int ompi_coll_tuned_bcast_intra_do_this(void *buf, int count,
struct ompi_datatype_t *dtype,
int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module,
int algorithm, int faninout, int segsize)
{
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:bcast_intra_do_this algorithm %d topo faninout %d segsize %d",
algorithm, faninout, segsize));
switch (algorithm) {
case (0):
return ompi_coll_tuned_bcast_intra_dec_fixed( buf, count, dtype, root, comm, module );
case (1):
return ompi_coll_tuned_bcast_intra_basic_linear( buf, count, dtype, root, comm, module );
case (2):
return ompi_coll_tuned_bcast_intra_chain( buf, count, dtype, root, comm, module, segsize, faninout );
case (3):
return ompi_coll_tuned_bcast_intra_pipeline( buf, count, dtype, root, comm, module, segsize );
case (4):
return ompi_coll_tuned_bcast_intra_split_bintree( buf, count, dtype, root, comm, module, segsize );
case (5):
return ompi_coll_tuned_bcast_intra_bintree( buf, count, dtype, root, comm, module, segsize );
case (6):
return ompi_coll_tuned_bcast_intra_binomial( buf, count, dtype, root, comm, module, segsize );
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:bcast_intra_do_this attempt to select algorithm %d when only 0-%d is valid?",
algorithm, ompi_coll_tuned_forced_max_algorithms[BCAST]));
} /* switch */
return (MPI_ERR_ARG);
}
int ompi_coll_tuned_bcast_intra_do_forced(void *buf, int count,
struct ompi_datatype_t *dtype,
int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:bcast_intra_do_forced algorithm %d",
data->user_forced[BCAST].algorithm));
switch (data->user_forced[BCAST].algorithm) {
case (0):
return ompi_coll_tuned_bcast_intra_dec_fixed( buf, count, dtype, root, comm, module );
case (1):
return ompi_coll_tuned_bcast_intra_basic_linear( buf, count, dtype, root, comm, module );
case (2):
return ompi_coll_tuned_bcast_intra_chain( buf, count, dtype, root, comm, module,
data->user_forced[BCAST].segsize,
data->user_forced[BCAST].chain_fanout );
case (3):
return ompi_coll_tuned_bcast_intra_pipeline( buf, count, dtype, root, comm, module,
data->user_forced[BCAST].segsize );
case (4):
return ompi_coll_tuned_bcast_intra_split_bintree( buf, count, dtype, root, comm, module,
data->user_forced[BCAST].segsize );
case (5):
return ompi_coll_tuned_bcast_intra_bintree( buf, count, dtype, root, comm, module,
data->user_forced[BCAST].segsize );
case (6):
return ompi_coll_tuned_bcast_intra_binomial( buf, count, dtype, root, comm, module,
data->user_forced[BCAST].segsize );
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:bcast_intra_do_forced attempt to select algorithm %d when only 0-%d is valid?",
data->user_forced[BCAST].algorithm, ompi_coll_tuned_forced_max_algorithms[BCAST]));
} /* switch */
return (MPI_ERR_ARG);
}
/*
* bcast_intra_dec
*
* Function: - seletects broadcast algorithm to use
* Accepts: - same arguments as MPI_Bcast()
* Returns: - MPI_SUCCESS or error code (passed from the bcast implementation)
*/
int ompi_coll_tuned_bcast_intra_dec_fixed(void *buff, int count,
struct ompi_datatype_t *datatype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
/* Decision function based on MX results for
messages up to 36MB and communicator sizes up to 64 nodes */
const size_t small_message_size = 2048;
const size_t intermediate_message_size = 370728;
const double a_p16 = 3.2118e-6; /* [1 / byte] */
const double b_p16 = 8.7936;
const double a_p64 = 2.3679e-6; /* [1 / byte] */
const double b_p64 = 1.1787;
const double a_p128 = 1.6134e-6; /* [1 / byte] */
const double b_p128 = 2.1102;
int communicator_size;
int segsize = 0;
size_t message_size, dsize;
communicator_size = ompi_comm_size(comm);
/* else we need data size for decision function */
ompi_ddt_type_size(datatype, &dsize);
message_size = dsize * (unsigned long)count; /* needed for decision */
OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_bcast_intra_dec_fixed"
" root %d rank %d com_size %d msg_length %lu",
root, ompi_comm_rank(comm), communicator_size, (unsigned long)message_size));
/* Handle messages of small and intermediate size, and
single-element broadcasts */
if ((message_size < small_message_size) || (count <= 1)) {
/* Binomial without segmentation */
segsize = 0;
return ompi_coll_tuned_bcast_intra_binomial (buff, count, datatype,
root, comm, module,
segsize);
} else if (message_size < intermediate_message_size) {
/* SplittedBinary with 1KB segments */
segsize = 1024;
return ompi_coll_tuned_bcast_intra_split_bintree(buff, count, datatype,
root, comm, module,
segsize);
}
/* Handle large message sizes */
else if (communicator_size < (a_p128 * message_size + b_p128)) {
/* Pipeline with 128KB segments */
segsize = 1024 << 7;
return ompi_coll_tuned_bcast_intra_pipeline (buff, count, datatype,
root, comm, module,
segsize);
} else if (communicator_size < 13) {
/* Split Binary with 8KB segments */
segsize = 1024 << 3;
return ompi_coll_tuned_bcast_intra_split_bintree(buff, count, datatype,
root, comm, module,
segsize);
} else if (communicator_size < (a_p64 * message_size + b_p64)) {
/* Pipeline with 64KB segments */
segsize = 1024 << 6;
return ompi_coll_tuned_bcast_intra_pipeline (buff, count, datatype,
root, comm, module,
segsize);
} else if (communicator_size < (a_p16 * message_size + b_p16)) {
/* Pipeline with 16KB segments */
segsize = 1024 << 4;
return ompi_coll_tuned_bcast_intra_pipeline (buff, count, datatype,
root, comm, module,
segsize);
}
/* Pipeline with 8KB segments */
segsize = 1024 << 3;
return ompi_coll_tuned_bcast_intra_pipeline (buff, count, datatype,
root, comm, module,
segsize);
#if 0
/* this is based on gige measurements */
if (communicator_size < 4) {
return ompi_coll_tuned_bcast_intra_basic_linear (buff, count, datatype, root, comm, module);
}
if (communicator_size == 4) {
if (message_size < 524288) segsize = 0;
else segsize = 16384;
return ompi_coll_tuned_bcast_intra_bintree (buff, count, datatype, root, comm, module, segsize);
}
if (communicator_size <= 8 && message_size < 4096) {
return ompi_coll_tuned_bcast_intra_basic_linear (buff, count, datatype, root, comm, module);
}
if (communicator_size > 8 && message_size >= 32768 && message_size < 524288) {
segsize = 16384;
return ompi_coll_tuned_bcast_intra_bintree (buff, count, datatype, root, comm, module, segsize);
}
if (message_size >= 524288) {
segsize = 16384;
return ompi_coll_tuned_bcast_intra_pipeline (buff, count, datatype, root, comm, module, segsize);
}
segsize = 0;
/* once tested can swap this back in */
/* return ompi_coll_tuned_bcast_intra_bmtree (buff, count, datatype, root, comm, segsize); */
return ompi_coll_tuned_bcast_intra_bintree (buff, count, datatype, root, comm, module, segsize);
#endif /* 0 */
}
