Implement the PUT protocol over MX. The send/receive approach give the best

performance on a 2G Myrinet card, as it look like pipelining the messages
by 1M is faster than a simple send/receive. However, when using a 10G card
the send/receive will limit the maximum bandwidth to 2.5Gbs. The reason is
the scarce bus resources that have to be shared between the Myrinet hardware
and the memcpy operation. The PUT protocol remove the memcpy, we now have a 
true zero-copy mechanism. But, there is no pipelining yet as it look like the
RDMA pipeline somehow disappeared from the OB1 PML ...

This commit was SVN r12925.

ompi/mca/btl/mx/btl_mx.c

@@ -142,7 +142,7 @@ int mca_btl_mx_register( struct mca_btl_base_module_t* btl,
         mx_segment.segment_ptr = frag->base.des_dst->seg_addr.pval;
         mx_segment.segment_length = frag->base.des_dst->seg_len;
         mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1, (uint64_t)tag,
-                              0xffffffffffffffffULL,
+                              BTL_MX_RECV_MASK,
                               frag, &(frag->mx_request) );
         if( MX_SUCCESS != mx_return ) {
             return OMPI_ERROR;
@@ -284,6 +284,44 @@ mca_btl_base_descriptor_t* mca_btl_mx_prepare_src( struct mca_btl_base_module_t*
     return &frag->base;
 }
 
+/**
+ * Initiate an asynchronous put. 
+ *
+ * @param btl (IN)         BTL module
+ * @param endpoint (IN)    BTL addressing information
+ * @param descriptor (IN)  Description of the data to be transferred
+ */
+static int mca_btl_mx_put( struct mca_btl_base_module_t* btl,
+                           struct mca_btl_base_endpoint_t* endpoint,
+                           struct mca_btl_base_descriptor_t* descriptor )
+{
+    mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
+    mca_btl_mx_frag_t* frag = (mca_btl_mx_frag_t*)descriptor;
+    mx_segment_t mx_segment;
+    mx_return_t mx_return;
+
+    if( MCA_BTL_MX_CONNECTED != ((mca_btl_mx_endpoint_t*)endpoint)->status ) {
+        if( MCA_BTL_MX_NOT_REACHEABLE == ((mca_btl_mx_endpoint_t*)endpoint)->status )
+            return OMPI_ERROR;
+        if( OMPI_SUCCESS != mca_btl_mx_proc_connect( (mca_btl_mx_endpoint_t*)endpoint ) )
+            return OMPI_ERROR;
+    }
+
+    frag->endpoint  = endpoint;
+    frag->tag       = 0xff;
+
+    mx_segment.segment_ptr    = descriptor->des_src[1].seg_addr.pval;
+    mx_segment.segment_length = descriptor->des_src[1].seg_len;
+
+    mx_return = mx_isend( mx_btl->mx_endpoint, &mx_segment, 1/*descriptor->des_src_cnt*/,
+                          endpoint->mx_peer_addr,
+                          descriptor->des_dst[0].seg_key.key64, frag, &frag->mx_request );
+    if( MX_SUCCESS != mx_return ) {
+        opal_output( 0, "mx_isend fails with error %s\n", mx_strerror(mx_return) );
+        return OMPI_ERROR;
+    }
+    return OMPI_SUCCESS;
+}
 
 /**
  * Prepare a descriptor for send/rdma using the supplied
@@ -306,7 +344,10 @@ mca_btl_base_descriptor_t* mca_btl_mx_prepare_dst( struct mca_btl_base_module_t*
                                                    size_t reserve,
                                                    size_t* size)
 {
+    mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
     mca_btl_mx_frag_t* frag;
+    mx_return_t mx_return;
+    mx_segment_t mx_segment;
     int rc;
 
     MCA_BTL_MX_FRAG_ALLOC_USER(btl, frag, rc);
@@ -314,8 +355,19 @@ mca_btl_base_descriptor_t* mca_btl_mx_prepare_dst( struct mca_btl_base_module_t*
         return NULL;
     }
 
-    frag->segment[0].seg_len = *size;
+    frag->segment[0].seg_len       = *size;
     frag->segment[0].seg_addr.pval = convertor->pBaseBuf + convertor->bConverted;
+    frag->segment[0].seg_key.key64 = (uint64_t)(intptr_t)frag;
+
+    mx_segment.segment_ptr    = frag->segment[0].seg_addr.pval;
+    mx_segment.segment_length = frag->segment[0].seg_len;
+    mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1, frag->segment[0].seg_key.key64, 
+                          BTL_MX_PUT_MASK, NULL, &(frag->mx_request) );
+    if( MX_SUCCESS != mx_return ) {
+        opal_output( 0, "Fail to re-register a fragment with the MX NIC ...\n" );
+        MCA_BTL_MX_FRAG_RETURN( btl, frag );
+        return NULL;
+    }
 
     frag->base.des_src = NULL;
     frag->base.des_src_cnt = 0;
@@ -342,7 +394,7 @@ int mca_btl_mx_send( struct mca_btl_base_module_t* btl,
                      mca_btl_base_tag_t tag )
    
 {
-    mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*) btl;
+    mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
     mca_btl_mx_frag_t* frag = (mca_btl_mx_frag_t*)descriptor;
     mx_segment_t mx_segment[2];
     mx_return_t mx_return;
@@ -424,7 +476,7 @@ mca_btl_mx_module_t mca_btl_mx_module = {
         mca_btl_mx_prepare_src,
         mca_btl_mx_prepare_dst,
         mca_btl_mx_send,
-        NULL, /* put */
+        mca_btl_mx_put, /* put */
         NULL, /* get */
         mca_btl_base_dump,
         NULL, /* mpool */

ompi/mca/btl/mx/btl_mx.h

@@ -43,6 +43,12 @@
 extern "C" {
 #endif
 
+/**
+ * The mask used for receive and for the PUT protocol
+ */
+#define BTL_MX_RECV_MASK 0xffffffffffffffffULL
+#define BTL_MX_PUT_MASK  0xffffffffffffffffULL
+
 /**
  * MX BTL component.
  */

ompi/mca/btl/mx/btl_mx_component.c

@@ -447,7 +447,7 @@ int mca_btl_mx_component_progress()
             mx_segment.segment_ptr = frag->base.des_dst->seg_addr.pval;
             mx_segment.segment_length = mca_btl_mx_module.super.btl_eager_limit;
             mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1, (uint64_t)frag->tag, 
-                                  (uint64_t)0xffffffffffffffffULL,
+                                  BTL_MX_RECV_MASK,
                                   frag, &(frag->mx_request) );
             if( MX_SUCCESS != mx_return ) {
                 opal_output( 0, "Fail to re-register a fragment with the MX NIC ...\n" );