Merge pull request pmodels#7637 from hzhou/2510_ofi_pipeline

ch4/ofi: force allocating cells in the pipeline algorithm

Approved-by: Ken Raffenetti

Author: hzhou

src/mpid/ch4/netmod/ofi/ofi_init.c

@@ -475,6 +475,16 @@ categories :
       description : >-
         Specifies the number of chunk buffers for pipeline data transfer.
 
+    - name        : MPIR_CVAR_CH4_OFI_PIPELINE_MAX_CHUNKS
+      category    : CH4_OFI
+      type        : int
+      default     : 1024
+      class       : none
+      verbosity   : MPI_T_VERBOSITY_USER_BASIC
+      scope       : MPI_T_SCOPE_LOCAL
+      description : >-
+        Specifies the max number of chunk buffers to be reserved for pipeline data transfer.
+
     - name        : MPIR_CVAR_CH4_OFI_GPU_SEND_ENGINE_TYPE
       category    : CH4_OFI
       type        : enum
@@ -1509,7 +1519,7 @@ int MPIDI_OFI_init_per_vci(int vci)
     /* Create chunk buffer pool (for pipeline etc.) */
     mpi_errno = MPIDU_genq_private_pool_create(MPIR_CVAR_CH4_OFI_PIPELINE_CHUNK_SZ,
                                                MPIR_CVAR_CH4_OFI_PIPELINE_NUM_CHUNKS,
-                                               MPIR_CVAR_CH4_OFI_PIPELINE_NUM_CHUNKS,
+                                               MPIR_CVAR_CH4_OFI_PIPELINE_MAX_CHUNKS,
                                                host_alloc_registered,
                                                host_free_registered,
                                                &MPIDI_OFI_global.per_vci[vci].pipeline_pool);

src/mpid/ch4/netmod/ofi/ofi_pipeline.c

@@ -197,8 +197,8 @@ static int pipeline_send_poll(MPIX_Async_thing thing)
         }
 
         /* alloc a chunk */
-        MPIDU_genq_private_pool_alloc_cell(MPIDI_OFI_global.per_vci[p->vci_local].pipeline_pool,
-                                           &chunk_buf);
+        MPIDU_genq_private_pool_force_alloc_cell(MPIDI_OFI_global.
+                                                 per_vci[p->vci_local].pipeline_pool, &chunk_buf);
         if (!chunk_buf) {
             goto fn_exit;
         }
@@ -371,8 +371,8 @@ static int pipeline_recv_poll(MPIX_Async_thing thing)
         }
 
         /* alloc a chunk */
-        MPIDU_genq_private_pool_alloc_cell(MPIDI_OFI_global.per_vci[p->vci_local].pipeline_pool,
-                                           &chunk_buf);
+        MPIDU_genq_private_pool_force_alloc_cell(MPIDI_OFI_global.
+                                                 per_vci[p->vci_local].pipeline_pool, &chunk_buf);
         if (!chunk_buf) {
             goto fn_exit;
         }

src/mpid/common/genq/mpidu_genq_private_pool.c

@@ -43,7 +43,9 @@ typedef struct MPIDU_genq_private_pool {
     MPIDU_genq_free_fn free_fn;
 
     intptr_t num_blocks;
-    intptr_t max_num_blocks;
+    intptr_t max_num_blocks;    /* as cells are alloc'ed, the pool may grow up to max_num_blocks.
+                                 * NOTE: MPIDU_genq_private_pool_force_alloc_cell may increase max_num_blocks */
+    intptr_t min_num_blocks;    /* as cells are freed, the pool may shrink back to min_num_blocks */
     cell_block_s *cell_blocks;
     cell_block_s *free_blocks_head;
     cell_block_s *free_blocks_tail;
@@ -69,8 +71,10 @@ int MPIDU_genq_private_pool_create(intptr_t cell_size, intptr_t num_cells_in_blo
     if (max_num_cells == 0) {
         /* 0 means unlimited */
         pool_obj->max_num_blocks = 0;
+        pool_obj->min_num_blocks = 1;
     } else {
         pool_obj->max_num_blocks = max_num_cells / num_cells_in_block;
+        pool_obj->min_num_blocks = pool_obj->max_num_blocks;
     }
 
     pool_obj->malloc_fn = malloc_fn;
@@ -278,9 +282,9 @@ int MPIDU_genq_private_pool_free_cell(MPIDU_genq_private_pool_t pool, void *cell
     if (block->num_used_cells == pool_obj->num_cells_in_block - 1) {
         append_free_blocks(pool_obj, block);
     } else if (block->num_used_cells == 0) {
-        /* Avoid frequent re-allocation by preserving the last block.
+        /* Avoid frequent re-allocation by preserving a minimum number of blocks.
          * All blocks will be freed when the pool is destroyed */
-        if (pool_obj->num_blocks > 1 && pool_obj->max_num_blocks == 0) {
+        if (pool_obj->num_blocks > pool_obj->min_num_blocks) {
             remove_free_blocks(pool_obj, block);
             DL_DELETE(pool_obj->cell_blocks, block);
             cell_block_free(pool_obj, block);