Implementing POSIX Barriers and MPI Custom Data Types

Custom POSIX Barrier Implementation

This structure defines a custom barrier for thread synchronization:

typedef struct {
    int total;          // Expected threads
    int count;          // Arrived threads
    int phase;          // Current phase (for reuse)
    pthread_mutex_t mutex;
    pthread_cond_t  cond;
} barrier_t;

Barrier Initialization and Destruction

• Init: Sets the total thread count and initializes synchronization primitives.
• Destroy: Cleans up mutex and condition variables.

Barrier Wait Logic

The barrier_wait function ensures all threads reach the synchronization point before proceeding, utilizing a phase-based approach to allow barrier reuse.

MPI Vector Types for Chessboard Patterns

Using MPI_Type_vector to handle non-contiguous memory layouts in odd-sized matrices (N=111).

#define N 111
int rank;
int mat[N][N];

The exchChess function demonstrates how to create a derived data type to exchange alternating elements between processes efficiently.

Von Neumann Neighborhood (Radius 2)

Optimized boundary exchange using a single MPI_Send/Recv operation:

• Top/Bottom Edges: MPI_Type_vector(2, NCOLS/C_PARTITIONS, NCOLS/C_PARTITIONS+4, MPI_INT, &UpDownDataType);
• Left/Right Edges: MPI_Type_vector(NROWS/R_PARTITIONS, 2, NCOLS/C_PARTITIONS+4, MPI_INT, &LeftRightDataType);

Moore Neighborhood on 2D Grids

Defining custom data types for generic radius-based communication:

• Vertical Edges (Columns): MPI_Type_vector(N/numPartition, radius, N + 2*radius, MPI_INT, &colDT);
• Horizontal Edges (Rows): MPI_Type_vector(radius, N + 2*radius, N + 2*radius, MPI_INT, &rowDT);