Align eager RDMA buffer so that last byte of the buffer is on the last byte of

the CPU cache line. Improves zero byte latency a little bit because of L1 cache
miss reduction.

This commit was SVN r11465.

ompi/mca/btl/openib/btl_openib_component.c

@@ -562,7 +562,7 @@ btl_openib_component_init(int *num_btl_modules,
             sizeof(mca_btl_openib_footer_t) + 
             openib_btl->super.btl_eager_limit;
 	
-        openib_btl->eager_rdma_frag_size = length;
+        openib_btl->eager_rdma_frag_size = (length + mca_btl_openib_component.buffer_alignment) & ~(mca_btl_openib_component.buffer_alignment-1);
  
         ompi_free_list_init_ex(&openib_btl->send_free_eager,
                             length,
@@ -879,7 +879,7 @@ static int btl_openib_component_progress(void)
                         size + sizeof(mca_btl_openib_footer_t));
                 frag->segment.seg_addr.pval = ((unsigned char* )frag->hdr) + 
                     sizeof(mca_btl_openib_header_t);
-                
+
                 ret = btl_openib_handle_incoming_hp(openib_btl,
                         frag->endpoint, frag, 
                         size - sizeof(mca_btl_openib_footer_t));
@@ -1011,7 +1011,7 @@ static int btl_openib_component_progress(void)
                 break;
             }
         }
-        
+
         ne=ibv_poll_cq(openib_btl->ib_cq_lp, 1, &wc );
         if(ne < 0){ 
             BTL_ERROR(("error polling LP CQ with %d errno says %s", ne, strerror(errno))); 

ompi/mca/btl/openib/btl_openib_endpoint.c

@@ -1256,12 +1256,17 @@ void mca_btl_openib_endpoint_connect_eager_rdma(
 
     buf = openib_btl->super.btl_mpool->mpool_alloc(openib_btl->super.btl_mpool,
             openib_btl->eager_rdma_frag_size * 
-            mca_btl_openib_component.eager_rdma_num, 0, 0,
+            mca_btl_openib_component.eager_rdma_num +
+            mca_btl_openib_component.buffer_alignment +
+            sizeof(mca_btl_openib_recv_frag_eager_t), 0, 0,
             (mca_mpool_base_registration_t**)&endpoint->eager_rdma_local.reg);
 
     if(!buf)
        goto unlock_rdma_local;
 
+    buf = (char*)(((uintptr_t)buf+mca_btl_openib_component.buffer_alignment) & ~(mca_btl_openib_component.buffer_alignment-1));
+    buf = buf + openib_btl->eager_rdma_frag_size - sizeof(mca_btl_openib_footer_t) - openib_btl->super.btl_eager_limit - sizeof(mca_btl_openib_header_t) - sizeof(mca_btl_openib_frag_t);
+
     for(i = 0; i < mca_btl_openib_component.eager_rdma_num; i++) {
         ompi_free_list_item_t *item = (ompi_free_list_item_t *)(buf +
                 i*openib_btl->eager_rdma_frag_size);

ompi/mca/btl/openib/btl_openib_frag.h

@@ -78,8 +78,10 @@ typedef enum mca_btl_openib_frag_type_t mca_btl_openib_frag_type_t;
  */
 struct mca_btl_openib_frag_t {
     mca_btl_base_descriptor_t base; 
-    mca_btl_base_segment_t segment; 
     struct mca_btl_base_endpoint_t *endpoint; 
+    mca_btl_openib_footer_t *ftr;
+    mca_btl_openib_header_t *hdr;
+    mca_btl_base_segment_t segment; 
     size_t size; 
     int rc;
     mca_btl_openib_frag_type_t type; 
@@ -89,8 +91,6 @@ struct mca_btl_openib_frag_t {
     } wr_desc;  
     struct ibv_sge sg_entry;  
     struct ibv_mr *mr; 
-    mca_btl_openib_header_t *hdr;
-    mca_btl_openib_footer_t *ftr;
     mca_mpool_openib_registration_t * openib_reg; 
 }; 
 typedef struct mca_btl_openib_frag_t mca_btl_openib_frag_t;