Merge branch '1-mpi-non-blocking-slides' into 'master'

Resolve "MPI: non-blocking slides"

Closes #1

See merge request training/sohpc-training-2021!1

D4-MPI/02-MPI-Non-blocking/demo/blocking-send-recv.c

+#include <mpi.h>
+#include <unistd.h>
+#include <stdio.h>
+
+int main(int argc, char* argv[]) {
+
+    MPI_Init(&argc, &argv);
+
+    int rank, size;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+    double before_time = MPI_Wtime();
+    if (rank == 0) {
+        /* simulate 2 units of work */
+        sleep(2);
+
+        /* send rank to processor 1 */
+        MPI_Ssend(&rank, 1, MPI_INT, 1, 0, MPI_COMM_WORLD);
+
+        /* simulate 4 units of work */
+        sleep(4);
+
+        /* send COMM_WORLD size to processor 1 */
+        MPI_Ssend(&rank, 1, MPI_INT, 1, 1, MPI_COMM_WORLD);
+    } else if (rank == 1) {
+        int recv_rank;
+
+        /* simulate 4 units of work */
+        sleep(4);
+
+        /* send rank to processor 1 */
+        MPI_Recv(&recv_rank, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+
+        /* simulate 2 units of work */
+        sleep(2);
+
+        /* send rank to processor 1 again */
+        MPI_Recv(&recv_rank, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+    }
+
+    printf("Processor %d elapsed time: %lfs\n", rank, MPI_Wtime() - before_time);
+
+    MPI_Finalize();
+    return 0;
+}

D4-MPI/02-MPI-Non-blocking/demo/non-blocking-loop.c

+#include <mpi.h>
+#include <unistd.h>
+#include <stdio.h>
+
+int main(int argc, char* argv[]) {
+    MPI_Init(&argc, &argv);
+
+    int rank;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+    MPI_Request req1, req2;
+
+    double before_time = MPI_Wtime();
+    const int n = 100000;
+    if (rank == 0) {
+        for (int i = 0; i < n; i++) {
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* send rank to processor 1 */
+            MPI_Issend(&rank, 1, MPI_INT, 1, 0, MPI_COMM_WORLD, &req1);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for first communication to finish */
+            MPI_Wait(&req1, MPI_STATUS_IGNORE);
+
+            /* send rank to processor 1 again */
+            MPI_Issend(&rank, 1, MPI_INT, 1, 1, MPI_COMM_WORLD, &req2);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for second communication to finish */
+            MPI_Wait(&req2, MPI_STATUS_IGNORE);
+        }
+    } else if (rank == 1) {
+        for (int i = 0; i < n; i++) {
+            int recv_rank;
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* send rank to processor 1 */
+            MPI_Irecv(&recv_rank, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &req1);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for first communication to finish */
+            MPI_Wait(&req1, MPI_STATUS_IGNORE);
+
+            /* send rank to processor 1 again */
+            MPI_Irecv(&recv_rank, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &req2);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for second communication to finish */
+            MPI_Wait(&req2, MPI_STATUS_IGNORE);
+        }
+    }
+
+    printf("Processor %d elapsed time: %lfs\n", rank, MPI_Wtime() - before_time);
+
+    MPI_Finalize();
+    return 0;
+}

D4-MPI/02-MPI-Non-blocking/demo/non-blocking-send-recv.c

+#include <mpi.h>
+#include <unistd.h>
+#include <stdio.h>
+
+int main(int argc, char* argv[]) {
+
+    MPI_Init(&argc, &argv);
+
+    int rank;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+    MPI_Request req1, req2;
+
+    double before_time = MPI_Wtime();
+    if (rank == 0) {
+        /* simulate 2 units of work */
+        sleep(2);
+
+        /* send rank to processor 1 */
+        MPI_Issend(&rank, 1, MPI_INT, 1, 0, MPI_COMM_WORLD, &req1);
+
+        /* simulate 6 units of work */
+        sleep(4);
+
+        /* wait for first communication to finish */
+        MPI_Wait(&req1, MPI_STATUS_IGNORE);
+
+        /* send rank to processor 1 again */
+        MPI_Issend(&rank, 1, MPI_INT, 1, 1, MPI_COMM_WORLD, &req2);
+
+        /* wait for second communication to finish */
+        MPI_Wait(&req2, MPI_STATUS_IGNORE);
+    } else if (rank == 1) {
+        int recv_rank;
+
+        /* simulate 2 units of work */
+        sleep(4);
+
+        /* send rank to processor 1 */
+        MPI_Irecv(&recv_rank, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &req1);
+
+        /* simulate 6 units of work */
+        sleep(2);
+
+        /* wait for first communication to finish */
+        MPI_Wait(&req1, MPI_STATUS_IGNORE);
+
+        /* send rank to processor 1 again */
+        MPI_Irecv(&recv_rank, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &req2);
+
+        /* wait for second communication to finish */
+        MPI_Wait(&req2, MPI_STATUS_IGNORE);
+    }
+
+    printf("Processor %d elapsed time: %lfs\n", rank, MPI_Wtime() - before_time);
+
+    MPI_Finalize();
+    return 0;
+}

D4-MPI/02-MPI-Non-blocking/demo/persistent-loop.c

+#include <mpi.h>
+#include <unistd.h>
+#include <stdio.h>
+
+int main(int argc, char* argv[]) {
+    MPI_Init(&argc, &argv);
+
+    int rank;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+    MPI_Request req1, req2;
+
+    double before_time = MPI_Wtime();
+    const int n = 100000;
+    if (rank == 0) {
+        MPI_Send_init(&rank, 1, MPI_INT, 1, 0, MPI_COMM_WORLD, &req1);
+        MPI_Send_init(&rank, 1, MPI_INT, 1, 1, MPI_COMM_WORLD, &req2);
+
+        for (int i = 0; i < n; i++) {
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* send rank to processor 1 */
+            MPI_Start(&req1);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for first communication to finish */
+            MPI_Wait(&req1, MPI_STATUS_IGNORE);
+
+            /* send rank to processor 1 again */
+            MPI_Start(&req2);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for second communication to finish */
+            MPI_Wait(&req2, MPI_STATUS_IGNORE);
+        }
+    } else if (rank == 1) {
+        int recv_rank;
+        MPI_Recv_init(&recv_rank, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &req1);
+        MPI_Recv_init(&recv_rank, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &req2);
+
+        for (int i = 0; i < n; i++) {
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* send rank to processor 1 */
+            MPI_Start(&req1);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for first communication to finish */
+            MPI_Wait(&req1, MPI_STATUS_IGNORE);
+
+            /* send rank to processor 1 again */
+            MPI_Start(&req2);
+
+            /* simulate 10 units of work */
+            usleep(10);
+
+            /* wait for second communication to finish */
+            MPI_Wait(&req2, MPI_STATUS_IGNORE);
+        }
+    }
+
+    printf("Processor %d elapsed time: %lfs\n", rank, MPI_Wtime() - before_time);
+
+    MPI_Finalize();
+    return 0;
+}

D4-MPI/02-MPI-Non-blocking/exercise/README.md

-Fill with exercise instructions.
+# Non-blocking Communication Exercises
+
+## Ring Exchange
+
+Modify your ring exchange implementation from the previous set of exercises to
+use non-blocking communication.
+
+Don't expect (much of) a speed up since there is computation to hide latency.
+
+## Heat equation
+
+The provided [heat equation code](./heat_equation/main.c) is a working
+implementation of the two-dimensional heat equation with periodic boundary
+conditions, parallelised in one dimension (the y-dimension). Replace the
+blocking communication pattern with a non-blocking one. Think carefully about
+where to place completion routines.
+
+Compare the execution time of your implementation with the blocking version.
+You should realise a speedup while the grid should still tend towards
+uniformity.
+
+***Note:*** periodic boundary conditions are handled in the x-dimension locally.
+MPI handles periodic boundary conditions in the y-dimension as `rank 0` and
+`rank N-1` exchange boundary data with one another.
+
+***Hint:*** To perform communication independent work before completion
+routines, the evolve function in [evolve.h](./heat_equation/evolve.h) must be
+modified.
+
+## Persistent Heat Equation
+
+Further modify the heat equation code to use persistent communication.

D4-MPI/02-MPI-Non-blocking/solution/README.md

-Fill with exercise instructions.

D4-MPI/02-MPI-Non-blocking/solution/pingpong.c

+#include <mpi.h>
+#include <omp.h>
+#include <stdio.h>
+
+int main(int argc, char **argv) {
+    int i, n;
+    int myRank, uniSize;
+    double time;
+
+    // Must call MPI_Init before any other MPI calls
+    MPI_Init(&argc, &argv);
+    if (argc != 2) {
+        printf("Usage: ./pingPong n\n n is the number of times to run the "
+               "simulation\n");
+        MPI_Abort(MPI_COMM_WORLD, 10);
+    }
+    // We take a command line argument to decide the number of loops to do
+    n = atoi(argv[1]);
+    // Get the size of the communicator
+    MPI_Comm_size(MPI_COMM_WORLD, &uniSize);
+    // Get our rank in this communicator
+    MPI_Comm_rank(MPI_COMM_WORLD, &myRank);
+
+    if (uniSize == 2) {
+        // Initialise the message to 10
+        int msg = 10;
+        // Start the timer, we will take the average for a number of loops
+        time = MPI_Wtime();
+        // time = omp_get_wtime();
+
+        /* declare requests */
+        MPI_Request req[2];
+        for (i = 0; i < n; i++) {
+            if (myRank == 0) {
+                // P 0 increments this message by one and passes it to P 1
+                // (ping)
+                msg++;
+                MPI_Isend(&msg, 1, MPI_INT, 1, 0, MPI_COMM_WORLD, &req[0]);
+                // P 0 also has to receive the msg from P 1
+                MPI_Irecv(&msg, 1, MPI_INT, 1, 1, MPI_COMM_WORLD, &req[1]);
+            }
+            if (myRank == 1) {
+                // P 1 receives the message it increments it by one and passes
+                // it back to P 0
+                MPI_Irecv(&msg, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &req[1]);
+                msg++;
+                MPI_Isend(&msg, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &req[0]);
+            }
+            MPI_Waitall(2, req, MPI_STATUSES_IGNORE);
+        }
+        time = MPI_Wtime() - time;
+        // time = omp_get_wtime()-time;
+
+        // Print the value of msg and the average time
+        if (myRank == 0) {
+            printf("value of msg after %d exchanges is %d\n", n, msg);
+            printf("On average it took %lf seconds\n", time / n);
+            // printf("Resolution of MPI_Wtime = %lf seconds\n", MPI_Wtick());
+        }
+    } else if (myRank == 0) {
+        printf("Ping pong can be played with exactly 2 processes.\n");
+    }
+
+    // Remember to always call MPI_Finalize()
+    MPI_Finalize();
+    return 0;
+}

D4-MPI/03-MPI-Collective-Comms/exercise/README.md

-Fill with exercise instructions.
+# Collective Communication Exercises
+
+## Ordered Hello World
+
+The [hello world](./hello_world/main.c) code provided prints a simple hello
+world message on each process containing its rank. Modify the code so that the
+messages are printed in order of increasing rank.
+
+## Find the Deadlock
+
+The [find the deadlock](./find_the_deadlock/main.c) code provided contains a
+deadlock. The expected behaviour of the program is;
+
+- processor 0 reads 25 integers from a [file](./find_the_deadlock/values.dat),
+- processor 0 broadcasts the integers to the other processors,
+- each processor saves the data in separate files.
+
+You should;
+
+- find and fix the deadlock,
+- correct some other errors in the program (the expected wall time of the
+  corrected program is less than 30 seconds).
+
+## Complex Reduction
+
+The goal of this exercise is to perform a summation of a set of complex numbers
+distributed across several processes. The [skeleton
+code](./complex_reduction/main.c) provided;
+
+- defines the complex number data type,
+- initialises arrays a complex number on each process,
+- implements a function for summing complex numbers,
+- registers the complex number data type with MPI, storing its handle as
+  `MPI_COMPLEX`.
+
+Calculate the sum of the complex numbers using `MPI_Reduce`.

D4-MPI/03-MPI-Collective-Comms/solution/README.md

-Fill with exercise instructions.