+
+
#include "private.h"
+#include "smpi_coll_private.h"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi,
"Logging specific to SMPI (mpi)");
int SMPI_MPI_Init(int *argc, char ***argv)
{
- smpi_process_init(argc,argv);
+ smpi_process_init(argc, argv);
smpi_bench_begin();
return MPI_SUCCESS;
}
int SMPI_MPI_Barrier(MPI_Comm comm)
{
int retval = MPI_SUCCESS;
+ int arity=4;
smpi_bench_end();
if (NULL == comm) {
retval = MPI_ERR_COMM;
} else {
- retval = smpi_mpi_barrier(comm);
+
+ /*
+ * original implemantation:
+ * retval = smpi_mpi_barrier(comm);
+ * this one is unrealistic: it just cond_waits, means no time.
+ */
+ retval = nary_tree_barrier( comm, arity );
}
smpi_bench_begin();
return retval;
}
+/**
+ * MPI_Wait and friends
+ **/
int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status)
{
return smpi_mpi_wait(*request, status);
}
+int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
+{
+ return smpi_mpi_waitall(count, requests, status);
+}
+
+int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index,
+ MPI_Status status[])
+{
+ return smpi_mpi_waitany(count, requests, index, status);
+}
+
/**
* MPI_Bcast
**/
+
+/**
+ * flat bcast
+ **/
+int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
+int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
+ MPI_Comm comm)
+{
+ int rank;
+ int retval = MPI_SUCCESS;
+ smpi_mpi_request_t request;
+
+ rank = smpi_mpi_comm_rank(comm);
+ if (rank == root) {
+ retval = smpi_create_request(buf, count, datatype, root,
+ (root + 1) % comm->size, 0, comm, &request);
+ request->forward = comm->size - 1;
+ smpi_mpi_isend(request);
+ } else {
+ retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank,
+ 0, comm, &request);
+ smpi_mpi_irecv(request);
+ }
+
+ smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+
+ return(retval);
+
+}
+
+/**
+ * Bcast user entry point
+ **/
int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
MPI_Comm comm)
{
-
int retval = MPI_SUCCESS;
- int rank;
- smpi_mpi_request_t request;
smpi_bench_end();
- rank = smpi_mpi_comm_rank(comm);
-
- if (rank == root) {
- retval = smpi_create_request(buf, count, datatype, root,
- (root + 1) % comm->size, 0, comm, &request);
- request->forward = comm->size - 1;
- smpi_mpi_isend(request);
- } else {
- retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank,
- 0, comm, &request);
- smpi_mpi_irecv(request);
- }
-
- smpi_mpi_wait(request, MPI_STATUS_IGNORE);
- xbt_mallocator_release(smpi_global->request_mallocator, request);
+ //retval = flat_tree_bcast(buf, count, datatype, root, comm);
+ retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 );
smpi_bench_begin();
return retval;
}
+
+
+#ifdef DEBUG_REDUCE
+/**
+ * debugging helper function
+ **/
+static void print_buffer_int(void *buf, int len, const char *msg, int rank)
+{
+ int tmp, *v;
+ printf("**[%d] %s: ", rank, msg);
+ for (tmp = 0; tmp < len; tmp++) {
+ v = buf;
+ printf("[%d]", v[tmp]);
+ }
+ printf("\n");
+ free(msg);
+}
+#endif
+
/**
* MPI_Reduce
**/
+int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
+{
+ int retval = MPI_SUCCESS;
+ int rank;
+ int size;
+ int i;
+ int tag = 0;
+ smpi_mpi_request_t *requests;
+ smpi_mpi_request_t request;
+
+ smpi_bench_end();
+
+ rank = smpi_mpi_comm_rank(comm);
+ size = comm->size;
+
+ if (rank != root) { // if i am not ROOT, simply send my buffer to root
+
+#ifdef DEBUG_REDUCE
+ print_buffer_int(sendbuf, count, xbt_strdup("sndbuf"), rank);
+#endif
+ retval =
+ smpi_create_request(sendbuf, count, datatype, rank, root, tag, comm,
+ &request);
+ smpi_mpi_isend(request);
+ smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+
+ } else {
+ // i am the ROOT: wait for all buffers by creating one request by sender
+ int src;
+ requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t));
+
+ void **tmpbufs = xbt_malloc((size-1) * sizeof(void *));
+ for (i = 0; i < size-1; i++) {
+ // we need 1 buffer per request to store intermediate receptions
+ tmpbufs[i] = xbt_malloc(count * datatype->size);
+ }
+ // root: initiliaze recv buf with my own snd buf
+ memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char));
+
+ // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
+ // since we should op values as soon as one receiving request matches.
+ for (i = 0; i < size-1; i++) {
+ // reminder: for smpi_create_request() the src is always the process sending.
+ src = i < root ? i : i + 1;
+ retval = smpi_create_request(tmpbufs[i], count, datatype,
+ src, root, tag, comm, &(requests[i]));
+ if (NULL != requests[i] && MPI_SUCCESS == retval) {
+ if (MPI_SUCCESS == retval) {
+ smpi_mpi_irecv(requests[i]);
+ }
+ }
+ }
+ // now, wait for completion of all irecv's.
+ for (i = 0; i < size-1; i++) {
+ int index = MPI_UNDEFINED;
+ smpi_mpi_waitany( size-1, requests, &index, MPI_STATUS_IGNORE);
+#ifdef DEBUG_REDUCE
+ printf ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",index);
+ print_buffer_int(tmpbufs[index], count, bprintf("tmpbufs[index=%d] (value received)", index),
+ rank);
+#endif
+
+ // arg 2 is modified
+ op->func(tmpbufs[index], recvbuf, &count, &datatype);
+#ifdef DEBUG_REDUCE
+ print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank);
+#endif
+ xbt_free(tmpbufs[index]);
+ /* FIXME: with the following line, it generates an
+ * [xbt_ex/CRITICAL] Conditional list not empty 162518800.
+ */
+ // xbt_mallocator_release(smpi_global->request_mallocator, requests[index]);
+ }
+ xbt_free(requests);
+ xbt_free(tmpbufs);
+ }
+ smpi_bench_begin();
+ return retval;
+}
-int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
- MPI_Op op, int root, MPI_Comm comm )
+/**
+ * MPI_Allreduce
+ *
+ * Same as MPI_REDUCE except that the result appears in the receive buffer of all the group members.
+ **/
+int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+ MPI_Op op, MPI_Comm comm );
+int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+ MPI_Op op, MPI_Comm comm )
+{
+ int retval = MPI_SUCCESS;
+ int root=1; // arbitrary choice
+
+ smpi_bench_end();
+
+ retval = SMPI_MPI_Reduce( sendbuf, recvbuf, count, datatype, op, root, comm);
+ if (MPI_SUCCESS != retval)
+ return(retval);
+
+ retval = SMPI_MPI_Bcast( sendbuf, count, datatype, root, comm);
+ smpi_bench_begin();
+ return( retval );
+}
+
+
+/**
+ * MPI_Scatter user entry point
+ **/
+//int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype,
+// void *recvbuf, int recvcount, MPI_Datatype recvtype,int root,
+// MPI_Comm comm);
+int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype,
+ void *recvbuf, int recvcount, MPI_Datatype recvtype,
+ int root, MPI_Comm comm)
{
int retval = MPI_SUCCESS;
- int rank;
- int size;
int i;
- smpi_mpi_request_t *tabrequest;
+ int cnt=0;
+ int rank;
+ int tag=0;
+ char *cbuf; // to manipulate the void * buffers
+ smpi_mpi_request_t *requests;
+ smpi_mpi_request_t request;
+ smpi_mpi_status_t status;
+
smpi_bench_end();
rank = smpi_mpi_comm_rank(comm);
- size = comm->size;
- tabrequest = malloc((size)*sizeof(smpi_mpi_request_t));
- if (NULL==tabrequest) {
- fprintf(stderr,"[smpi] %s:%d : cannot alloc memory for %d requests. Exiting.\n",__FILE__,__LINE__,size);
- exit(1);
- }
-
- if (rank != root) { // if i am not root, simply send my buffer to root
- retval = smpi_create_request(sendbuf, count, datatype,
- rank, root, 0, comm, &(tabrequest[rank]));
- smpi_mpi_isend(tabrequest[rank]);
- smpi_mpi_wait(tabrequest[rank], MPI_STATUS_IGNORE);
- //printf("DEBUG: rank %d sent my sendbuf to root (rank %d)\n",rank,root);
- } else {
- // i am the root: wait for all buffers by creating requests
- // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
- // since we should op values as soon as one receiving request matches.
- for (i=0; i<comm->size; i++) {
- if ( rank != i ) { // except for me
- // reminder: for smpi_create_request() the src is always the process sending.
- retval = smpi_create_request(recvbuf, count, datatype, MPI_ANY_SOURCE, root,
- 0, comm, &(tabrequest[i]));
- if (NULL != tabrequest[i] && MPI_SUCCESS == retval) {
+ requests = xbt_malloc((comm->size-1) * sizeof(smpi_mpi_request_t));
+ if (rank == root) {
+ // i am the root: distribute my sendbuf
+ for (i=0; i < comm->size; i++) {
+ cbuf = sendbuf;
+ cbuf += i*sendcount*datatype->size;
+ if ( i!=root ) { // send to processes ...
+
+ retval = smpi_create_request((void *)cbuf, sendcount,
+ datatype, root, i, tag, comm, &(requests[cnt++]));
+ if (NULL != requests[cnt] && MPI_SUCCESS == retval) {
if (MPI_SUCCESS == retval) {
- smpi_mpi_irecv(tabrequest[i]);
+ smpi_mpi_isend(requests[cnt]);
}
}
+ cnt++;
+ }
+ else { // ... except if it's me.
+ memcpy(recvbuf, (void *)cbuf, recvcount*recvtype->size*sizeof(char));
}
}
- // now, wait for completion of all irecv's.
- // FIXME: we should implement smpi_waill_all for a more asynchronous behavior
- for (i=0; i<comm->size; i++) {
- if ( rank != i ) { // except for me
- smpi_mpi_wait(tabrequest[i], MPI_STATUS_IGNORE);
-
- // FIXME: the core part is here. To be written ...
- fprintf(stderr,"[smpi] %s:%d : MPI_Reduce *Not yet implemented*.\n",__FILE__,__LINE__);
+ for(i=0; i<cnt; i++) { // wait for send to complete
+ /* FIXME: waitall() should be slightly better */
+ smpi_mpi_wait(requests[i], &status);
+ xbt_mallocator_release(smpi_global->request_mallocator, requests[i]);
+ }
+ }
+ else { // i am a non-root process: wait data from the root
+ retval = smpi_create_request(recvbuf,recvcount,
+ recvtype, root, rank, tag, comm, &request);
+ if (NULL != request && MPI_SUCCESS == retval) {
+ if (MPI_SUCCESS == retval) {
+ smpi_mpi_irecv(request);
}
-
}
+ smpi_mpi_wait(request, &status);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
}
- for (i=0; i<comm->size; i++)
- xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]);
+ xbt_free(requests);
smpi_bench_begin();
return retval;
}
+
+
+
+
+
// used by comm_split to sort ranks based on key values
int smpi_compare_rankkeys(const void *a, const void *b);
int smpi_compare_rankkeys(const void *a, const void *b)
return retval;
}
-double SMPI_MPI_Wtime( void )
-{
- return ( SIMIX_get_clock() );
+double SMPI_MPI_Wtime(void)
+{
+ return (SIMIX_get_clock());
}
+
+
