Add runtime support to turn off CUDA IPC support.

This commit was SVN r29444.

ompi/mca/btl/smcuda/btl_smcuda.c

@@ -921,7 +921,7 @@ int mca_btl_smcuda_sendi( struct mca_btl_base_module_t* btl,
 
 #if OMPI_CUDA_SUPPORT
     /* Initiate setting up CUDA IPC support. */
-    if (mca_common_cuda_enabled && (IPC_INIT == endpoint->ipcstate)) {
+    if (mca_common_cuda_enabled && (IPC_INIT == endpoint->ipcstate) && mca_btl_smcuda_component.use_cuda_ipc) {
         mca_btl_smcuda_send_cuda_ipc_request(btl, endpoint);
     }
 #endif /* OMPI_CUDA_SUPPORT */
@@ -1004,10 +1004,12 @@ int mca_btl_smcuda_send( struct mca_btl_base_module_t* btl,
         mca_btl_smcuda_component_progress();
     }
 
+#if OMPI_CUDA_SUPPORT
     /* Initiate setting up CUDA IPC support */
-    if (mca_common_cuda_enabled && (IPC_INIT == endpoint->ipcstate)) {
+    if (mca_common_cuda_enabled && (IPC_INIT == endpoint->ipcstate) && mca_btl_smcuda_component.use_cuda_ipc) {
         mca_btl_smcuda_send_cuda_ipc_request(btl, endpoint);
     }
+#endif /* OMPI_CUDA_SUPPORT */
 
     /* available header space */
     frag->hdr->len = frag->segment.base.seg_len;

ompi/mca/btl/smcuda/btl_smcuda.h

@@ -13,7 +13,7 @@
  * Copyright (c) 2009-2010 Cisco Systems, Inc.  All rights reserved.
  * Copyright (c) 2010      Los Alamos National Security, LLC.  
  *                         All rights reserved. 
- * Copyright (c) 2012      NVIDIA Corporation.  All rights reserved.
+ * Copyright (c) 2012-2013 NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -205,6 +205,8 @@ struct mca_btl_smcuda_component_t {
 #if OMPI_CUDA_SUPPORT
     int cuda_ipc_verbose;
     int cuda_ipc_output;
+    int use_cuda_ipc;
+    int use_cuda_ipc_same_gpu;
 #endif /* OMPI_CUDA_SUPPORT */
 };
 typedef struct mca_btl_smcuda_component_t mca_btl_smcuda_component_t;

ompi/mca/btl/smcuda/btl_smcuda_component.c

@@ -170,6 +170,8 @@ static int smcuda_register(void)
     } else {
         mca_btl_smcuda.super.btl_exclusivity = MCA_BTL_EXCLUSIVITY_LOW;
     }
+    mca_btl_smcuda_param_register_int("use_cuda_ipc", 1, &mca_btl_smcuda_component.use_cuda_ipc);
+    mca_btl_smcuda_param_register_int("use_cuda_ipc_same_gpu", 1, &mca_btl_smcuda_component.use_cuda_ipc_same_gpu);
     mca_btl_smcuda_param_register_int("cuda_ipc_verbose", 0, &mca_btl_smcuda_component.cuda_ipc_verbose);
     mca_btl_smcuda_component.cuda_ipc_output = opal_output_open(NULL);
     opal_output_set_verbosity(mca_btl_smcuda_component.cuda_ipc_output, mca_btl_smcuda_component.cuda_ipc_verbose);
@@ -734,12 +736,36 @@ static void btl_smcuda_control(mca_btl_base_module_t* btl,
                 return;
             }
 
-            /* Check for IPC support between devices. If the CUDA API call fails, then
+            /* Check for IPC support between devices. If they are the
-             * just move endpoint into bad state.  No need to send a reply. */
+             * same device and use_cuda_ipc_same_gpu is 1 (default),
-            res = mca_common_cuda_device_can_access_peer(&ipcaccess, mydevnum, ctrlhdr.cudev);
+             * then assume CUDA IPC is possible.  This could be a
-            if (0 != res) {
+             * device running in DEFAULT mode or running under MPS.
-                endpoint->ipcstate = IPC_BAD;
+             * Otherwise, check peer acces to determine CUDA IPC
-                return;
+             * support.  If the CUDA API call fails, then just move
+             * endpoint into bad state.  No need to send a reply. */
+            if (mydevnum == ctrlhdr.cudev) {
+                if (mca_btl_smcuda_component.use_cuda_ipc_same_gpu) {
+                    ipcaccess = 1;
+                } else {
+                    opal_output_verbose(10, mca_btl_smcuda_component.cuda_ipc_output,
+                                        "Analyzed CUDA IPC request: myrank=%d, mydev=%d, peerrank=%d, "
+                                        "peerdev=%d --> Access is disabled by btl_smcuda_use_cuda_ipc_same_gpu",
+                                        endpoint->my_smp_rank, mydevnum, endpoint->peer_smp_rank,
+                                        ctrlhdr.cudev);
+                    endpoint->ipcstate = IPC_BAD;
+                    return;
+                }
+            } else {
+                res = mca_common_cuda_device_can_access_peer(&ipcaccess, mydevnum, ctrlhdr.cudev);
+                if (0 != res) {
+                    opal_output_verbose(10, mca_btl_smcuda_component.cuda_ipc_output,
+                                        "Analyzed CUDA IPC request: myrank=%d, mydev=%d, peerrank=%d, "
+                                        "peerdev=%d --> Access is disabled because peer check failed with err=%d",
+                                        endpoint->my_smp_rank, mydevnum, endpoint->peer_smp_rank,
+                                        ctrlhdr.cudev, res);
+                    endpoint->ipcstate = IPC_BAD;
+                    return;
+                }
             }
 
             assert(endpoint->peer_smp_rank == frag->hdr->my_smp_rank);