Add demo for nonblocking collectives

mpi/demos/nonblocking_collectives.c

@@ -0,0 +1,47 @@
+#include <stdio.h>
+#include <unistd.h>  // for sleep()
+#include <mpi.h>
+
+int main(int argc, char *argv[])
+{
+  int size, rank;
+  MPI_Init(&argc, &argv);
+  MPI_Comm_size(MPI_COMM_WORLD, &size);
+  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+  printf("Hello from rank %d of %d\n", rank, size);
+  fflush(stdout);
+
+  double t0 = MPI_Wtime();
+
+  // Computing 1
+  usleep(1000 * rank);
+  double data1 = 1.0 * rank;
+
+  // Allreduce
+  double reduced;
+#ifdef NONBLOCKING
+  MPI_Request request;
+  MPI_Iallreduce(&data1, &reduced, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD, &request);
+#else
+  MPI_Allreduce(&data1, &reduced, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+#endif
+
+  // Computing 2 (independent of 1)
+  usleep(1000 * (size - rank));
+  double data2 = 2.0 * rank;
+
+#ifdef NONBLOCKING
+  MPI_Wait(&request, MPI_STATUS_IGNORE);
+#endif
+
+  // Computing 3 (requiring 1 and 2)
+  usleep(1000);
+  double data3 = reduced + data2;
+
+  double t1 = MPI_Wtime();
+
+  printf("Rank %d computing took %.2f ms, data: %6.1f, %6.1f, %6.1f\n", rank, (t1-t0) * 1e3, data1, data2, data3);
+
+  MPI_Finalize();
+}

mpi/docs/06-non-blocking-communication.md

@@ -182,6 +182,10 @@ MPI_Ibcast(`buf`{.input}`fer`{.output}, `count`{.input}, `datatype`{.input}, `ro
 ![](img/green_arrow.png){width=1%} (Computation) work 2, not
 involving data in the `MPI_Allreduce` operation
 
+# Non-blocking collectives
+
+- Demo: `nonblocking_collectives.c`
+
 
 # Summary