Update 10-further-mpi-topics.md

mpi/docs/10-further-mpi-topics.md

@@ -12,3 +12,60 @@ lang:   en
 - Intercommunicators
 - Point-to-point communication modes
 - Dynamic process creation
+
+
+
+# Persistent communication {.section}
+
+# Persistent communication
+
+- Often a communication with same argument list is repeatedly executed
+- It may be possible to optimize such pattern by persistent communication requests
+    - Can be thought as a ”communication port”
+- Three separate phases:
+    1. Inititation of communication
+    2. Starting of communication
+    3. Completing communication
+- Recently published MPI 4.0 includes also persistent collectives
+    - Not supported by all implementations yet
+
+# Persistent communication
+
+- Inititiate communication by creating requests
+    - `MPI_Send_init` and `MPI_Recv_init`
+    - Same arguments as in `MPI_Isend` and `MPI_Irecv`
+- Start communication
+    - `MPI_Start` / `MPI_Startall`
+    - Request or array of requests as argument
+- Complete communication
+    - `MPI_Wait` / `MPI_Waitall`
+    - Same as in standard non-blocking communication
+
+# Persistent point-to-point communication
+
+```c
+MPI_Request recv_req, send_req;
+...
+// Initialize send/request objects
+MPI_Recv_init(buf1, cnt, MPI_DOUBLE, src, tag, MPI_COMM_WORLD, &recv_req);
+MPI_Send_init(buf2, cnt, MPI_DOUBLE, dst, tag, MPI_COMM_WORLD, &send_req);
+for (int i=1; i<BIGNUM; i++){
+// Start communication described by recv_obj and send_obj
+    MPI_Start(&recv_req);
+    MPI_Start(&send_req);
+    // Do work, e.g. update the interior domains
+    ...
+    // Wait for send and receive to complete
+    MPI_Wait(&send_req, MPI_STATUS_IGNORE);
+    MPI_Wait(&recv_req, MPI_STATUS_IGNORE);
+}
+//Clean up the requests
+MPI_Request_free (&recv_req); MPI_Request_free (&send_req);
+```
+
+# Summary
+
+- In persistent communication the communication pattern remains constant
+- All the parameters for the communication are set up in the initialization phase
+    - Communication is started and finalized in separate steps
+- Persistent communication provides optimization opportunities for MPI library