MPI

Homework/Q1/Makefile

+# Makefile for MPI program
+
+# Source file
+SRC = prefix_sum_mpi.cpp
+
+# number of process
+NP = 4
+
+# Compiler
+CXX = mpic++
+
+# Executable name
+EXE = ./output
+
+all: $(EXE)
+
+$(EXE): $(SRC)
+	$(CXX) -o $(EXE) $(SRC)
+
+clean:
+	rm -f $(EXE)
+
+run:
+	mpirun -np $(NP) -f hosts $(EXE)
\ No newline at end of file

Homework/Q1/hosts

+master 
+slave1 user=mpiuser 
+slave2 user=mpiuser

Homework/Q1/prefix_sum_mpi.cpp

+#include <mpi.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <math.h>  
+#include <stdbool.h>
+
+void prefix_mpi(int* block_array, int block_size, int* block_prefix, MPI_Comm communicator)
+{
+    int my_rank, com_size;
+    MPI_Comm_rank(communicator, &my_rank);
+    MPI_Comm_size(communicator, &com_size);
+
+    int* local_prefix = (int*)malloc(block_size * sizeof(int));
+    local_prefix[0] = block_array[0];
+
+    // Calculate local prefix sum
+    for (int i = 1; i < block_size; i++) {
+        local_prefix[i] = local_prefix[i - 1] + block_array[i];
+    }
+
+    // Perform prefix sum across processes
+    int* prefix_sums = (int*)malloc(com_size * sizeof(int));
+    MPI_Gather(&local_prefix[block_size - 1], 1, MPI_INT,prefix_sums, 1, MPI_INT, 0, communicator);
+
+    int global_accum = 0;
+
+    // Calculate the total prefix sum from prefix sums of all processes
+    if (my_rank == 0) {
+        for (int i = 0; i < com_size; i++) {
+            global_accum += prefix_sums[i];
+            prefix_sums[i] = global_accum;
+        }
+    }
+
+    // Broadcast the global prefix sum to all processes
+    MPI_Bcast(prefix_sums, com_size, MPI_INT, 0, communicator);
+
+    // Adjust local prefix sums with the global prefix sum
+    for (int i = 0; i < block_size; i++) {
+        block_prefix[i] = local_prefix[i] + prefix_sums[my_rank] - local_prefix[block_size - 1];
+    }
+
+    free(local_prefix);
+    free(prefix_sums);
+}
+
+
+int main(int argc, char** args)
+{
+    // Initialize MPI
+    MPI_Init(&argc, &args);
+
+    int my_rank;
+    int com_size;
+    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
+    MPI_Comm_size(MPI_COMM_WORLD, &com_size);
+
+    // Set the total array size and adjust it for an even distribution among processes
+    int total_array_size = 64;
+    if (total_array_size % com_size != 0)
+        total_array_size = (total_array_size / com_size + 1) * com_size;
+
+    // Calculate block size for each process
+    int block_size = total_array_size / com_size;
+
+    // Arrays to store the total array and its prefix sum
+    int* total_array = NULL;
+    int* total_prefix = NULL;
+
+    // Master process (rank 0) initializes the total array with random values
+    if (my_rank == 0)
+    {
+        total_array = (int*)malloc(total_array_size * sizeof(int));
+        total_prefix = (int*)malloc(total_array_size * sizeof(int));
+        for (int i = 0; i < total_array_size; i++)
+            total_array[i] = rand() % 11;
+
+        // Display the generated total array
+        printf("Total Array: ");
+        for (int i = 0; i < total_array_size; i++)
+            printf("%d ", total_array[i]);
+        printf("\n\n");
+    }
+
+    // Arrays to store the block array and its prefix sum for each process
+    int* block_array = (int*)malloc(block_size * sizeof(int));
+    int* block_prefix = (int*)malloc(block_size * sizeof(int));
+
+    // Scatter the total array among processes
+    MPI_Scatter(total_array, block_size, MPI_INT,
+        block_array, block_size, MPI_INT, 0, MPI_COMM_WORLD);
+
+    // Calculate the prefix sum for the block array
+    prefix_mpi(block_array, block_size, block_prefix, MPI_COMM_WORLD);
+
+    // Gather the block prefixes back to the total prefix array on the master process
+    MPI_Gather(block_prefix, block_size, MPI_INT,
+        total_prefix, block_size, MPI_INT, 0, MPI_COMM_WORLD);
+
+    // Validate the computed prefix sum against the sequential calculation on the master process
+    int accum = 0;
+    if (my_rank == 0)
+    {
+    
+      // Display block prefixes for all processes
+        printf("total Prefixes:\n");
+            for (int j = 0; j < total_array_size; j++)
+                printf("%d ", total_prefix[j]);
+            printf("\n");
+    
+    
+        // Display block prefixes for each process
+        printf("\n for each process :\n");
+        for (int i = 0; i < com_size; i++)
+        {
+            printf("Process %d: ", i);
+            for (int j = 0; j < block_size; j++)
+                printf("%d ", total_prefix[i * block_size + j]);
+            printf("\n");
+        }
+
+        // Validate the computed prefix sum against the sequential calculation
+        for (int i = 1; i < total_array_size; i++)
+        {
+            accum += total_array[i - 1];
+            if (total_prefix[i-1] != accum)
+                printf("Error at index %i: %i expected, %i computed\n", i, accum, total_prefix[i-1]);
+        }
+
+        
+        printf("Test completed!\n"); 
+        free(total_array);
+        free(total_prefix);
+    }
+
+    // Free allocated memory for block arrays
+    free(block_array);
+    free(block_prefix);
+
+    // Finalize MPI
+    MPI_Finalize();
+
+    return 0;
+}

Homework/q2/Makefile

+# Makefile for MPI program
+
+# Source file
+SRC = Mpi_reduce.cpp
+
+# number of process
+NP = 4
+
+# Compiler
+CXX = mpic++
+
+# Executable name
+EXE = ./output
+
+all: $(EXE)
+
+$(EXE): $(SRC)
+	$(CXX) -o $(EXE) $(SRC)
+
+clean:
+	rm -f $(EXE)
+
+run:
+	mpirun -np $(NP) -f hosts $(EXE)
\ No newline at end of file

Homework/q2/Mpi_reduce.cpp

+#include <mpi.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <math.h>  
+#include <stdbool.h>
+#include <unistd.h>
+
+
+#include <mpi.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <math.h>
+#include <stdbool.h>
+
+void reduce_tree(int* send_data, int* recv_data, int count, MPI_Comm communicator) {
+    int my_rank, com_size;
+    MPI_Comm_rank(communicator, &my_rank);
+    MPI_Comm_size(communicator, &com_size);
+
+    // Determine the parent and children ranks in the virtual tree
+    int parent_rank = (my_rank - 1) / 2;
+    int left_child_rank = 2 * my_rank + 1;
+    int right_child_rank = 2 * my_rank + 2;
+    
+     if(my_rank){
+      recv_data = (int*)malloc(count * sizeof(int));
+    }
+    
+    memset(recv_data, 0, count * sizeof(int));
+   
+    // Allocate buffers for received data from children
+    int* left_child_data = (int*)malloc(count * sizeof(int));
+    int* right_child_data = (int*)malloc(count * sizeof(int));
+
+    // Receive data from left child if it exists
+    if (left_child_rank < com_size) {
+        MPI_Recv(left_child_data, count, MPI_INT, left_child_rank, 0, communicator, MPI_STATUS_IGNORE);
+        
+        for (int i = 0; i < count; i++) {
+        recv_data[i] = left_child_data[i] + recv_data[i];
+    }   
+    }
+
+    // Receive data from right child if it exists
+    if (right_child_rank < com_size) {
+        MPI_Recv(right_child_data, count, MPI_INT, right_child_rank, 0, communicator, MPI_STATUS_IGNORE);
+     
+        for (int i = 0; i < count; i++) {
+        recv_data[i] = right_child_data[i] + recv_data[i];
+    }
+    
+  
+    }
+   
+    // Combine data with own data (e.g., using addition)
+    for (int i = 0; i < count; i++) {
+        recv_data[i] = recv_data[i] + send_data[i];
+    }
+
+    // Send the result to the parent if not the root process
+    if (my_rank != 0) {
+        MPI_Send(recv_data, count, MPI_INT, parent_rank, 0, communicator);
+        
+    }
+
+    // Free allocated memory
+    free(left_child_data); 
+    free(right_child_data);
+}
+
+
+
+
+
+void reduce_sequential(
+    int* send_data,
+    int* recv_data,
+    int count,
+    MPI_Comm communicator)
+{
+    int my_rank;
+    int com_size;
+    MPI_Comm_rank(communicator, &my_rank);
+    MPI_Comm_size(communicator, &com_size);
+
+    int* gather_buffer = NULL;
+    if (my_rank == 0)
+    {
+        gather_buffer = (int*) calloc(count * com_size, sizeof(int));
+    }
+
+    MPI_Gather(send_data, count, MPI_INT, gather_buffer, count, MPI_INT, 0, communicator);
+    
+    if (my_rank == 0)
+    {
+        memset(recv_data, 0, count * sizeof(int));
+        for (int p = 0; p < com_size; p++)
+            for (int i = 0; i < count; i++)
+                recv_data[i] += gather_buffer[count * p + i];
+        free(gather_buffer);
+    }
+ 
+}
+
+
+
+int main(int argc, char** args)
+{
+
+    MPI_Init(&argc, &args);
+    int max_value = 10000000;
+    int* recv_array_tree = NULL;
+    int* recv_array_sequential = NULL;
+
+    int my_rank;
+    double sequential_time = 0.0;
+    double tree_time = 0.0;
+    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
+
+    for (int count = 100; count <= max_value; count *= 10) {
+        if (my_rank == 0) {
+            recv_array_tree = (int*) malloc(count * sizeof(int));
+            recv_array_sequential = (int*) malloc(count * sizeof(int));
+        }
+
+        int* send_array = (int*) malloc(count * sizeof(int));
+        for (int i = 0; i < count; i++)
+            send_array[i] = my_rank;
+
+        double time1 = 0.0;
+        double time2 = 0.0;
+        MPI_Barrier(MPI_COMM_WORLD);
+        time1 -= MPI_Wtime();
+        reduce_tree(send_array, recv_array_tree, count, MPI_COMM_WORLD);
+        time1 += MPI_Wtime();
+        tree_time = time1;
+
+        MPI_Barrier(MPI_COMM_WORLD);
+        time2 -= MPI_Wtime();
+        reduce_sequential(send_array, recv_array_sequential, count, MPI_COMM_WORLD);
+        time2 += MPI_Wtime();
+        sequential_time = time2;
+
+        if (my_rank == 0) {
+            
+            printf("\nsize of array : %d \n", count);
+            printf("Processing time for reduce tree is: %lf\n", tree_time);
+            printf("Processing time for reduce sequential is: %lf\n\n", sequential_time);
+            printf("-------------------------------------------------------------\n\n");
+            for (int i = 0; i < count; i++) {
+                if (recv_array_tree[i] != recv_array_sequential[i]) {
+                    printf("At index %i: reduce_tree is %i, reduce_sequential is %i\n",
+                        i, recv_array_tree[i], recv_array_sequential[i]);
+                }
+            }
+ 
+            free(recv_array_tree);
+            free(recv_array_sequential);
+        }
+        
+        free(send_array);
+    }
+
+    MPI_Finalize();
+    return 0;
+    
+    
+}
+
+
+ 

Homework/q2/hosts

+master 
+slave1 user=mpiuser 
+slave2 user=mpiuser