int down_search(int rank, int parent, int me, int num_procs,
int *num_children, opal_list_t *children, opal_bitmap_t *relatives)
{
int i, bitmap, peer, hibit, mask, found;
orte_routed_tree_t *child;
opal_bitmap_t *relations;
/* is this me? */
if (me == rank) {
bitmap = opal_cube_dim(num_procs);
hibit = opal_hibit(rank, bitmap);
--bitmap;
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
peer = rank | mask;
if (peer < num_procs) {
child = OBJ_NEW(orte_routed_tree_t);
child->vpid = peer;
if (NULL != children) {
/* this is a direct child - add it to my list */
opal_list_append(children, &child->super);
(*num_children)++;
/* setup the relatives bitmap */
opal_bitmap_init(&child->relatives, num_procs);
/* point to the relatives */
relations = &child->relatives;
} else {
/* we are recording someone's relatives - set the bit */
opal_bitmap_set_bit(relatives, peer);
/* point to this relations */
relations = relatives;
}
/* search for this child's relatives */
down_search(0, 0, peer, num_procs, NULL, NULL, relations);
}
}
return parent;
}
/* find the children of this rank */
bitmap = opal_cube_dim(num_procs);
hibit = opal_hibit(rank, bitmap);
--bitmap;
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
peer = rank | mask;
if (peer < num_procs) {
/* execute compute on this child */
if (0 <= (found = down_search(peer, rank, me, num_procs, num_children, children, relatives))) {
return found;
}
}
}
return -1;
}
/*
* bcast_log_intra
*
* Function: - broadcast using O(log(N)) algorithm
* Accepts: - same arguments as MPI_Bcast()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_bcast_log_intra(void *buff, int count,
struct ompi_datatype_t *datatype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int size;
int rank;
int vrank;
int peer;
int dim;
int hibit;
int mask;
int err;
int nreqs;
ompi_request_t **preq;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
ompi_request_t **reqs = basic_module->mccb_reqs;
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
vrank = (rank + size - root) % size;
dim = comm->c_cube_dim;
hibit = opal_hibit(vrank, dim);
--dim;
/* Receive data from parent in the tree. */
if (vrank > 0) {
peer = ((vrank & ~(1 << hibit)) + root) % size;
err = MCA_PML_CALL(recv(buff, count, datatype, peer,
MCA_COLL_BASE_TAG_BCAST,
comm, MPI_STATUS_IGNORE));
if (MPI_SUCCESS != err) {
return err;
}
}
/* Send data to the children. */
err = MPI_SUCCESS;
preq = reqs;
nreqs = 0;
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
peer = vrank | mask;
if (peer < size) {
peer = (peer + root) % size;
++nreqs;
err = MCA_PML_CALL(isend_init(buff, count, datatype, peer,
MCA_COLL_BASE_TAG_BCAST,
MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) {
mca_coll_basic_free_reqs(reqs, nreqs);
return err;
}
}
}
/* Start and wait on all requests. */
if (nreqs > 0) {
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, reqs));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error.
* The PML will finish all requests, even if one or more of them
* fail. i.e., by the end of this call, all the requests are
* free-able. So free them anyway -- even if there was an
* error, and return the error after we free everything. */
err = ompi_request_wait_all(nreqs, reqs, MPI_STATUSES_IGNORE);
/* Free the reqs */
mca_coll_basic_free_reqs(reqs, nreqs);
}
/* All done */
return err;
}
static int mca_oob_xcast_binomial_tree(orte_jobid_t job,
bool process_first,
orte_buffer_t* buffer,
orte_gpr_trigger_cb_fn_t cbfunc)
{
orte_std_cntr_t i;
int rc;
int tag = ORTE_RML_TAG_XCAST;
int peer, size, rank, hibit, mask;
orte_buffer_t rbuf, sbuf;
orte_gpr_notify_message_t *msg;
orte_process_name_t target;
/* check to see if there is something to send - this is only true on the HNP end.
* However, we cannot just test to see if we are the HNP since, if we are a singleton,
* we are the HNP *and* we still need to handle both ends of the xcast
*/
if (NULL != buffer) {
/* this is the HNP end, so it starts the procedure. Accordingly, it sends its
* message to the first process in the job in the peer list, which takes it from there
*/
OBJ_CONSTRUCT(&xcastmutex, opal_mutex_t);
OPAL_THREAD_LOCK(&xcastmutex);
target.cellid = ORTE_PROC_MY_NAME->cellid;
target.jobid = job;
target.vpid = 0;
if (0 > (rc = mca_oob_send_packed(&target, buffer, tag, 0))) {
ORTE_ERROR_LOG(rc);
OPAL_THREAD_UNLOCK(&xcastmutex);
OBJ_DESTRUCT(&xcastmutex);
return rc;
}
OPAL_THREAD_UNLOCK(&xcastmutex);
OBJ_DESTRUCT(&xcastmutex);
return ORTE_SUCCESS;
}
/* this process is one of the application procs - accordingly, it will
* receive the message from its "parent" in the broadcast tree, and
* then send it along to some set of children
*/
/* compute the bitmap, if we haven't already done so */
if (!bitmap_init) {
bitmap_save = opal_cube_dim((int)orte_process_info.num_procs);
bitmap_init = true;
}
xcast_bitmap = bitmap_save;
rank = (int)(ORTE_PROC_MY_NAME->vpid);
size = (int)orte_process_info.num_procs;
hibit = opal_hibit(rank, xcast_bitmap);
--xcast_bitmap;
/* regardless of who we are, we first have to receive the message */
OBJ_CONSTRUCT(&rbuf, orte_buffer_t);
if (0 > (rc = mca_oob_recv_packed(ORTE_NAME_WILDCARD, &rbuf, tag))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&rbuf);
return rc;
}
msg = OBJ_NEW(orte_gpr_notify_message_t);
if (NULL == msg) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
i=1;
if (ORTE_SUCCESS != (rc = orte_dss.unpack(&rbuf, &msg, &i, ORTE_GPR_NOTIFY_MSG))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_DESTRUCT(&rbuf);
/* repack the message so we can send it on */
OBJ_CONSTRUCT(&sbuf, orte_buffer_t);
if (ORTE_SUCCESS != (rc = orte_dss.pack(&sbuf, &msg, 1, ORTE_GPR_NOTIFY_MSG))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&sbuf);
return rc;
}
/* since the OOB contact info for our peers is in the STG1 message, we have to
* process it BEFORE we can relay the message to any "children"
*/
if (cbfunc != NULL && process_first) {
/* process the message */
cbfunc(msg);
}
/* send data to any children */
target.cellid = ORTE_PROC_MY_NAME->cellid;
target.jobid = ORTE_PROC_MY_NAME->jobid;
for (i = hibit + 1, mask = 1 << i; i <= xcast_bitmap; ++i, mask <<= 1) {
peer = rank | mask;
if (peer < size) {
target.vpid = (orte_vpid_t)peer;
if (0 > (rc = mca_oob_send_packed(&target, &sbuf, tag, 0))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
}
}
OBJ_DESTRUCT(&sbuf);
/* if it wasn't the STG1 message, then process it here */
if (cbfunc != NULL && !process_first) {
cbfunc(msg);
}
OBJ_RELEASE(msg);
return ORTE_SUCCESS;
}
static int binomial_tree(int rank, int parent, int me, int num_procs,
int *nchildren, opal_list_t *childrn,
opal_bitmap_t *relatives, bool mine)
{
int i, bitmap, peer, hibit, mask, found;
orte_routed_tree_t *child;
opal_bitmap_t *relations;
OPAL_OUTPUT_VERBOSE((3, orte_routed_base_framework.framework_output,
"%s routed:binomial rank %d parent %d me %d num_procs %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
rank, parent, me, num_procs));
/* is this me? */
if (me == rank) {
bitmap = opal_cube_dim(num_procs);
hibit = opal_hibit(rank, bitmap);
--bitmap;
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
peer = rank | mask;
if (peer < num_procs) {
child = OBJ_NEW(orte_routed_tree_t);
child->vpid = peer;
OPAL_OUTPUT_VERBOSE((3, orte_routed_base_framework.framework_output,
"%s routed:binomial %d found child %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
rank,
ORTE_VPID_PRINT(child->vpid)));
if (mine) {
/* this is a direct child - add it to my list */
opal_list_append(childrn, &child->super);
(*nchildren)++;
/* setup the relatives bitmap */
opal_bitmap_init(&child->relatives, num_procs);
/* point to the relatives */
relations = &child->relatives;
} else {
/* we are recording someone's relatives - set the bit */
opal_bitmap_set_bit(relatives, peer);
/* point to this relations */
relations = relatives;
}
/* search for this child's relatives */
binomial_tree(0, 0, peer, num_procs, nchildren, childrn, relations, false);
}
}
return parent;
}
/* find the children of this rank */
OPAL_OUTPUT_VERBOSE((5, orte_routed_base_framework.framework_output,
"%s routed:binomial find children of rank %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), rank));
bitmap = opal_cube_dim(num_procs);
hibit = opal_hibit(rank, bitmap);
--bitmap;
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
peer = rank | mask;
OPAL_OUTPUT_VERBOSE((5, orte_routed_base_framework.framework_output,
"%s routed:binomial find children checking peer %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), peer));
if (peer < num_procs) {
OPAL_OUTPUT_VERBOSE((5, orte_routed_base_framework.framework_output,
"%s routed:binomial find children computing tree",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* execute compute on this child */
if (0 <= (found = binomial_tree(peer, rank, me, num_procs, nchildren, childrn, relatives, mine))) {
OPAL_OUTPUT_VERBOSE((5, orte_routed_base_framework.framework_output,
"%s routed:binomial find children returning found value %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), found));
return found;
}
}
}
return -1;
}
/** adapted from openmpi see,
~/openmpi-1.6.4/ompi/mca/coll/basic/coll_basic_bcast.c
This does a binomial tree based broadcast.
*/
int MPI_Bcast_log( void *buff, int count, MPI_Datatype datatype, int root,
MPI_Comm comm )
{
int i;
int size;
int rank;
int vrank;
int peer;
int dim;
int hibit;
int mask;
int err;
int rcnt=0;
MPI_Request *areqs = NULL;
MPI_CHECK( err = MPI_Comm_rank( comm, &rank ) );
MPI_CHECK( err = MPI_Comm_size( comm, &size ) );
vrank = (rank + size - root) % size;
dim = opal_cube_dim(size);
hibit = opal_hibit(vrank, dim);
--dim;
/* malloc the maximum possible send request, this is the dimension
minus the depth then one more the leaf, this might allocate a
more slots than necessary if the number of nodes do not evenly
divide into the tree.
*/
NULL_CHECK( areqs = malloc( sizeof( MPI_Request ) * ( dim-(hibit+1)+1 ) ) );
/* Receive data from parent in the tree. */
if (vrank > 0) {
peer = ((vrank & ~(1 << hibit)) + root) % size;
// printf("Getting ready to receive from %d at %d\n", peer, rank);
MPI_CHECK( err = MPI_Recv( buff, count, datatype, peer, BCAST_TAG, comm, MPI_STATUS_IGNORE ) );
if (MPI_SUCCESS != err) {
return err;
}
}
/* Send data to the children. */
// just in case there are no children, return should be success
err = MPI_SUCCESS;
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
peer = vrank | mask;
if (peer < size) {
peer = (peer + root) % size;
MPI_CHECK( err = MPI_Isend( buff, count, datatype, peer,
BCAST_TAG, comm,
&areqs[ rcnt ] ) );
rcnt++;
if (MPI_SUCCESS != err) {
free(areqs);
return err;
}
}
}
/* Wait on all requests. */
if (rcnt > 0) {
MPI_CHECK( err = MPI_Waitall( rcnt, areqs, MPI_STATUSES_IGNORE ) );
/* Free the reqs */
free(areqs);
}
/* All done */
return err;
}
/*
* barrier_intra_log
*
* Function: - barrier using O(log(N)) algorithm
* Accepts: - same as MPI_Barrier()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_barrier_intra_log(struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int err;
int peer;
int dim;
int hibit;
int mask;
int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm);
/* Send null-messages up and down the tree. Synchronization at the
* root (rank 0). */
dim = comm->c_cube_dim;
hibit = opal_hibit(rank, dim);
--dim;
/* Receive from children. */
for (i = dim, mask = 1 << i; i > hibit; --i, mask >>= 1) {
peer = rank | mask;
if (peer < size) {
err = MCA_PML_CALL(recv(NULL, 0, MPI_BYTE, peer,
MCA_COLL_BASE_TAG_BARRIER,
comm, MPI_STATUS_IGNORE));
if (MPI_SUCCESS != err) {
return err;
}
}
}
/* Send to and receive from parent. */
if (rank > 0) {
peer = rank & ~(1 << hibit);
err =
MCA_PML_CALL(send
(NULL, 0, MPI_BYTE, peer,
MCA_COLL_BASE_TAG_BARRIER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != err) {
return err;
}
err = MCA_PML_CALL(recv(NULL, 0, MPI_BYTE, peer,
MCA_COLL_BASE_TAG_BARRIER,
comm, MPI_STATUS_IGNORE));
if (MPI_SUCCESS != err) {
return err;
}
}
/* Send to children. */
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
peer = rank | mask;
if (peer < size) {
err = MCA_PML_CALL(send(NULL, 0, MPI_BYTE, peer,
MCA_COLL_BASE_TAG_BARRIER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != err) {
return err;
}
}
}
/* All done */
return MPI_SUCCESS;
}
int
ompi_coll_portals4_barrier_intra(struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_portals4_module_t *portals4_module = (mca_coll_portals4_module_t*) module;
int ret, i, dim, hibit, mask, num_msgs;
int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm);
ptl_ct_event_t ct;
ptl_handle_ct_t ct_h;
ptl_handle_me_t me_h;
ptl_me_t me;
size_t count;
ptl_match_bits_t match_bits;
ptl_handle_md_t md_h;
void *base;
ompi_coll_portals4_get_md(0, &md_h, &base);
count = opal_atomic_add_size_t(&portals4_module->barrier_count, 1);
ret = PtlCTAlloc(mca_coll_portals4_component.ni_h,
&ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm),
0, COLL_PORTALS4_BARRIER, count);
/* Build "tree" out of hypercube */
dim = comm->c_cube_dim;
hibit = opal_hibit(rank, dim);
--dim;
/* receive space */
me.start = NULL;
me.length = 0;
me.ct_handle = ct_h;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT | PTL_ME_EVENT_SUCCESS_DISABLE |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM | PTL_ME_EVENT_CT_OVERFLOW;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = match_bits;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* calculate number of children to receive from */
num_msgs = ompi_coll_portals4_get_nchildren(dim + 1, hibit, rank, size);
/* send to parent when children have sent to us */
if (rank > 0) {
int parent = rank & ~(1 << hibit);
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, parent),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* we'll need to wait for the parent response before the next set of comms */
num_msgs++;
}
/* send to children when parent (or all children if root) has sent to us */
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
int peer = rank | mask;
if (peer < size) {
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, peer),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
}
}
/* Wait for all incoming messages */
ret = PtlCTWait(ct_h, num_msgs, &ct);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTWait failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* cleanup */
ret = PtlMEUnlink(me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEUnlink failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlCTFree(ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTFree failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
/*
The Binomial Spanning Tree algorithm.
Outlay:
The game scales with log2(NP) and uses 1 byte of memory.
*/
static int __algorithm_binomial_tree(struct oshmem_group_t *group,
int PE_root,
void *target,
const void *source,
size_t nlong,
long *pSync)
{
int rc = OSHMEM_SUCCESS;
long value = SHMEM_SYNC_INIT;
int root_id = oshmem_proc_group_find_id(group, PE_root);
int my_id = oshmem_proc_group_find_id(group, group->my_pe);
int peer_id = 0;
int peer_pe = 0;
int vrank;
int dim = opal_cube_dim(group->proc_count);
int hibit;
int mask;
int i = 0;
SCOLL_VERBOSE(12, "[#%d] Broadcast algorithm: Tree", group->my_pe);
SCOLL_VERBOSE(15,
"[#%d] pSync[0] = %ld root = #%d",
group->my_pe, pSync[0], PE_root);
vrank = (my_id + group->proc_count - root_id) % group->proc_count;
hibit = opal_hibit(vrank, dim);
SCOLL_VERBOSE(15,
"[#%d] dim = %d vrank = %d hibit = %d",
group->my_pe, dim, vrank, hibit);
dim--;
pSync[0] = SHMEM_SYNC_READY;
/* Receive data from parent in the tree. */
if (vrank > 0) {
value = SHMEM_SYNC_READY;
SCOLL_VERBOSE(14, "[#%d] wait", group->my_pe);
rc = MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_NE, (void*)&value, SHMEM_LONG));
while ((value = pSync[0]) < 0) {
SCOLL_VERBOSE(14,
"[#%d] Broadcast size is a negative value (%li)\n",
group->my_pe, pSync[0]);
MCA_SPML_CALL(wait((void*)pSync, SHMEM_CMP_NE, (void*)&value, SHMEM_LONG));
}
if (OSHMEM_SUCCESS != rc) {
return rc;
}
nlong = (size_t) pSync[0];
}
/* Send data to the children. */
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
peer_id = vrank | mask;
if (peer_id < group->proc_count) {
/* Wait for the child to be ready to receive (pSync must have the initial value) */
peer_id = (peer_id + root_id) % group->proc_count;
peer_pe = oshmem_proc_pe(group->proc_array[peer_id]);
SCOLL_VERBOSE(14,
"[#%d] check remote pe is ready to receive #%d",
group->my_pe, peer_pe);
do {
rc = MCA_SPML_CALL(get((void*)pSync, sizeof(long), (void*)pSync, peer_pe));
} while ((OSHMEM_SUCCESS == rc) && (pSync[0] != SHMEM_SYNC_READY));
SCOLL_VERBOSE(14, "[#%d] send data to #%d", group->my_pe, peer_pe);
rc = MCA_SPML_CALL(put(target, nlong, (my_id == root_id ? (void *)source : target), peer_pe));
MCA_SPML_CALL(fence());
SCOLL_VERBOSE(14, "[#%d] signals to #%d", group->my_pe, peer_pe);
value = nlong;
rc = MCA_SPML_CALL(put((void*)pSync, sizeof(value), (void*)&value, peer_pe));
if (OSHMEM_SUCCESS != rc) {
break;
}
}
}
return rc;
}
