--- /dev/null
+#define DEFAULT_POWER 100
+
+#define MPI_ANY_SOURCE -1
+
+// errorcodes
+#define MPI_SUCCESS 0
+#define MPI_ERR_COMM 1
+#define MPI_ERR_ARG 2
+#define MPI_ERR_TYPE 3
+#define MPI_ERR_REQUEST 4
+#define MPI_ERR_INTERN 5
+#define MPI_ERR_COUNT 6
+#define MPI_ERR_RANK 7
+#define MPI_ERR_TAG 8
+
+#include <stdlib.h>
+#include <msg/msg.h>
+
+typedef enum { MPI_PORT = 0, SEND_SYNC_PORT, RECV_SYNC_PORT, MAX_CHANNEL } channel_t;
+
+// MPI_Comm
+struct smpi_mpi_communicator_t {
+ int id;
+ int size;
+ int barrier;
+ m_host_t *hosts;
+ m_process_t *processes;
+};
+typedef struct smpi_mpi_communicator_t smpi_mpi_communicator_t;
+typedef smpi_mpi_communicator_t *MPI_Comm;
+extern smpi_mpi_communicator_t smpi_mpi_comm_world;
+#define MPI_COMM_WORLD (&smpi_mpi_comm_world)
+
+// MPI_Status
+struct smpi_mpi_status_t {
+ int MPI_SOURCE;
+};
+typedef struct smpi_mpi_status_t smpi_mpi_status_t;
+typedef smpi_mpi_status_t MPI_Status;
+extern smpi_mpi_status_t smpi_mpi_status_ignore;
+#define MPI_STATUS_IGNORE (&smpi_mpi_status_ignore)
+
+// MPI_Datatype
+struct smpi_mpi_datatype_t {
+// int type;
+ size_t size;
+};
+typedef struct smpi_mpi_datatype_t smpi_mpi_datatype_t;
+typedef smpi_mpi_datatype_t *MPI_Datatype;
+// FIXME: add missing datatypes
+extern smpi_mpi_datatype_t smpi_mpi_byte;
+#define MPI_BYTE (&smpi_mpi_byte)
+extern smpi_mpi_datatype_t smpi_mpi_int;
+#define MPI_INT (&smpi_mpi_int)
+extern smpi_mpi_datatype_t smpi_mpi_double;
+#define MPI_DOUBLE (&smpi_mpi_double)
+
+struct smpi_waitlist_node_t {
+ m_process_t process;
+ struct smpi_waitlist_node_t *next;
+};
+typedef struct smpi_waitlist_node_t smpi_waitlist_node_t;
+
+// FIXME: maybe it isn't appropriate to have the next pointer inside
+// MPI_Request
+struct smpi_mpi_request_t {
+ void *buf;
+ int count;
+ smpi_mpi_datatype_t *datatype;
+ int src;
+ int dst;
+ int tag;
+ smpi_mpi_communicator_t *comm;
+ short int completed;
+ smpi_waitlist_node_t *waitlist;
+ struct smpi_mpi_request_t *next;
+ int fwdthrough;
+};
+typedef struct smpi_mpi_request_t smpi_mpi_request_t;
+typedef smpi_mpi_request_t *MPI_Request;
+
+// MPI_Op
+struct smpi_mpi_op_t {
+ void (*func)(void *x, void *y, void *z);
+};
+typedef struct smpi_mpi_op_t smpi_mpi_op_t;
+typedef smpi_mpi_op_t *MPI_Op;
+extern smpi_mpi_op_t smpi_mpi_land;
+#define MPI_LAND (&smpi_mpi_land)
+extern smpi_mpi_op_t smpi_mpi_sum;
+#define MPI_SUM (&smpi_mpi_sum)
+
+// smpi_received_t
+struct smpi_received_t {
+ int commid;
+ int src;
+ int dst;
+ int tag;
+ int fwdthrough;
+ void *data;
+ struct smpi_received_t *next;
+};
+typedef struct smpi_received_t smpi_received_t;
+
+// sender/receiver (called by main routine)
+int smpi_sender(int argc, char *argv[]);
+int smpi_receiver(int argc, char *argv[]);
+
+// smpi functions
+int smpi_comm_rank(smpi_mpi_communicator_t *comm, m_host_t host);
+void smpi_isend(smpi_mpi_request_t*);
+void smpi_irecv(smpi_mpi_request_t*);
+void smpi_barrier(smpi_mpi_communicator_t *comm);
+void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status);
+void smpi_wait_all(int count, smpi_mpi_request_t **requests, smpi_mpi_status_t *statuses);
+void smpi_wait_all_nostatus(int count, smpi_mpi_request_t **requests);
+void smpi_bench_begin();
+void smpi_bench_end();
+int smpi_create_request(void *buf, int count, smpi_mpi_datatype_t *datatype, int src, int dst, int tag, smpi_mpi_communicator_t *comm, smpi_mpi_request_t **request);
+unsigned int smpi_sleep(unsigned int);
+void smpi_exit(int);
--- /dev/null
+#include <stdio.h>
+#include <stdlib.h>
+#include <mpi.h>
+
+int main(int argc, char *argv[]) {
+ int rank, size;
+ int i;
+ int *sendbuf, *recvbuf;
+ MPI_Init(&argc, &argv);
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+ sendbuf = malloc(sizeof(int) * size);
+ recvbuf = malloc(sizeof(int) * size);
+ for (i = 0; i < size; i++) {
+ sendbuf[i] = 0;
+ recvbuf[i] = 0;
+ }
+ sendbuf[rank] = rank + 1;
+ MPI_Allreduce(sendbuf, recvbuf, size, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+ printf("node %d has: ", rank);
+ for (i = 0; i < size; i++) {
+ printf("%d ", recvbuf[i]);
+ }
+ printf("\n");
+ free(sendbuf);
+ free(recvbuf);
+ MPI_Finalize();
+ return 0;
+}
--- /dev/null
+#include "mpi.h"
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <errno.h>
+
+#ifndef EXIT_SUCCESS
+#define EXIT_SUCCESS 0
+#define EXIT_FAILURE 1
+#endif
+
+// sandler says, compile with mpicc -v alltoalldemo.c
+// run with mpirun -np 3 a.out -m 5
+
+int main( int argc, char *argv[] )
+{
+ int rank, size;
+ int chunk = 128;
+ int i;
+ int j; // added by sandler
+ int *sb;
+ int *rb;
+ int status, gstatus;
+
+ MPI_Init(&argc,&argv);
+ MPI_Comm_rank(MPI_COMM_WORLD,&rank);
+ MPI_Comm_size(MPI_COMM_WORLD,&size);
+ if (rank==0) {
+ printf("size: %d\n", size);
+ }
+ for ( i=1 ; i < argc ; ++i ) {
+ if ( argv[i][0] != '-' ) {
+ // added by sandler
+ fprintf(stderr, "Unrecognized option %s\n", argv[i]);fflush(stderr);
+ continue;
+ }
+ switch(argv[i][1]) {
+ case 'm':
+ chunk = atoi(argv[++i]);
+ if (rank==0) {
+ printf("chunk: %d\n", chunk);
+ }
+ break;
+ default:
+ fprintf(stderr, "Unrecognized argument %s\n", argv[i]);fflush(stderr);
+ MPI_Abort(MPI_COMM_WORLD,EXIT_FAILURE);
+ }
+ }
+ sb = (int *)malloc(size*chunk*sizeof(int));
+ if ( !sb ) {
+ perror( "can't allocate send buffer" );fflush(stderr);
+ MPI_Abort(MPI_COMM_WORLD,EXIT_FAILURE);
+ }
+ rb = (int *)malloc(size*chunk*sizeof(int));
+ if ( !rb ) {
+ perror( "can't allocate recv buffer");fflush(stderr);
+ free(sb);
+ MPI_Abort(MPI_COMM_WORLD, EXIT_FAILURE);
+ }
+
+
+ /* original deino.net:
+ for ( i=0 ; i < size*chunk ; ++i ) {
+ sb[i] = sb[i] = rank + 1;
+ rb[i] = 0;
+ }
+ */
+ // written by sandler
+
+ if (rank==0) printf("note in the following:\n"
+ "if you were to compare the sending buffer and the receiving buffer on the SAME processor, \n"
+ "you might think that the values were getting wiped out. However, each row IS going somewhere. \n"
+ "The 0th row of processor 0 goes to the 0th row of processor 0\n"
+ "The 1st row of processor 0 goes to the 0th row of processor 1. (Go look at rb for processor 1!)\n"
+ "\n"
+ "Too bad the values don't come out in a deterministic order. That's life!\n"
+ "\n"
+ "Now look at the receiving buffer for processor 0.\n"
+ "The 0th row is from processor 0 (itself).\n"
+ "The 1st row on processor 0 is from the 0th row on processor 1. (Go look at the sb of processor 1!)\n"
+ "\n"
+ "Apparently this is the intended behavior.\n"
+ "\n"
+ "Note that each row is always moved as one chunk, unchangeable.\n"
+ "\n"
+ "TODO: draw a diagram\n"
+ );
+
+ for (i=0; i<size; i++) {
+ for (j=0; j<chunk; j++) {
+ int offset = i*chunk + j; // note the multiplier is chunk, not size
+
+ sb[offset] = rank*100 + i*10 + j;
+ rb[offset] = -1;
+ }
+ }
+
+
+ // this clearly shows what is NOT indended to be done, in that the rb on a processor is the same as the sb on the processor
+ // in this intialization: on processor 0, only the 0th row gets normal values.
+ // on processor 1, only the 1st row gets normal values.
+ // when you look the rb, it looks like nothing happened. this is because, say, for processor 1, the 1st row got sent to itself.
+ /*
+ for (i=0; i<size; i++) {
+ for (j=0; j<chunk; j++) {
+ int offset = i*chunk + j; // note the multiplier is chunk, not size
+
+ if (i==rank)
+ sb[offset] = rank*100 + i*10 + j;
+ else
+ sb[offset] = 999;
+
+ rb[i*chunk + j] = 999;
+ }
+ }
+ */
+
+ // this does printgrid("sb", rank, size, chunk, sb);
+ //added by sandler
+ printf("[processor %d] To send:\n", rank);
+ for (i=0; i<size; i++) {
+ for (j=0; j<chunk; j++) {
+ // note the multiplier is chunk, not size
+ printf("%03d ", sb[i*chunk+j]);
+ }
+ printf("\n");
+ }
+ printf("\n");
+
+/*
+ // for another variation, could send out a bunch of characters, like
+ p r o c e s s o r 0 r o w 0
+ p r o c e s s o r 0 r o w 1
+ p r o c e s s o r 0 r o w 2
+
+ then you'd get back
+
+ p r o c e s s o r 0 r o w 0
+ p r o c e s s o r 1 r o w 0
+ p r o c e s s o r 2 r o w 0
+*/
+
+
+ status = MPI_Alltoall(sb, chunk, MPI_INT, rb, chunk, MPI_INT, MPI_COMM_WORLD);
+
+ // this does printgrid("rb", rank, size, chunk, rb);
+ // added by sandler
+ printf("[processor %d] Received:\n", rank);
+ for (i=0; i<size; i++) {
+ for (j=0; j<chunk; j++) {
+ printf("%03d ", rb[i*chunk+j]);
+ }
+ printf("\n");
+ }
+ printf("\n");
+
+
+ MPI_Allreduce( &status, &gstatus, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD );
+ if (rank == 0) {
+ if (gstatus != 0) {
+ printf("all_to_all returned %d\n",gstatus);fflush(stdout);
+ }
+
+
+ }
+
+
+
+ free(sb);
+ free(rb);
+ MPI_Finalize();
+ return(EXIT_SUCCESS);
+}
--- /dev/null
+#include <stdio.h>
+#include <mpi.h>
+
+int main (int argc, char **argv) {
+ int size, rank;
+ int value = 3;
+ MPI_Init(&argc, &argv);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ if (0 == rank) {
+ value = 17;
+ }
+ printf("node %d has value %d\n", rank, value);
+ MPI_Bcast(&value, 1, MPI_INT, 0, MPI_COMM_WORLD);
+ printf("node %d has value %d\n", rank, value);
+ MPI_Finalize();
+ return 0;
+}
--- /dev/null
+#include <stdio.h>
+#include <stdlib.h>
+#include <mpi.h>
+
+#define GETTIMEOFDAY_ERROR 1
+
+#define N_START 1
+#define N_STOP 1024*1024
+#define N_NEXT (N*2)
+#define ITER 100
+#define ONE_MILLION 1000000.0
+#define RAND_SEED 842270
+
+int main(int argc, char* argv[]) {
+
+ int size, rank;
+ int N, I;
+ struct timeval *start_time, *stop_time;
+ double seconds;
+ int i, j, k;
+ char *buffer;
+ int check;
+
+ srandom(RAND_SEED);
+
+ MPI_Init(&argc, &argv);
+
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+
+ if (0 == rank) {
+ start_time = (struct timeval *)malloc(sizeof(struct timeval));
+ stop_time = (struct timeval *)malloc(sizeof(struct timeval));
+ }
+
+ for (N = N_START ; N <= N_STOP ; N = N_NEXT) {
+
+ buffer = malloc(sizeof(char) * N);
+
+ if (0 == rank) {
+ for (j = 0; j < N; j++) {
+ buffer[j] = (char)(random() % 256);
+ }
+ if (-1 == gettimeofday(start_time, NULL)) {
+ printf("couldn't set start_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(EXIT_FAILURE);
+ }
+ }
+
+ for (i = 0; i < ITER; i++) {
+ MPI_Bcast(buffer, N, MPI_BYTE, 0, MPI_COMM_WORLD);
+ if (0 == rank) {
+ for (j = 1; j < size; j++) {
+ MPI_Recv(&check, 1, MPI_INT, MPI_ANY_SOURCE, 1, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+ }
+ } else {
+ MPI_Send(&rank, 1, MPI_INT, 0, 1, MPI_COMM_WORLD);
+ }
+ }
+
+ if (0 == rank) {
+ if (-1 == gettimeofday(stop_time, NULL)) {
+ printf("couldn't set start_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(EXIT_FAILURE);
+ }
+ seconds = (double)(stop_time->tv_sec - start_time->tv_sec) + (double)(stop_time->tv_usec - start_time->tv_usec) / ONE_MILLION;
+ }
+
+ free(buffer);
+
+ if (0 == rank) {
+ printf("N: %10d, iter: %d, time: %10f s, avg rate: %12f Mbps\n", N, ITER, seconds, ((double)N * ITER * 8) / (1024.0 * 1024.0 * seconds));
+ }
+
+ }
+
+ if (0 == rank) {
+ free(start_time);
+ free(stop_time);
+ }
+
+ MPI_Finalize();
+
+ return 0;
+}
--- /dev/null
+/* A first simple SPMD example program using MPI */
+
+/* The program consists of on receiver process and N-1 sender */
+/* processes. The sender processes send a message consisting */
+/* of their process identifier (id) and the total number of */
+/* processes (ntasks) to the receiver. The receiver process */
+/* prints out the values it receives in the messeges from the */
+/* senders. */
+
+/* Compile the program with 'mpicc first.c -o first' */
+/* To run the program, using four of the computers specified in */
+/* your hostfile, do 'mpirun -machinefile hostfile -np 4 first */
+
+#include <stdio.h>
+#include <mpi.h>
+main(int argc, char *argv[]) {
+ const int tag = 42; /* Message tag */
+ int id, ntasks, source_id, dest_id, err, i;
+ MPI_Status status;
+ int msg[2]; /* Message array */
+
+ err = MPI_Init(&argc, &argv); /* Initialize MPI */
+ if (err != MPI_SUCCESS) {
+ printf("MPI initialization failed!\n");
+ exit(1);
+ }
+ err = MPI_Comm_size(MPI_COMM_WORLD, &ntasks); /* Get nr of tasks */
+ err = MPI_Comm_rank(MPI_COMM_WORLD, &id); /* Get id of this process */
+ if (ntasks < 2) {
+ printf("You have to use at least 2 processors to run this program\n");
+ MPI_Finalize(); /* Quit if there is only one processor */
+ exit(0);
+ }
+
+ if (id == 0) { /* Process 0 (the receiver) does this */
+ for (i=1; i<ntasks; i++) {
+ err = MPI_Recv(msg, 2, MPI_INT, MPI_ANY_SOURCE, tag, MPI_COMM_WORLD, \
+ &status); /* Receive a message */
+ source_id = status.MPI_SOURCE; /* Get id of sender */
+ printf("Received message %d %d from process %d\n", msg[0], msg[1], \
+ source_id);
+ }
+ }
+ else { /* Processes 1 to N-1 (the senders) do this */
+ msg[0] = id; /* Put own identifier in the message */
+ msg[1] = ntasks; /* and total number of processes */
+ dest_id = 0; /* Destination address */
+ sleep(3);
+ err = MPI_Send(msg, 2, MPI_INT, dest_id, tag, MPI_COMM_WORLD);
+ }
+
+ err = MPI_Finalize(); /* Terminate MPI */
+ if (id==0) printf("Ready\n");
+ return 0;
+}
+
--- /dev/null
+/*
+ * Mark Stillwell
+ * ICS691: High Performance Computing
+ * Fall 2006
+ * Homework 3, Exercise 2, Step 1
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <libgen.h>
+#include <mpi.h>
+
+#define ITERATIONS 10
+#define STEPS 1
+#define STEP_SIZE 0
+
+#define USAGE_ERROR 1
+#define MALLOC_ERROR 2
+#define GETTIMEOFDAY_ERROR 3
+
+void * checked_malloc(int rank, const char * varname, size_t size) {
+ void * ptr;
+ ptr = malloc(size);
+ if (NULL == ptr) {
+ printf("node %d could not malloc memory for %s.\n", rank, varname);
+ MPI_Abort(MPI_COMM_WORLD, MALLOC_ERROR);
+ exit(MALLOC_ERROR);
+ }
+ return ptr;
+}
+
+int main(int argc, char* argv[]) {
+
+ // timing/system variables
+ int iteration, iterations = ITERATIONS;
+ int step, steps = STEPS, step_size = STEP_SIZE;
+ long usecs, total_usecs;
+ struct timeval *start_time, *stop_time;
+ char *program;
+
+ // mpi/communications variables
+ int rank;
+ int row, col;
+ MPI_Comm row_comm, col_comm;
+
+ // algorithm variables
+ int N_start, N, P;
+ int *A, *A_t, *B, *C, *D, *a, *b, *abuf, *bbuf;
+ int n, i, j, k, I, J;
+
+ MPI_Init(&argc, &argv);
+
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+ if (0 == rank) {
+ int size;
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+
+ program = basename(argv[0]);
+
+ // root node parses cmdline args
+ /*
+ if (3 > argc || !isdigit(*argv[1]) || !isdigit(*argv[2])) {
+ printf("usage:\n%s <N> <P> [<iterations>]\n", program);
+ MPI_Abort(MPI_COMM_WORLD, USAGE_ERROR);
+ exit(USAGE_ERROR);
+ }
+ */
+
+ //N_start = atoi(argv[1]);
+ //P = atoi(argv[2]);
+ N_start = 100;
+ P = 2;
+
+ /*
+ if (4 <= argc && isdigit(*argv[3])) {
+ iterations = atoi(argv[3]);
+ }
+
+ if (5 <= argc && isdigit(*argv[4])) {
+ steps = atoi(argv[4]);
+ }
+
+ if (6 <= argc && isdigit(*argv[5])) {
+ step_size = atoi(argv[5]);
+ }
+ */
+
+ if (P*P != size) {
+ printf("P^2 must equal size.\n");
+ MPI_Abort(MPI_COMM_WORLD, USAGE_ERROR);
+ exit(USAGE_ERROR);
+ }
+
+ start_time = (struct timeval *)checked_malloc(rank, "start_time", sizeof(struct timeval));
+ stop_time = (struct timeval *)checked_malloc(rank, "stop_time", sizeof(struct timeval));
+
+ }
+
+ // send command line parameters except N, since it can vary
+ MPI_Bcast(&P, 1, MPI_INT, 0, MPI_COMM_WORLD);
+ MPI_Bcast(&iterations, 1, MPI_INT, 0, MPI_COMM_WORLD);
+ MPI_Bcast(&steps, 1, MPI_INT, 0, MPI_COMM_WORLD);
+ MPI_Bcast(&step_size, 1, MPI_INT, 0, MPI_COMM_WORLD);
+
+ row = rank / P;
+ col = rank % P;
+
+ // create row/column communicators
+ MPI_Comm_split(MPI_COMM_WORLD, row, col, &row_comm);
+ MPI_Comm_split(MPI_COMM_WORLD, col, row, &col_comm);
+
+ for (step = 0; step < steps; step++) {
+
+ total_usecs = 0;
+
+ if (0 == rank) {
+ N = N_start + step * step_size;
+ if ((N/P)*P != N) {
+ printf("P must divide N and %d does not divide %d.\n", P, N);
+ N = -1;
+ }
+ }
+
+ MPI_Bcast(&N, 1, MPI_INT, 0, MPI_COMM_WORLD);
+
+ // if root passes N = -1, skip this round
+ if (-1 == N) continue;
+
+ n = N / P;
+
+ // initialize matrix components
+ A = (int *)checked_malloc(rank, "A", n*n*sizeof(int));
+ A_t = (int *)checked_malloc(rank, "A_t", n*n*sizeof(int));
+ B = (int *)checked_malloc(rank, "B", n*n*sizeof(int));
+ C = (int *)checked_malloc(rank, "C", n*n*sizeof(int));
+ D = (int *)checked_malloc(rank, "D", n*n*sizeof(int));
+
+ for (i = 0; i < n; i++) {
+ for (j = 0; j < n; j++) {
+
+ I = n*row+i;
+ J = n*col+j;
+
+ A[n*i+j] = I+J;
+ B[n*i+j] = I;
+
+ // d is the check matrix
+ D[n*i+j] = 0;
+ for (k = 0; k < N; k++) {
+ // A[I,k] = I+k
+ // B[k,J] = k
+ D[n*i+j] += (I+k) * k;
+ }
+
+ }
+ }
+
+ // buffers
+ abuf = (int *)checked_malloc(rank, "abuf", n*sizeof(int));
+ bbuf = (int *)checked_malloc(rank, "bbuf", n*sizeof(int));
+
+ for (iteration = 0; iteration < iterations; iteration++) {
+
+ for (i = 0; i < n*n; i++) {
+ C[i] = 0;
+ }
+
+ // node zero sets start time
+ if (0 == rank && -1 == gettimeofday(start_time, NULL)) {
+ printf("couldn't set start_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(GETTIMEOFDAY_ERROR);
+ }
+
+ // populate transpose of A
+ for (i = 0; i < n; i++) {
+ for (j = 0; j < n; j++) {
+ A_t[n*i+j] = A[n*j+i];
+ }
+ }
+
+ // perform calculations
+ for (k = 0; k < N; k++) {
+
+ if (k/n == col) {
+ a = A_t + n*(k%n);
+ } else {
+ a = abuf;
+ }
+
+ if (k/n == row) {
+ b = B + n*(k%n);
+ } else {
+ b = bbuf;
+ }
+
+ MPI_Bcast(a, n, MPI_INT, k/n, row_comm);
+ MPI_Bcast(b, n, MPI_INT, k/n, col_comm);
+
+ for (i = 0; i < n; i++) {
+ for (j = 0; j < n; j++) {
+ C[n*i+j] += a[i] * b[j];
+ }
+ }
+
+ } // for k
+
+ // node zero sets stop time
+ if (0 == rank && -1 == gettimeofday(stop_time, NULL)) {
+ printf("couldn't set stop_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(GETTIMEOFDAY_ERROR);
+ }
+
+ // check calculation
+ for (i = 0; i < n*n && C[i] == D[i]; i++);
+ j = (n*n == i);
+ MPI_Reduce(&j, &k, 1, MPI_INT, MPI_LAND, 0, MPI_COMM_WORLD);
+
+ // node zero prints stats
+ if (0 == rank) {
+ usecs = (stop_time->tv_sec*1000000+stop_time->tv_usec) - (start_time->tv_sec*1000000+start_time->tv_usec);
+ printf("prog: %s, N: %d, P: %d, procs: %d, time: %d us, check: %d\n", program, N, P, P*P, usecs, k);
+ total_usecs += usecs;
+ }
+
+ }
+
+ // node 0 prints final stats
+ if (0 == rank) {
+ printf("prog: %s, N: %d, P: %d, procs: %d, iterations: %d, avg. time: %d us\n",
+ program, N, P, P*P, iterations, total_usecs / iterations);
+ }
+
+ // free data structures
+ free(A);
+ free(A_t);
+ free(B);
+ free(C);
+ free(D);
+ free(abuf);
+ free(bbuf);
+
+ }
+
+ if (0 == rank) {
+ free(start_time);
+ free(stop_time);
+ }
+
+ MPI_Finalize();
+
+ return 0;
+}
--- /dev/null
+#include <stdio.h>
+#include <stdlib.h>
+#include <mpi.h>
+
+#define ITERATIONS 10
+
+#define USAGE_ERROR 1
+#define SANITY_ERROR 2
+#define GETTIMEOFDAY_ERROR 3
+
+int main(int argc, char* argv[]) {
+
+ int size, rank;
+ int N, n, i, j, k, current_iteration, successful_iterations = 0;
+ double *matrix, *vector, *vcalc, *vcheck;
+ MPI_Status status;
+ struct timeval *start_time, *stop_time;
+ long parallel_usecs, parallel_usecs_total = 0, sequential_usecs, sequential_usecs_total = 0;
+
+ MPI_Init(&argc, &argv);
+
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+
+ if (0 == rank) {
+
+ // root node parses cmdline args
+ if (2 > argc || !isdigit(*argv[1])) {
+ printf("usage:\n%s <size>\n", argv[0]);
+ MPI_Abort(MPI_COMM_WORLD, USAGE_ERROR);
+ exit(USAGE_ERROR);
+ }
+
+ N = atoi(argv[1]);
+
+ start_time = (struct timeval *)malloc(sizeof(struct timeval));
+ stop_time = (struct timeval *)malloc(sizeof(struct timeval));
+
+ }
+
+ for(current_iteration = 0; current_iteration < ITERATIONS; current_iteration++) {
+
+ if (0 == rank) {
+
+ matrix = (double *)malloc(N*N*sizeof(double));
+ vector = (double *)malloc(N*sizeof(double));
+
+ for(i = 0; i < N*N; i++) {
+ matrix[i] = (double)rand()/((double)RAND_MAX + 1);
+ }
+
+ for(i = 0; i < N; i++) {
+ vector[i] = (double)rand()/((double)RAND_MAX + 1);
+ }
+
+ // for the sake of argument, the parallel algorithm begins
+ // when the root node begins to transmit the matrix to the
+ // workers.
+ if (-1 == gettimeofday(start_time, NULL)) {
+ printf("couldn't set start_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(GETTIMEOFDAY_ERROR);
+ }
+
+ for(i = 1; i < size; i++) {
+ MPI_Send(&N, 1, MPI_INT, i, 0, MPI_COMM_WORLD);
+ }
+
+ } else {
+ MPI_Recv(&N, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
+ }
+
+ // this algorithm uses at most N processors...
+ if (rank < N) {
+
+ if (size > N) size = N;
+ n = N / size + ((rank < (N % size)) ? 1 : 0);
+
+ if (0 == rank) {
+
+ for(i = 1, j = n; i < size && j < N; i++, j+=k) {
+ k = N / size + ((i < (N % size)) ? 1 : 0);
+ MPI_Send(matrix+N*j, N*k, MPI_DOUBLE, i, 0, MPI_COMM_WORLD);
+ MPI_Send(vector, N, MPI_DOUBLE, i, 0, MPI_COMM_WORLD);
+ }
+
+ // sanity check
+ #ifdef DEBUG
+ if(i != size || j != N) {
+ printf("index calc error: i = %d, size = %d, j = %d, N = %d\n", i, size, j, N);
+ MPI_Abort(MPI_COMM_WORLD, SANITY_ERROR);
+ exit(SANITY_ERROR);
+ }
+ #endif
+
+ vcalc = (double *)malloc(N*sizeof(double));
+
+ } else {
+
+ matrix = (double *)malloc(N*n*sizeof(double));
+ vector = (double *)malloc(N*sizeof(double));
+ vcalc = (double *)malloc(n*sizeof(double));
+
+ MPI_Recv(matrix, N*n, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD, &status);
+ MPI_Recv(vector, N, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD, &status);
+
+ }
+
+ for(i = 0; i < n; i++) {
+ vcalc[i] = 0.0;
+ for(j = 0; j < N; j++) {
+ vcalc[i] += matrix[N*i+j] * vector[j];
+ }
+ }
+
+ if (0 != rank) {
+ MPI_Send(vcalc, n, MPI_DOUBLE, 0, 0, MPI_COMM_WORLD);
+ } else {
+
+ for(i = 1, j = n; i < size && j < N; i++, j+=k) {
+ k = N / size + ((i < (N % size)) ? 1 : 0);
+ MPI_Recv(vcalc+j, k, MPI_DOUBLE, i, 0, MPI_COMM_WORLD, &status);
+ }
+
+ // sanity check
+ #ifdef DEBUG
+ if(i != size || j != N) {
+ printf("index calc error 2: i = %d, size = %d, j = %d, N = %d\n", i, size, j, N);
+ MPI_Abort(MPI_COMM_WORLD, SANITY_ERROR);
+ exit(SANITY_ERROR);
+ }
+ #endif
+
+ if (-1 == gettimeofday(stop_time, NULL)) {
+ printf("couldn't set stop_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(GETTIMEOFDAY_ERROR);
+ }
+
+ parallel_usecs = (stop_time->tv_sec*1000000+stop_time->tv_usec) - (start_time->tv_sec*1000000+start_time->tv_usec);
+
+ if (-1 == gettimeofday(start_time, NULL)) {
+ printf("couldn't set start_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(GETTIMEOFDAY_ERROR);
+ }
+
+ // calculate serially
+ vcheck = (double *)malloc(N*sizeof(double));
+ for(i = 0; i < N; i++) {
+ vcheck[i] = 0.0;
+ for(j = 0; j < N; j++) {
+ vcheck[i] += matrix[N*i+j] * vector[j];
+ }
+ }
+
+ if (-1 == gettimeofday(stop_time, NULL)) {
+ printf("couldn't set stop_time on node 0!\n");
+ MPI_Abort(MPI_COMM_WORLD, GETTIMEOFDAY_ERROR);
+ exit(GETTIMEOFDAY_ERROR);
+ }
+
+ sequential_usecs = (stop_time->tv_sec*1000000+stop_time->tv_usec) - (start_time->tv_sec*1000000+start_time->tv_usec);
+
+ // verify correctness
+ for(i = 0; i < N && vcalc[i] == vcheck[i]; i++);
+
+ printf("prog: blocking, i: %d ", current_iteration);
+
+ if (i == N) {
+ printf("ptime: %d us, stime: %d us, speedup: %.3f, nodes: %d, efficiency: %.3f\n",
+ parallel_usecs,
+ sequential_usecs,
+ (double)sequential_usecs / (double)parallel_usecs,
+ size,
+ (double)sequential_usecs / ((double)parallel_usecs * (double)size));
+
+ parallel_usecs_total += parallel_usecs;
+ sequential_usecs_total += sequential_usecs;
+ successful_iterations++;
+ } else {
+ printf("parallel calc != serial calc, ");
+ }
+
+ free(vcheck);
+
+ }
+
+ free(matrix);
+ free(vector);
+ free(vcalc);
+ }
+
+ }
+
+ if(0 == rank) {
+ printf("prog: blocking, ");
+ if(0 < successful_iterations) {
+ printf("iterations: %d, avg. ptime: %d us, avg. stime: %d us, avg. speedup: %.3f, nodes: %d, avg. efficiency: %.3f\n",
+ successful_iterations,
+ parallel_usecs_total / successful_iterations,
+ sequential_usecs_total / successful_iterations,
+ (double)sequential_usecs_total / (double)parallel_usecs_total,
+ size,
+ (double)sequential_usecs_total / ((double)parallel_usecs_total * (double)size));
+ } else {
+ printf("no successful iterations!\n");
+ }
+
+ free(start_time);
+ free(stop_time);
+
+ }
+
+ MPI_Finalize();
+
+ return 0;
+}
--- /dev/null
+#include <stdio.h>
+#include <stdlib.h>
+#include <mpi.h>
+
+int main(int argc, char *argv[]) {
+ int rank, size;
+ int i;
+ int *sendbuf, *recvbuf;
+ MPI_Init(&argc, &argv);
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+ sendbuf = malloc(sizeof(int) * size);
+ recvbuf = malloc(sizeof(int) * size);
+ for (i = 0; i < size; i++) {
+ sendbuf[i] = 0;
+ recvbuf[i] = 0;
+ }
+ sendbuf[rank] = rank + 1;
+ MPI_Reduce(sendbuf, recvbuf, size, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
+ if (0 == rank) {
+ printf("nodes: ", rank);
+ for (i = 0; i < size; i++) {
+ printf("%d ", recvbuf[i]);
+ }
+ printf("\n");
+ }
+ free(sendbuf);
+ free(recvbuf);
+ MPI_Finalize();
+ return 0;
+}
--- /dev/null
+/*
+ * Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana
+ * University Research and Technology
+ * Corporation. All rights reserved.
+ * Copyright (c) 2006 Cisco Systems, Inc. All rights reserved.
+ *
+ * Simple ring test program
+ */
+
+#include <stdio.h>
+#include "mpi.h"
+
+int main(int argc, char *argv[])
+{
+ int rank, size, next, prev, message, tag = 201;
+
+ /* Start up MPI */
+
+ MPI_Init(&argc, &argv);
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+
+ /* Calculate the rank of the next process in the ring. Use the
+ modulus operator so that the last process "wraps around" to
+ rank zero. */
+
+ next = (rank + 1) % size;
+ prev = (rank + size - 1) % size;
+
+ /* If we are the "master" process (i.e., MPI_COMM_WORLD rank 0),
+ put the number of times to go around the ring in the
+ message. */
+
+ if (0 == rank) {
+ message = 10;
+
+ printf("Process 0 sending %d to %d, tag %d (%d processes in ring)\n",
+ message, next, tag, size);
+ MPI_Send(&message, 1, MPI_INT, next, tag, MPI_COMM_WORLD);
+ printf("Process 0 sent to %d\n", next);
+ }
+
+ /* Pass the message around the ring. The exit mechanism works as
+ follows: the message (a positive integer) is passed around the
+ ring. Each time it passes rank 0, it is decremented. When
+ each processes receives a message containing a 0 value, it
+ passes the message on to the next process and then quits. By
+ passing the 0 message first, every process gets the 0 message
+ and can quit normally. */
+
+ sleep(3);
+
+ while (1) {
+ MPI_Recv(&message, 1, MPI_INT, prev, tag, MPI_COMM_WORLD,
+ MPI_STATUS_IGNORE);
+
+ if (0 == rank) {
+ --message;
+ printf("Process 0 decremented value: %d\n", message);
+ }
+
+ MPI_Send(&message, 1, MPI_INT, next, tag, MPI_COMM_WORLD);
+ if (0 == message) {
+ printf("Process %d exiting\n", rank);
+ break;
+ }
+ }
+
+ /* The last process does one extra send to process 0, which needs
+ to be received before the program can exit */
+
+ if (0 == rank) {
+ MPI_Recv(&message, 1, MPI_INT, prev, tag, MPI_COMM_WORLD,
+ MPI_STATUS_IGNORE);
+ }
+
+ /* All done */
+
+ MPI_Finalize();
+ return 0;
+}
--- /dev/null
+#!/bin/sh
+#FIXME: .. paths...
+SIMGRID_INCLUDE="${SIMGRID_HOME}/include"
+SIMGRID_LIB="${SIMGRID_HOME}/lib"
+CC="gcc"
+
+SMPI_INCLUDE="${SMPI_HOME}/include"
+SMPI_LIB="${SMPI_HOME}/lib"
+SEED="221238"
+
+TMPDIR="$(mktemp -d tmpXXXXXXX)"
+
+function modsource {
+ SOURCE="$1"
+ SOURCEFILE="$(basename ${SOURCE})"
+ SOURCEDIR="${SOURCE%${SOURCEFILE}}"
+ if [ -n "${SOURCEDIR}" ]; then
+ mkdir -p ${TMPDIR}${SOURCEDIR}
+ fi
+ TMPSOURCE="${TMPDIR}${SOURCE}"
+ cat > ${TMPSOURCE} <<HEADER
+#define SEED ${SEED}
+#include "smpi.h"
+#include "msg/msg.h"
+#include "xbt/sysdep.h"
+#include "xbt/log.h"
+#include "xbt/asserts.h"
+#define sleep(x) smpi_sleep(x)
+#define gettimeofday(x, y) smpi_gettimeofday(x, y)
+HEADER
+ # very simplistic transform, will probably want full parser for next version
+ grep -v "mpi.h" < ${SOURCE} | perl -pe 's/main/smpi_main/;' >> ${TMPSOURCE}
+ grep -q "smpi_main" ${TMPSOURCE}
+ if [ $? -eq 0 ]; then
+ cat >> ${TMPSOURCE} <<FOOTER
+XBT_LOG_NEW_DEFAULT_CATEGORY(msg_simulation, "Messages specific to this simulation");
+int main(int argc, char *argv[]) {
+ MSG_error_t result;
+ srand(SEED);
+ MSG_global_init(&argc, argv);
+ MSG_set_channel_number(MAX_CHANNEL);
+ MSG_function_register("smpi_main", smpi_main);
+ MSG_function_register("smpi_sender", smpi_sender);
+ MSG_function_register("smpi_receiver", smpi_receiver);
+ MSG_create_environment(argv[1]);
+ MSG_launch_application(argv[2]);
+ result = MSG_main();
+ INFO1("simulation time %g", MSG_get_clock());
+ MSG_clean();
+ return (MSG_OK != result);
+}
+FOOTER
+ fi
+}
+
+INCLUDEARGS=""
+LINKARGS="-L${SMPI_LIB} -lsmpi -L${SIMGRID_LIB} -lsimgrid"
+
+CMDLINE=""
+while [ -n "$1" ]; do
+ ARG="$1"
+ shift
+ if [ "${ARG}" = "-c" ]; then
+ LINKARGS=""
+ CMDLINE="${CMDLINE} -c "
+ elif [ "${ARG%.c}" != "${ARG}" ]; then
+ INCLUDEARGS="-I${SMPI_INCLUDE} -I${SIMGRID_INCLUDE} "
+ SRCFILE="$(realpath ${ARG})"
+ modsource ${SRCFILE}
+ CMDLINE="${CMDLINE} ${TMPDIR}${SRCFILE} "
+ else
+ CMDLINE="${CMDLINE} ${ARG} "
+ fi
+done
+
+CMDLINE="${CC} ${INCLUDEARGS}${CMDLINE}${LINKARGS}"
+
+#echo "${CMDLINE}"
+${CMDLINE}
+rm -rf ${TMPDIR}
--- /dev/null
+#!/bin/sh
+DEFAULT_LOOPBACK_BANDWIDTH="498000000"
+DEFAULT_LOOPBACK_LATENCY="0.000004"
+DEFAULT_NETWORK_BANDWIDTH="$((26 * 1024 * 1024))"
+DEFAULT_NETWORK_LATENCY="0.000005"
+DEFAULT_NUMPROCS="4"
+DEFAULT_POWER="100"
+
+LOOPBACK_BANDWIDTH="${DEFAULT_LOOPBACK_BANDWIDTH}"
+LOOPBACK_LATENCY="${DEFAULT_LOOPBACK_LATENCY}"
+NETWORK_BANDWIDTH="${DEFAULT_NETWORK_BANDWIDTH}"
+NETWORK_LATENCY="${DEFAULT_NETWORK_LATENCY}"
+NUMPROCS="${DEFAULT_NUMPROCS}"
+POWER="${DEFAULT_POWER}"
+
+while true; do
+ case "$1" in
+ "-np")
+ NUMPROCS="$2"
+ shift 2
+ ;;
+ "-bandwidth")
+ NETWORK_BANDWIDTH="$2"
+ shift 2
+ ;;
+ "-latency")
+ NETWORK_LATENCY="$2"
+ shift 2
+ ;;
+ *)
+ break
+ ;;
+ esac
+done
+
+EXEC="$1"
+shift
+
+PLATFORMTMP="$(mktemp tmpXXXXXX)"
+#PLATFORMTMP="pla.xml"
+
+cat > ${PLATFORMTMP} <<PLATFORMHEAD
+<?xml version='1.0'?>
+<!DOCTYPE platform_description SYSTEM "surfxml.dtd">
+<platform_description version="1">
+PLATFORMHEAD
+
+for (( i=${NUMPROCS}; $i ; i=$i-1 )) do
+ echo " <cpu name=\"host$i\" power=\"${POWER}\"/>" >> ${PLATFORMTMP}
+ echo " <network_link name=\"loop$i\" bandwidth=\"${LOOPBACK_BANDWIDTH}\" latency=\"${LOOPBACK_LATENCY}\"/>" >> ${PLATFORMTMP}
+ echo " <network_link name=\"link$i\" bandwidth=\"${NETWORK_BANDWIDTH}\" latency=\"${NETWORK_LATENCY}\"/>" >> ${PLATFORMTMP}
+done
+
+for (( i=${NUMPROCS}; $i ; i=$i-1 )) do
+ for (( j=${NUMPROCS}; $j ; j=$j-1 )) do
+ if [ $i -eq $j ]; then
+ echo " <route src=\"host$i\" dst=\"host$j\"><route_element name=\"loop$i\"/></route>" >> ${PLATFORMTMP}
+ else
+ echo " <route src=\"host$i\" dst=\"host$j\"><route_element name=\"link$i\"/><route_element name=\"link$j\"/></route>" >> ${PLATFORMTMP}
+ fi
+ done
+done
+
+cat >> ${PLATFORMTMP} <<PLATFORMFOOT
+</platform_description>
+PLATFORMFOOT
+
+APPLICATIONTMP="$(mktemp tmpXXXXXX)"
+#APPLICATIONTMP="app.xml"
+
+cat > ${APPLICATIONTMP} <<APPLICATIONHEAD
+<?xml version='1.0'?>
+<!DOCTYPE platform_description SYSTEM "surfxml.dtd">
+<platform_description version="1">
+APPLICATIONHEAD
+
+for (( i=${NUMPROCS}; $i ; i=$i-1 )) do
+ echo " <process host=\"host$i\" function=\"smpi_main\">" >> ${APPLICATIONTMP}
+ for ARG in $*; do
+ echo " <argument value=\"${ARG}\"/>" >> ${APPLICATIONTMP}
+ done
+ echo " </process>" >> ${APPLICATIONTMP}
+done
+
+for (( i=${NUMPROCS}; $i ; i=$i-1 )) do
+ echo " <process host=\"host$i\" function=\"smpi_sender\"/>" >> ${APPLICATIONTMP}
+done
+
+for (( i=${NUMPROCS}; $i ; i=$i-1 )) do
+ echo " <process host=\"host$i\" function=\"smpi_receiver\"/>" >> ${APPLICATIONTMP}
+done
+
+cat >> ${APPLICATIONTMP} <<APPLICATIONFOOT
+</platform_description>
+APPLICATIONFOOT
+
+${EXEC} ${PLATFORMTMP} ${APPLICATIONTMP}
+rm ${PLATFORMTMP} ${APPLICATIONTMP}
--- /dev/null
+#include <stdio.h>
+#include <sys/time.h>
+#include "msg/msg.h"
+#include "xbt/sysdep.h"
+#include "xbt/xbt_portability.h"
+#include "smpi.h"
+
+smpi_mpi_request_t **smpi_pending_send_requests = NULL;
+smpi_mpi_request_t **smpi_last_pending_send_requests = NULL;
+
+smpi_mpi_request_t **smpi_pending_recv_requests = NULL;
+smpi_mpi_request_t **smpi_last_pending_recv_requests = NULL;
+
+smpi_received_t **smpi_received = NULL;
+smpi_received_t **smpi_last_received = NULL;
+
+m_process_t *smpi_sender_processes = NULL;
+m_process_t *smpi_receiver_processes = NULL;
+
+int smpi_running_hosts = 0;
+
+smpi_mpi_communicator_t smpi_mpi_comm_world;
+
+smpi_mpi_status_t smpi_mpi_status_ignore;
+
+smpi_mpi_datatype_t smpi_mpi_byte;
+smpi_mpi_datatype_t smpi_mpi_int;
+smpi_mpi_datatype_t smpi_mpi_double;
+
+smpi_mpi_op_t smpi_mpi_land;
+smpi_mpi_op_t smpi_mpi_sum;
+
+static xbt_os_timer_t smpi_timer;
+static int smpi_benchmarking;
+static double smpi_reference;
+
+void smpi_mpi_land_func(void *x, void *y, void *z) {
+ *(int *)z = *(int *)x && *(int *)y;
+}
+
+void smpi_mpi_sum_func(void *x, void *y, void *z) {
+ *(int *)z = *(int *)x + *(int *)y;
+}
+
+void smpi_mpi_init() {
+ int i;
+ int size, rank;
+ m_host_t *hosts;
+ m_host_t host;
+ double duration;
+ m_task_t mtask;
+
+ // will eventually need mutex
+ smpi_running_hosts++;
+
+ // initialize some local variables
+ size = MSG_get_host_number();
+ host = MSG_host_self();
+ hosts = MSG_get_host_table();
+ for(i = 0; i < size && host != hosts[i]; i++);
+ rank = i;
+
+ // node 0 sets the globals
+ if (0 == rank) {
+
+ // global communicator
+ smpi_mpi_comm_world.id = 0;
+ smpi_mpi_comm_world.size = size;
+ smpi_mpi_comm_world.barrier = 0;
+ smpi_mpi_comm_world.hosts = hosts;
+ smpi_mpi_comm_world.processes = xbt_malloc(sizeof(m_process_t) * size);
+ smpi_mpi_comm_world.processes[0] = MSG_process_self();
+
+ // mpi datatypes
+ smpi_mpi_byte.size = (size_t)1;
+ smpi_mpi_int.size = sizeof(int);
+ smpi_mpi_double.size = sizeof(double);
+
+ // mpi operations
+ smpi_mpi_land.func = &smpi_mpi_land_func;
+ smpi_mpi_sum.func = &smpi_mpi_sum_func;
+
+ // smpi globals
+ smpi_pending_send_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
+ smpi_last_pending_send_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
+ smpi_pending_recv_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
+ smpi_last_pending_recv_requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * size);
+ smpi_received = xbt_malloc(sizeof(smpi_received_t*) * size);
+ smpi_last_received = xbt_malloc(sizeof(smpi_received_t*) * size);
+ smpi_sender_processes = xbt_malloc(sizeof(m_process_t) * size);
+ smpi_receiver_processes = xbt_malloc(sizeof(m_process_t) * size);
+ for(i = 0; i < size; i++) {
+ smpi_pending_send_requests[i] = NULL;
+ smpi_last_pending_send_requests[i] = NULL;
+ smpi_pending_recv_requests[i] = NULL;
+ smpi_last_pending_recv_requests[i] = NULL;
+ smpi_received[i] = NULL;
+ smpi_last_received[i] = NULL;
+ }
+ smpi_timer = xbt_os_timer_new();
+ smpi_reference = DEFAULT_POWER;
+ smpi_benchmarking = 0;
+
+ // tell send/recv nodes to begin
+ for(i = 0; i < size; i++) {
+ mtask = MSG_task_create("READY", 0, 0, NULL);
+ MSG_task_put(mtask, hosts[i], SEND_SYNC_PORT);
+ mtask = (m_task_t)0;
+ MSG_task_get_from_host(&mtask, SEND_SYNC_PORT, hosts[i]);
+ MSG_task_destroy(mtask);
+ mtask = MSG_task_create("READY", 0, 0, NULL);
+ MSG_task_put(mtask, hosts[i], RECV_SYNC_PORT);
+ mtask = (m_task_t)0;
+ MSG_task_get_from_host(&mtask, RECV_SYNC_PORT, hosts[i]);
+ MSG_task_destroy(mtask);
+ }
+
+ // now everyone else
+ for(i = 1; i < size; i++) {
+ mtask = MSG_task_create("READY", 0, 0, NULL);
+ MSG_task_put(mtask, hosts[i], MPI_PORT);
+ }
+
+ } else {
+ // everyone needs to wait for node 0 to finish
+ mtask = (m_task_t)0;
+ MSG_task_get(&mtask, MPI_PORT);
+ MSG_task_destroy(mtask);
+ smpi_mpi_comm_world.processes[rank] = MSG_process_self();
+ }
+
+ // now that mpi_comm_world_processes is set, it's safe to set a barrier
+ smpi_barrier(&smpi_mpi_comm_world);
+}
+
+void smpi_mpi_finalize() {
+ int i;
+ smpi_running_hosts--;
+ if (0 <= smpi_running_hosts) {
+ for(i = 0; i < smpi_mpi_comm_world.size; i++) {
+ if(MSG_process_is_suspended(smpi_sender_processes[i])) {
+ MSG_process_resume(smpi_sender_processes[i]);
+ }
+ if(MSG_process_is_suspended(smpi_receiver_processes[i])) {
+ MSG_process_resume(smpi_receiver_processes[i]);
+ }
+ }
+ } else {
+ xbt_free(smpi_mpi_comm_world.processes);
+ xbt_free(smpi_pending_send_requests);
+ xbt_free(smpi_last_pending_send_requests);
+ xbt_free(smpi_pending_recv_requests);
+ xbt_free(smpi_last_pending_recv_requests);
+ xbt_free(smpi_received);
+ xbt_free(smpi_last_received);
+ xbt_free(smpi_sender_processes);
+ xbt_free(smpi_receiver_processes);
+ xbt_os_timer_free(smpi_timer);
+ }
+}
+
+void smpi_complete(smpi_mpi_request_t *request) {
+ smpi_waitlist_node_t *current, *next;
+ request->completed = 1;
+ request->next = NULL;
+ current = request->waitlist;
+ while(NULL != current) {
+ if(MSG_process_is_suspended(current->process)) {
+ MSG_process_resume(current->process);
+ }
+ next = current->next;
+ xbt_free(current);
+ current = next;
+ }
+ request->waitlist = NULL;
+}
+
+int smpi_host_rank_self() {
+ return smpi_comm_rank(&smpi_mpi_comm_world, MSG_host_self());
+}
+
+void smpi_isend(smpi_mpi_request_t *sendreq) {
+ int rank = smpi_host_rank_self();
+ if (NULL == smpi_last_pending_send_requests[rank]) {
+ smpi_pending_send_requests[rank] = sendreq;
+ } else {
+ smpi_last_pending_send_requests[rank]->next = sendreq;
+ }
+ smpi_last_pending_send_requests[rank] = sendreq;
+ if (MSG_process_is_suspended(smpi_sender_processes[rank])) {
+ MSG_process_resume(smpi_sender_processes[rank]);
+ }
+}
+
+void smpi_match_requests(int rank) {
+ smpi_mpi_request_t *frequest, *prequest, *crequest;
+ smpi_received_t *freceived, *preceived, *creceived;
+ size_t dsize;
+ short int match;
+ frequest = smpi_pending_recv_requests[rank];
+ prequest = NULL;
+ crequest = frequest;
+ while(NULL != crequest) {
+ freceived = smpi_received[rank];
+ preceived = NULL;
+ creceived = freceived;
+ match = 0;
+ while(NULL != creceived && !match) {
+ if(crequest->comm->id == creceived->commid &&
+ (MPI_ANY_SOURCE == crequest->src || crequest->src == creceived->src) &&
+ crequest->tag == creceived->tag) {
+
+ // we have a match!
+ match = 1;
+
+ // pull the request from the queue
+ if(NULL == prequest) {
+ frequest = crequest->next;
+ smpi_pending_recv_requests[rank] = frequest;
+ } else {
+ prequest->next = crequest->next;
+ }
+ if(crequest == smpi_last_pending_recv_requests[rank]) {
+ smpi_last_pending_recv_requests[rank] = prequest;
+ }
+
+ // pull the received data from the queue
+ if(NULL == preceived) {
+ freceived = creceived->next;
+ smpi_received[rank] = freceived;
+ } else {
+ preceived->next = creceived->next;
+ }
+ if(creceived == smpi_last_received[rank]) {
+ smpi_last_received[rank] = preceived;
+ }
+
+ // for when request->src is any source
+ crequest->src = creceived->src;
+
+ // calculate data size
+ dsize = crequest->count * crequest->datatype->size;
+
+ // copy data to buffer
+ memcpy(crequest->buf, creceived->data, dsize);
+
+ // fwd through
+ crequest->fwdthrough = creceived->fwdthrough;
+
+ // get rid of received data node, no longer needed
+ xbt_free(creceived->data);
+ xbt_free(creceived);
+
+ if (crequest->fwdthrough == rank) {
+ smpi_complete(crequest);
+ } else {
+ crequest->src = rank;
+ crequest->dst = (rank + 1) % crequest->comm->size;
+ smpi_isend(crequest);
+ }
+
+ } else {
+ preceived = creceived;
+ creceived = creceived->next;
+ }
+ }
+ prequest = crequest;
+ crequest = crequest->next;
+ }
+}
+
+void smpi_bench_begin() {
+ xbt_assert0(!smpi_benchmarking, "Already benchmarking");
+ smpi_benchmarking = 1;
+ xbt_os_timer_start(smpi_timer);
+ return;
+}
+
+void smpi_bench_end() {
+ m_task_t ctask = NULL;
+ double duration;
+ xbt_assert0(smpi_benchmarking, "Not benchmarking yet");
+ smpi_benchmarking = 0;
+ xbt_os_timer_stop(smpi_timer);
+ duration = xbt_os_timer_elapsed(smpi_timer);
+ ctask = MSG_task_create("computation", duration * smpi_reference, 0 , NULL);
+ MSG_task_execute(ctask);
+ MSG_task_destroy(ctask);
+ return;
+}
+
+int smpi_create_request(void *buf, int count, smpi_mpi_datatype_t *datatype,
+ int src, int dst, int tag, smpi_mpi_communicator_t *comm, smpi_mpi_request_t **request) {
+ int retval = MPI_SUCCESS;
+ *request = NULL;
+ if (NULL == buf && 0 < count) {
+ retval = MPI_ERR_INTERN;
+ } else if (0 > count) {
+ retval = MPI_ERR_COUNT;
+ } else if (NULL == datatype) {
+ retval = MPI_ERR_TYPE;
+ } else if (NULL == comm) {
+ retval = MPI_ERR_COMM;
+ } else if (MPI_ANY_SOURCE != src && (0 > src || comm->size <= src)) {
+ retval = MPI_ERR_RANK;
+ } else if (0 > dst || comm->size <= dst) {
+ retval = MPI_ERR_RANK;
+ } else if (0 > tag) {
+ retval = MPI_ERR_TAG;
+ } else {
+ *request = xbt_malloc(sizeof(smpi_mpi_request_t));
+ (*request)->buf = buf;
+ (*request)->count = count;
+ (*request)->datatype = datatype;
+ (*request)->src = src;
+ (*request)->dst = dst;
+ (*request)->tag = tag;
+ (*request)->comm = comm;
+ (*request)->completed = 0;
+ (*request)->fwdthrough = dst;
+ (*request)->waitlist = NULL;
+ (*request)->next = NULL;
+ }
+ return retval;
+}
+
+void smpi_barrier(smpi_mpi_communicator_t *comm) {
+ int i;
+ comm->barrier++;
+ if(comm->barrier < comm->size) {
+ MSG_process_suspend(MSG_process_self());
+ } else {
+ comm->barrier = 0;
+ for(i = 0; i < comm->size; i++) {
+ if (MSG_process_is_suspended(comm->processes[i])) {
+ MSG_process_resume(comm->processes[i]);
+ }
+ }
+ }
+}
+
+int smpi_comm_rank(smpi_mpi_communicator_t *comm, m_host_t host) {
+ int i;
+ for(i = 0; i < comm->size && host != comm->hosts[i]; i++);
+ if (i >= comm->size) i = -1;
+ return i;
+}
+
+void smpi_irecv(smpi_mpi_request_t *recvreq) {
+ int rank = smpi_host_rank_self();
+ if (NULL == smpi_pending_recv_requests[rank]) {
+ smpi_pending_recv_requests[rank] = recvreq;
+ } else if (NULL != smpi_last_pending_recv_requests[rank]) {
+ smpi_last_pending_recv_requests[rank]->next = recvreq;
+ } else { // can't happen!
+ fprintf(stderr, "smpi_pending_recv_requests not null while smpi_last_pending_recv_requests null!\n");
+ }
+ smpi_last_pending_recv_requests[rank] = recvreq;
+ smpi_match_requests(rank);
+ if (MSG_process_is_suspended(smpi_receiver_processes[rank])) {
+ MSG_process_resume(smpi_receiver_processes[rank]);
+ }
+}
+
+void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status) {
+ smpi_waitlist_node_t *waitnode, *current;
+ if (NULL != request) {
+ if (!request->completed) {
+ waitnode = xbt_malloc(sizeof(smpi_waitlist_node_t));
+ waitnode->process = MSG_process_self();
+ waitnode->next = NULL;
+ if (NULL == request->waitlist) {
+ request->waitlist = waitnode;
+ } else {
+ for(current = request->waitlist; NULL != current->next; current = current->next);
+ current->next = waitnode;
+ }
+ MSG_process_suspend(waitnode->process);
+ }
+ if (NULL != status && MPI_STATUS_IGNORE != status) {
+ status->MPI_SOURCE = request->src;
+ }
+ }
+}
+
+void smpi_wait_all(int count, smpi_mpi_request_t **requests, smpi_mpi_status_t *statuses) {
+ int i;
+ for (i = 0; i < count; i++) {
+ smpi_wait(requests[i], &statuses[i]);
+ }
+}
+
+void smpi_wait_all_nostatus(int count, smpi_mpi_request_t **requests) {
+ int i;
+ for (i = 0; i < count; i++) {
+ smpi_wait(requests[i], MPI_STATUS_IGNORE);
+ }
+}
+
+int smpi_sender(int argc, char *argv[]) {
+ m_process_t process;
+ char taskname[50];
+ size_t dsize;
+ void *data;
+ m_host_t dhost;
+ m_task_t mtask;
+ int rank, fc, ft;
+ smpi_mpi_request_t *sendreq;
+
+ process = MSG_process_self();
+
+ // wait for init
+ mtask = (m_task_t)0;
+ MSG_task_get(&mtask, SEND_SYNC_PORT);
+
+ rank = smpi_host_rank_self();
+
+ smpi_sender_processes[rank] = process;
+
+ // ready!
+ MSG_task_put(mtask, MSG_task_get_source(mtask), SEND_SYNC_PORT);
+
+ while (0 < smpi_running_hosts) {
+ sendreq = smpi_pending_send_requests[rank];
+ if (NULL != sendreq) {
+
+ // pull from queue if not a fwd or no more to fwd
+ if (sendreq->dst == sendreq->fwdthrough) {
+ smpi_pending_send_requests[rank] = sendreq->next;
+ if(sendreq == smpi_last_pending_send_requests[rank]) {
+ smpi_last_pending_send_requests[rank] = NULL;
+ }
+ ft = sendreq->dst;
+ } else {
+ fc = ((sendreq->fwdthrough - sendreq->dst + sendreq->comm->size) % sendreq->comm->size) / 2;
+ ft = (sendreq->dst + fc) % sendreq->comm->size;
+ //printf("node %d sending broadcast to node %d through node %d\n", rank, sendreq->dst, ft);
+ }
+
+ // create task to send
+ sprintf(taskname, "comm:%d,src:%d,dst:%d,tag:%d,ft:%d", sendreq->comm->id, sendreq->src, sendreq->dst, sendreq->tag, ft);
+ dsize = sendreq->count * sendreq->datatype->size;
+ data = xbt_malloc(dsize);
+ memcpy(data, sendreq->buf, dsize);
+ mtask = MSG_task_create(taskname, 0, dsize, data);
+
+ // figure out which host to send it to
+ dhost = sendreq->comm->hosts[sendreq->dst];
+
+ // send task
+ #ifdef DEBUG
+ printf("host %s attempting to send to host %s\n", MSG_host_get_name(MSG_host_self()), MSG_host_get_name(dhost));
+ #endif
+ MSG_task_put(mtask, dhost, MPI_PORT);
+
+ if (sendreq->dst == sendreq->fwdthrough) {
+ smpi_complete(sendreq);
+ } else {
+ sendreq->dst = (sendreq->dst + fc + 1) % sendreq->comm->size;
+ }
+
+ } else {
+ MSG_process_suspend(process);
+ }
+ }
+ return 0;
+}
+
+int smpi_receiver(int argc, char **argv) {
+ m_process_t process;
+ m_task_t mtask;
+ smpi_received_t *received;
+ int rank;
+ smpi_mpi_request_t *recvreq;
+
+ process = MSG_process_self();
+
+ // wait for init
+ mtask = (m_task_t)0;
+ MSG_task_get(&mtask, RECV_SYNC_PORT);
+
+ rank = smpi_host_rank_self();
+
+ // potential race condition...
+ smpi_receiver_processes[rank] = process;
+
+ // ready!
+ MSG_task_put(mtask, MSG_task_get_source(mtask), RECV_SYNC_PORT);
+
+ while (0 < smpi_running_hosts) {
+ recvreq = smpi_pending_recv_requests[rank];
+ if (NULL != recvreq) {
+ mtask = (m_task_t)0;
+
+ #ifdef DEBUG
+ printf("host %s waiting to receive from anyone, but first in queue is (%d,%d,%d).\n",
+ MSG_host_get_name(MSG_host_self()), recvreq->src, recvreq->dst, recvreq->tag);
+ #endif
+ MSG_task_get(&mtask, MPI_PORT);
+
+ received = xbt_malloc(sizeof(smpi_received_t));
+
+ sscanf(MSG_task_get_name(mtask), "comm:%d,src:%d,dst:%d,tag:%d,ft:%d",
+ &received->commid, &received->src, &received->dst, &received->tag, &received->fwdthrough);
+ received->data = MSG_task_get_data(mtask);
+ received->next = NULL;
+
+ if (NULL == smpi_last_received[rank]) {
+ smpi_received[rank] = received;
+ } else {
+ smpi_last_received[rank]->next = received;
+ }
+ smpi_last_received[rank] = received;
+
+ MSG_task_destroy(mtask);
+
+ smpi_match_requests(rank);
+
+ } else {
+ MSG_process_suspend(process);
+ }
+ }
+ return 0;
+}
+
+// FIXME: move into own file
+int smpi_gettimeofday(struct timeval *tv, struct timezone *tz) {
+ double now;
+ int retval = 0;
+ smpi_bench_end();
+ if (NULL == tv) {
+ retval = -1;
+ } else {
+ now = MSG_get_clock();
+ tv->tv_sec = now;
+ tv->tv_usec = ((now - (double)tv->tv_sec) * 1000000.0);
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+unsigned int smpi_sleep(unsigned int seconds) {
+ m_task_t task = NULL;
+ smpi_bench_end();
+ task = MSG_task_create("sleep", seconds * DEFAULT_POWER, 0, NULL);
+ MSG_task_execute(task);
+ MSG_task_destroy(task);
+ smpi_bench_begin();
+ return 0;
+}
+
+void smpi_exit(int status) {
+ smpi_bench_end();
+ smpi_running_hosts--;
+ MSG_process_kill(MSG_process_self());
+ return;
+}
--- /dev/null
+#include <stdio.h>
+#include <sys/time.h>
+#include "msg/msg.h"
+#include "xbt/sysdep.h"
+#include "xbt/xbt_portability.h"
+#include "smpi.h"
+
+int MPI_Init(int *argc, char ***argv) {
+ smpi_mpi_init();
+ smpi_bench_begin();
+ return MPI_SUCCESS;
+}
+
+int MPI_Finalize() {
+ smpi_bench_end();
+ smpi_mpi_finalize();
+ return MPI_SUCCESS;
+}
+
+// right now this just exits the current node, should send abort signal to all
+// hosts in the communicator;
+int MPI_Abort(MPI_Comm comm, int errorcode) {
+ smpi_exit(errorcode);
+}
+
+int MPI_Comm_size(MPI_Comm comm, int *size) {
+ int retval = MPI_SUCCESS;
+ smpi_bench_end();
+ if (NULL == comm) {
+ retval = MPI_ERR_COMM;
+ } else if (NULL == size) {
+ retval = MPI_ERR_ARG;
+ } else {
+ *size = comm->size;
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Comm_rank(MPI_Comm comm, int *rank) {
+ int retval = MPI_SUCCESS;
+ smpi_bench_end();
+ if (NULL == comm) {
+ retval = MPI_ERR_COMM;
+ } else if (NULL == rank) {
+ retval = MPI_ERR_ARG;
+ } else {
+ *rank = smpi_comm_rank(comm, MSG_host_self());
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+/*
+int MPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm *newcomm) {
+ int retval = MPI_SUCCESS;
+ m_host_t host = MSG_host_self();
+ int rank = smpi_comm_rank(comm, host);
+ smpi_mpi_comm_split_table_node_t *split_table;
+ split_table = xbt_malloc(sizeof(smpi_mpi_comm_split_table_node_t) * comm->size);
+ split_table[rank].color = color;
+ split_table[rank].key = key;
+ split_table[rank].host = host;
+ smpi_mpi_alltoall(
+ return retval;
+}
+*/
+
+int MPI_Type_size(MPI_Datatype datatype, size_t *size) {
+ int retval = MPI_SUCCESS;
+ smpi_bench_end();
+ if (NULL == datatype) {
+ retval = MPI_ERR_TYPE;
+ } else if (NULL == size) {
+ retval = MPI_ERR_ARG;
+ } else {
+ *size = datatype->size;
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Wait(MPI_Request *request, MPI_Status *status) {
+ int retval = MPI_SUCCESS;
+ smpi_bench_end();
+ if (NULL == request) {
+ retval = MPI_ERR_REQUEST;
+ } else if (NULL == status) {
+ retval = MPI_ERR_ARG;
+ } else {
+ smpi_wait(*request, status);
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Waitall(int count, MPI_Request *requests, MPI_Status *statuses) {
+ int retval = MPI_SUCCESS;
+ smpi_bench_end();
+ if (NULL == requests) {
+ retval = MPI_ERR_REQUEST;
+ } else if (NULL == statuses) {
+ retval = MPI_ERR_ARG;
+ } else {
+ smpi_wait_all(count, requests, statuses);
+ }
+ smpi_bench_begin();
+ return MPI_ERR_INTERN;
+}
+
+int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request *request) {
+ int retval = MPI_SUCCESS;
+ int dst;
+ smpi_mpi_request_t *recvreq;
+ smpi_bench_end();
+ dst = smpi_comm_rank(comm, MSG_host_self());
+ retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &recvreq);
+ if (NULL != recvreq) {
+ smpi_irecv(recvreq);
+ if (NULL != request) {
+ *request = recvreq;
+ }
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status *status) {
+ int retval = MPI_SUCCESS;
+ int dst;
+ smpi_mpi_request_t *recvreq;
+ smpi_bench_end();
+ dst = smpi_comm_rank(comm, MSG_host_self());
+ retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &recvreq);
+ if (NULL != recvreq) {
+ smpi_irecv(recvreq);
+ smpi_wait(recvreq, status);
+ xbt_free(recvreq);
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request *request) {
+ int retval = MPI_SUCCESS;
+ int src;
+ smpi_mpi_request_t *sendreq;
+ smpi_bench_end();
+ src = smpi_comm_rank(comm, MSG_host_self());
+ retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &sendreq);
+ if (NULL != sendreq) {
+ smpi_isend(sendreq);
+ if (NULL != request) {
+ *request = sendreq;
+ }
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
+ int retval = MPI_SUCCESS;
+ int src;
+ smpi_mpi_request_t *sendreq;
+ smpi_bench_end();
+ src = smpi_comm_rank(comm, MSG_host_self());
+ retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &sendreq);
+ if (NULL != sendreq) {
+ smpi_isend(sendreq);
+ smpi_wait(sendreq, MPI_STATUS_IGNORE);
+ }
+ smpi_bench_begin();
+ return retval;
+}
+
+int MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) {
+ int i;
+ int rank;
+ smpi_mpi_request_t *request;
+ smpi_bench_end();
+
+ rank = smpi_comm_rank(comm, MSG_host_self());
+
+ if (root == rank) {
+ smpi_create_request(buf, count, datatype, root, (rank + 1) % comm->size, 0, comm, &request);
+ if (comm->size > 2) {
+ request->fwdthrough = (rank - 1 + comm->size) % comm->size;
+ }
+ smpi_isend(request);
+ smpi_wait(request, MPI_STATUS_IGNORE);
+ } else {
+ smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank, 0, comm, &request);
+ if (NULL != request) {
+ smpi_irecv(request);
+ smpi_wait(request, MPI_STATUS_IGNORE);
+ }
+ }
+
+ smpi_bench_begin();
+ return MPI_SUCCESS;
+}
+
+int MPI_Barrier(MPI_Comm comm) {
+ smpi_bench_end();
+ smpi_barrier(comm);
+ smpi_bench_begin();
+ return MPI_SUCCESS;
+}
+
+// FIXME: instead of everyone sending in order, might be a good idea to send to next...
+int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
+ int i;
+ int rank;
+ smpi_mpi_request_t **sendreqs, **recvreqs;
+ smpi_bench_end();
+
+ rank = smpi_comm_rank(comm, MSG_host_self());
+
+ sendreqs = xbt_malloc(sizeof(smpi_mpi_request_t*) * comm->size);
+ recvreqs = xbt_malloc(sizeof(smpi_mpi_request_t*) * comm->size);
+
+ for (i = 0; i < comm->size; i++) {
+ if (rank == i) {
+ sendreqs[i] = NULL;
+ recvreqs[i] = NULL;
+ memcpy(recvbuf + recvtype->size * recvcount * i, sendbuf + sendtype->size * sendcount * i, recvtype->size * recvcount);
+ } else {
+ smpi_create_request(sendbuf + sendtype->size * sendcount * i, sendcount, sendtype, rank, i, 0, comm, sendreqs + i);
+ smpi_isend(sendreqs[i]);
+ smpi_create_request(recvbuf + recvtype->size * recvcount * i, recvcount, recvtype, i, rank, 0, comm, recvreqs + i);
+ smpi_irecv(recvreqs[i]);
+ }
+ }
+
+ smpi_wait_all_nostatus(comm->size, sendreqs);
+ smpi_wait_all_nostatus(comm->size, recvreqs);
+
+ xbt_free(sendreqs);
+ xbt_free(recvreqs);
+
+ smpi_bench_begin();
+ return MPI_SUCCESS;
+}
+
+// FIXME: mpi routines shouldn't call mpi routines, complexity belongs at a lower level
+// also, there's probably a really clever way to overlap everything for this one...
+int MPI_Allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) {
+ MPI_Reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
+ MPI_Bcast(recvbuf, count, datatype, 0, comm);
+ return MPI_SUCCESS;
+}
+
+// FIXME: check if behavior defined when send/recv bufs are same...
+int MPI_Reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) {
+ int i, j;
+ int rank;
+ smpi_mpi_request_t **requests;
+ void **scratchbuf;
+ smpi_bench_end();
+
+ rank = smpi_comm_rank(comm, MSG_host_self());
+
+ if (root == rank) {
+ requests = xbt_malloc(sizeof(smpi_mpi_request_t*) * comm->size);
+ scratchbuf = xbt_malloc(sizeof(void*) * comm->size);
+ memcpy(recvbuf, sendbuf, datatype->size * count);
+ for (i = 0; i < comm->size; i++) {
+ if (rank == i) {
+ requests[i] = NULL;
+ scratchbuf[i] = NULL;
+ } else {
+ scratchbuf[i] = xbt_malloc(datatype->size * count);
+ smpi_create_request(scratchbuf[i], count, datatype, MPI_ANY_SOURCE, rank, 0, comm, requests + i);
+ smpi_irecv(requests[i]);
+ }
+ }
+ smpi_wait_all_nostatus(comm->size, requests); // FIXME: use wait_any for slight performance gain
+ for (i = 0; i < comm->size; i++) {
+ if (rank != i) {
+ for (j = 0; j < count; j++) {
+ op->func(scratchbuf[i] + datatype->size * j, recvbuf + datatype->size * j, recvbuf + datatype->size * j);
+ }
+ xbt_free(requests[i]);
+ xbt_free(scratchbuf[i]);
+ }
+ }
+ xbt_free(requests);
+ xbt_free(scratchbuf);
+ } else {
+ requests = xbt_malloc(sizeof(smpi_mpi_request_t*));
+ smpi_create_request(sendbuf, count, datatype, rank, root, 0, comm, requests);
+ smpi_isend(*requests);
+ smpi_wait(*requests, MPI_STATUS_IGNORE);
+ xbt_free(*requests);
+ xbt_free(requests);
+ }
+
+ smpi_bench_begin();
+ return MPI_SUCCESS;
+}
