 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
             if (OPAL_SUCCESS != ret) {
                 goto error;
             }
+
+            unlink_needed = true;
         }
 
 	ret = module->comm->c_coll->coll_bcast (&module->seg_ds, sizeof (module->seg_ds), MPI_BYTE, 0,
@@ -287,6 +309,17 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
 	    goto error;
 	}
 
+	ret = module->comm->c_coll->coll_bcast (&module->seg_ds, sizeof (module->seg_ds), MPI_BYTE, 0,
+					       module->comm, module->comm->c_coll->coll_bcast_module);
+	if (OMPI_SUCCESS != ret) {
+	    goto error;
+	}
+
+        if (0 == ompi_comm_rank (module->comm)) {
+            opal_shmem_unlink (&module->seg_ds);
+            unlink_needed = false;
+        }
+
         module->sizes = malloc(sizeof(size_t) * comm_size);
         if (NULL == module->sizes) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
         module->bases = malloc(sizeof(void*) * comm_size);
@@ -399,6 +432,11 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
     return OMPI_SUCCESS;
 
  error:
+
+    if (0 == ompi_comm_rank (module->comm) && unlink_needed) {
+        opal_shmem_unlink (&module->seg_ds);
+    }
+
     ompi_osc_sm_free (win);
 
     return ret;
@@ -477,10 +515,6 @@ ompi_osc_sm_free(struct ompi_win_t *win)
         module->comm->c_coll->coll_barrier(module->comm,
                                           module->comm->c_coll->coll_barrier_module);
 
-        if (0 == ompi_comm_rank (module->comm)) {
-            opal_shmem_unlink (&module->seg_ds);
-        }
-
 	opal_shmem_segment_detach (&module->seg_ds);
     } else {
         free(module->node_states);