#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <assert.h>
#include <signal.h>
#include <string.h>
#include <sys/time.h>
#include <mpi.h>

#define COMBUFSIZE (4 * 1024 * 1024)

int mpi_self;
int mpi_npes;

char* rbuffer;
char* sbuffer;

double
perf_alltoall(int size, int iter)
{
    MPI_Status	status;
    int err;
    struct timeval begin, end;
    int sec, usec;
    double total;
    int i;
    int stride;
    char* rbuf;
    char* sbuf;

    if (size * mpi_npes > COMBUFSIZE) {
	rbuf = (char*) valloc(size * mpi_npes);
	bzero(rbuf, size * mpi_npes);
    } else {
	rbuf = rbuffer;
    }
    if (size * mpi_npes > COMBUFSIZE / 2) {
	sbuf = (char*) valloc(size * mpi_npes);
	bzero(sbuf, size * mpi_npes);
	stride = 0;
    } else {
	sbuf = sbuffer;
	stride = (size * mpi_npes + 32) & ~31;
    }

    MPI_Barrier(MPI_COMM_WORLD);
    gettimeofday(&begin, 0);
    for (i = 0; i < iter; i++) {
	/*************************/
	/* do something when size*/
	/*************************/
	MPI_Alltoall(sbuf,
		     size,
		     MPI_BYTE,
		     rbuf,
		     size,
		     MPI_BYTE,
		     MPI_COMM_WORLD);
	/*************************/
	sbuf += stride;
	if (stride > 0 && sbuf > sbuffer + COMBUFSIZE - size * mpi_npes) {
	    sbuf = sbuffer;
	}
	MPI_Barrier(MPI_COMM_WORLD);
    }
    gettimeofday(&end, 0);
    sec = end.tv_sec - begin.tv_sec;
    usec = end.tv_usec - begin.tv_usec;
    if (usec < 0) {
	usec += 1000000;
	sec--;
    }
    total = sec * 1000000 + usec;

    return total / (double)iter;
}

int
main (int argc, char *argv[])
{
    int size, iter;
    int err;
    double time;

    MPI_Init (&argc, &argv);
    MPI_Comm_rank (MPI_COMM_WORLD, &mpi_self);
    MPI_Comm_size (MPI_COMM_WORLD, &mpi_npes);

    rbuffer = (char*)valloc(COMBUFSIZE);
    sbuffer = (char*)valloc(COMBUFSIZE);
    bzero(rbuffer, COMBUFSIZE);
    bzero(sbuffer, COMBUFSIZE);

    MPI_Barrier(MPI_COMM_WORLD);

    if (argc >= 3) {
	size = atoi(argv[1]);
	iter = atoi(argv[2]);
    } else {
	size = 16384;
	iter = 100;
    }

    if (mpi_self == 0) {
	time = perf_alltoall(size, iter);
    } else {
	perf_alltoall(size, iter);
    }

    if (mpi_self == 0) {
	printf("%d\t%d\t%f\n",
	       mpi_npes,
	       size,
	       time);
    }

    MPI_Finalize();
    return 0;
}