int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_logical;
ptl_pt_index_t logical_pt_index;
ptl_process_t myself;
struct timeval start, stop;
int potato = 0;
ENTRY_T potato_catcher;
HANDLE_T potato_catcher_handle;
ptl_md_t potato_launcher;
ptl_handle_md_t potato_launcher_handle;
int num_procs;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
num_procs = libtest_get_size();
if (NULL != getenv("MAKELEVEL") && num_procs > 2) {
return 77;
}
CHECK_RETURNVAL(PtlNIInit
(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL, PTL_PID_ANY,
NULL, NULL, &ni_logical));
CHECK_RETURNVAL(PtlSetMap(ni_logical, num_procs,
libtest_get_mapping(ni_logical)));
CHECK_RETURNVAL(PtlGetId(ni_logical, &myself));
CHECK_RETURNVAL(PtlPTAlloc
(ni_logical, 0, PTL_EQ_NONE, PTL_PT_ANY,
&logical_pt_index));
assert(logical_pt_index == 0);
/* Now do the initial setup on ni_logical */
potato_catcher.start = &potato;
potato_catcher.length = sizeof(potato);
potato_catcher.uid = PTL_UID_ANY;
potato_catcher.options = OPTIONS;
#if INTERFACE == 1
potato_catcher.match_id.rank = PTL_RANK_ANY;
potato_catcher.match_bits = 1;
potato_catcher.ignore_bits = ~potato_catcher.match_bits;
#endif
CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_catcher.ct_handle));
CHECK_RETURNVAL(APPEND
(ni_logical, logical_pt_index, &potato_catcher,
PTL_PRIORITY_LIST, NULL, &potato_catcher_handle));
/* Now do a barrier (on ni_physical) to make sure that everyone has their
* logical interface set up */
libtest_barrier();
/* now I can communicate between ranks with ni_logical */
/* set up the potato launcher */
potato_launcher.start = &potato;
potato_launcher.length = sizeof(potato);
potato_launcher.options =
PTL_MD_EVENT_CT_ACK | PTL_MD_EVENT_CT_SEND;
potato_launcher.eq_handle = PTL_EQ_NONE; // i.e. don't queue send events
CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_launcher.ct_handle));
CHECK_RETURNVAL(PtlMDBind
(ni_logical, &potato_launcher, &potato_launcher_handle));
/* rank 0 starts the potato going */
if (myself.rank == 0) {
ptl_process_t nextrank;
nextrank.rank = myself.rank + 1;
nextrank.rank *= (nextrank.rank <= num_procs - 1);
gettimeofday(&start, NULL);
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0, potato_launcher.length,
(LOOPS == 1) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 1, 0,
NULL, 1));
}
{ /* the potato-passing loop */
size_t waitfor;
ptl_ct_event_t ctc;
ptl_process_t nextrank;
nextrank.rank = myself.rank + 1;
nextrank.rank *= (nextrank.rank <= num_procs - 1);
for (waitfor = 1; waitfor <= LOOPS; ++waitfor) {
CHECK_RETURNVAL(PtlCTWait(potato_catcher.ct_handle, waitfor, &ctc)); // wait for potato
assert(ctc.failure == 0);
assert(ctc.success == waitfor);
/* I have the potato! */
++potato;
if (potato < LOOPS * (num_procs)) { // otherwise, the recipient may have exited
/* Bomb's away! */
if (myself.rank == 0) {
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
potato_launcher.length,
(waitfor == (LOOPS - 1)) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 3, 0, NULL, 2));
} else {
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
potato_launcher.length,
(waitfor == LOOPS) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 3, 0, NULL, 2));
}
}
}
// make sure that last send completed before exiting
CHECK_RETURNVAL(PtlCTWait(potato_launcher.ct_handle, LOOPS+1, &ctc));
assert(ctc.failure == 0);
}
if (myself.rank == 0) {
double accumulate = 0.0;
gettimeofday(&stop, NULL);
accumulate =
(stop.tv_sec + stop.tv_usec * 1e-6) - (start.tv_sec +
start.tv_usec * 1e-6);
/* calculate the average time waiting */
printf("Total time: %g secs\n", accumulate);
accumulate /= LOOPS;
printf("Average time around the loop: %g microseconds\n",
accumulate * 1e6);
accumulate /= num_procs;
printf("Average catch-to-toss latency: %g microseconds\n",
accumulate * 1e6);
}
/* cleanup */
CHECK_RETURNVAL(PtlMDRelease(potato_launcher_handle));
CHECK_RETURNVAL(PtlCTFree(potato_launcher.ct_handle));
CHECK_RETURNVAL(UNLINK(potato_catcher_handle));
CHECK_RETURNVAL(PtlCTFree(potato_catcher.ct_handle));
/* major cleanup */
CHECK_RETURNVAL(PtlPTFree(ni_logical, logical_pt_index));
CHECK_RETURNVAL(PtlNIFini(ni_logical));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
static int
component_init(bool enable_progress_threads, bool enable_mpi_threads)
{
int ret;
ptl_ni_limits_t actual;
ret = PtlInit();
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlInit failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlNIInit(PTL_IFACE_DEFAULT,
PTL_NI_PHYSICAL | PTL_NI_MATCHING,
PTL_PID_ANY,
NULL,
&actual,
&mca_osc_portals4_component.matching_ni_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlNIInit failed: %d\n",
__FILE__, __LINE__, ret);
return ret;
}
/* BWB: FIX ME: Need to make sure our ID matches with the MTL... */
mca_osc_portals4_component.matching_atomic_max = actual.max_atomic_size;
mca_osc_portals4_component.matching_fetch_atomic_max = actual.max_fetch_atomic_size;
mca_osc_portals4_component.matching_atomic_ordered_size =
MAX(actual.max_waw_ordered_size, actual.max_war_ordered_size);
ret = PtlEQAlloc(mca_osc_portals4_component.matching_ni_h,
4096,
&mca_osc_portals4_component.matching_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return ret;
}
ret = PtlPTAlloc(mca_osc_portals4_component.matching_ni_h,
0,
mca_osc_portals4_component.matching_eq_h,
4,
&mca_osc_portals4_component.matching_pt_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return ret;
}
OBJ_CONSTRUCT(&mca_osc_portals4_component.requests, opal_free_list_t);
ret = opal_free_list_init (&mca_osc_portals4_component.requests,
sizeof(ompi_osc_portals4_request_t),
opal_cache_line_size,
OBJ_CLASS(ompi_osc_portals4_request_t),
0, 0, 8, 0, 8, NULL, 0, NULL, NULL, NULL);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: opal_free_list_init failed: %d\n",
__FILE__, __LINE__, ret);
return ret;
}
ret = opal_progress_register(progress_callback);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: opal_progress_register failed: %d\n",
__FILE__, __LINE__, ret);
return ret;
}
return OMPI_SUCCESS;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_handle;
ptl_process_t *procs;
int rank;
ptl_pt_index_t pt_index, signal_pt_index;
HANDLE_T signal_e_handle;
HANDLE_T signal_e2_handle;
int num_procs;
ptl_handle_eq_t eq_handle;
ptl_handle_ct_t ct_handle;
ptl_handle_md_t md_handle;
ptl_ni_limits_t limits_reqd, limits_actual;
ENTRY_T value_e;
limits_reqd.max_entries = 1024;
limits_reqd.max_unexpected_headers = ITERS*2;
limits_reqd.max_mds = 1024;
limits_reqd.max_eqs = 1024;
limits_reqd.max_cts = 1024;
limits_reqd.max_pt_index = 64;
limits_reqd.max_iovecs = 1024;
limits_reqd.max_list_size = 1024;
limits_reqd.max_triggered_ops = 1024;
limits_reqd.max_msg_size = 1048576;
limits_reqd.max_atomic_size = 1048576;
limits_reqd.max_fetch_atomic_size = 1048576;
limits_reqd.max_waw_ordered_size = 1048576;
limits_reqd.max_war_ordered_size = 1048576;
limits_reqd.max_volatile_size = 1048576;
limits_reqd.features = 0;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
if (num_procs < 2) {
fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
return 77;
}
int iters;
if (num_procs < ITERS)
iters = ITERS*2+1;
else
iters = ITERS;
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, &limits_reqd, &limits_actual, &ni_handle));
procs = libtest_get_mapping(ni_handle);
CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));
if (0 == rank) {
/* create data PT space */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * iters + 64, &eq_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
&pt_index));
/* create signal ME */
CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 1, eq_handle, 6,
&signal_pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = ct_handle;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_OVERFLOW_LIST, NULL, &signal_e_handle));
} else {
ptl_md_t md;
/* 16 extra just in case... */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, iters*2 + 16, &eq_handle));
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = eq_handle;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
}
fprintf(stderr,"at barrier \n");
libtest_barrier();
if (0 == rank) {
ptl_ct_event_t ct;
ptl_event_t ev;
int ret, count = 0, saw_flowctl = 0;
fprintf(stderr,"begin ctwait \n");
/* wait for signal counts */
CHECK_RETURNVAL(PtlCTWait(ct_handle, iters / 2 , &ct));
if (ct.success != iters / 2 || ct.failure != 0) {
return 1;
}
fprintf(stderr,"done CT wait \n");
/* wait for event entries */
while (1) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_OK == ret) {
count++;
fprintf(stderr, "found EQ value \n");
} else if (ret == PTL_EQ_EMPTY) {
continue;
} else {
fprintf(stderr, "0: Unexpected return code from EQGet: %d\n", ret);
return 1;
}
if (ev.type == PTL_EVENT_PT_DISABLED) {
saw_flowctl++;
break;
}
}
fprintf(stderr, "0: Saw %d flowctl\n", saw_flowctl);
if (saw_flowctl == 0) {
return 1;
}
/* Now clear out all of the unexpected messages so we can clean up everything */
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e2_handle));
ret = PTL_OK;
while (ret != PTL_EQ_EMPTY)
ret = PtlEQGet(eq_handle, &ev);
} else {
ptl_process_t target;
ptl_event_t ev;
int ret, count = 0, fails = 0;
int i;
target.rank = 0;
printf("beginning puts \n");
for (i = 0 ; i < iters ; ++i) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
NULL,
0));
usleep(100);
}
while (count < iters) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_EQ_EMPTY == ret) {
continue;
} else if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
count++;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
count++;
fails++;
} else if (ev.ni_fail_type == PTL_EQ_EMPTY) {
continue;
} else if (ev.ni_fail_type == PTL_EQ_DROPPED) {
continue;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
fprintf(stderr, "%d: Saw %d of %d ACKs as fails\n", rank, fails, count);
}
fprintf(stderr,"at final barrier \n");
libtest_barrier();
if (0 == rank) {
CHECK_RETURNVAL(UNLINK(signal_e_handle));
CHECK_RETURNVAL(UNLINK(signal_e2_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
CHECK_RETURNVAL(PtlCTFree(ct_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
} else {
CHECK_RETURNVAL(PtlMDRelease(md_handle));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
}
fprintf(stderr,"final cleanup \n");
CHECK_RETURNVAL(PtlNIFini(ni_handle));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
/*
/!\ Called for each processes /!\
*/
static int
portals4_init_query(bool enable_progress_threads,
bool enable_mpi_threads)
{
int ret;
ptl_md_t md;
ptl_me_t me;
/* Initialize Portals and create a physical, matching interface */
ret = PtlInit();
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlInit failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlNIInit(PTL_IFACE_DEFAULT,
PTL_NI_PHYSICAL | PTL_NI_MATCHING,
PTL_PID_ANY,
NULL,
&mca_coll_portals4_component.ni_limits,
&mca_coll_portals4_component.ni_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlNIInit failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlGetId(mca_coll_portals4_component.ni_h, &mca_coll_portals4_component.id);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlGetid failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* FIX ME: Need to make sure our ID matches with the MTL... */
ret = PtlGetUid(mca_coll_portals4_component.ni_h, &mca_coll_portals4_component.uid);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlGetUid failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlEQAlloc(mca_coll_portals4_component.ni_h,
MCA_COLL_PORTALS4_EQ_SIZE,
&mca_coll_portals4_component.eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlPTAlloc(mca_coll_portals4_component.ni_h,
0,
mca_coll_portals4_component.eq_h,
REQ_COLL_TABLE_ID,
&mca_coll_portals4_component.pt_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
if (mca_coll_portals4_component.pt_idx != REQ_COLL_TABLE_ID) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc return wrong pt_idx: %d\n",
__FILE__, __LINE__,
mca_coll_portals4_component.finish_pt_idx);
return OMPI_ERROR;
}
ret = PtlPTAlloc(mca_coll_portals4_component.ni_h,
0,
mca_coll_portals4_component.eq_h,
REQ_COLL_FINISH_TABLE_ID,
&mca_coll_portals4_component.finish_pt_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
if (mca_coll_portals4_component.finish_pt_idx != REQ_COLL_FINISH_TABLE_ID) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc return wrong pt_idx: %d\n",
__FILE__, __LINE__,
mca_coll_portals4_component.finish_pt_idx);
return OMPI_ERROR;
}
/* Bind MD/MDs across all memory. We prefer (for obvious reasons)
to have a single MD across all of memory */
memset(&md, 0, sizeof(ptl_md_t));
md.start = 0;
md.length = 0;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(mca_coll_portals4_component.ni_h,
&md,
&mca_coll_portals4_component.zero_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(mca_coll_portals4_component.ni_h,
&md,
&mca_coll_portals4_component.data_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
OPAL_OUTPUT_VERBOSE((90, ompi_coll_base_framework.framework_output, "PtlMDBind start=%p length=%x\n", md.start, md.length));
/* setup finish ack ME */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = 0;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.finish_pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&mca_coll_portals4_component.finish_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend of barrier unexpected failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* This ME is used for RTR exchange only */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
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_OVER_DISABLE |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
/* Note : the RTR bit must be set to match this ME,
* this allows to discriminate the RTR from data flow
* (especially for the Barrier operations)
*/
COLL_PORTALS4_SET_BITS(me.match_bits, 0, 0, 1, 0, 0, 0);
me.ignore_bits = ~COLL_PORTALS4_RTR_MASK;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_OVERFLOW_LIST,
NULL,
&mca_coll_portals4_component.unex_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend of barrier unexpected failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* activate progress callback */
ret = opal_progress_register(portals4_progress);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: opal_progress_register failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
static int ptl_init(MPIDI_PG_t *pg_p, int pg_rank, char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
int ret;
ptl_md_t md;
ptl_ni_limits_t desired;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_PTL_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_PTL_INIT);
/* first make sure that our private fields in the vc and req fit into the area provided */
MPIR_Assert(sizeof(MPID_nem_ptl_vc_area) <= MPIDI_NEM_VC_NETMOD_AREA_LEN);
MPIR_Assert(sizeof(MPID_nem_ptl_req_area) <= MPIDI_NEM_REQ_NETMOD_AREA_LEN);
/* Make sure our IOV is the same as portals4's IOV */
MPIR_Assert(sizeof(ptl_iovec_t) == sizeof(MPL_IOV));
MPIR_Assert(((void*)&(((ptl_iovec_t*)0)->iov_base)) == ((void*)&(((MPL_IOV*)0)->MPL_IOV_BUF)));
MPIR_Assert(((void*)&(((ptl_iovec_t*)0)->iov_len)) == ((void*)&(((MPL_IOV*)0)->MPL_IOV_LEN)));
MPIR_Assert(sizeof(((ptl_iovec_t*)0)->iov_len) == sizeof(((MPL_IOV*)0)->MPL_IOV_LEN));
mpi_errno = MPIDI_CH3I_Register_anysource_notification(MPID_nem_ptl_anysource_posted, MPID_nem_ptl_anysource_matched);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIDI_Anysource_improbe_fn = MPID_nem_ptl_anysource_improbe;
/* init portals */
ret = PtlInit();
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlinit", "**ptlinit %s", MPID_nem_ptl_strerror(ret));
/* do an interface pre-init to get the default limits struct */
ret = PtlNIInit(PTL_IFACE_DEFAULT, PTL_NI_MATCHING | PTL_NI_PHYSICAL,
PTL_PID_ANY, NULL, &desired, &MPIDI_nem_ptl_ni);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
/* finalize the interface so we can re-init with our desired maximums */
ret = PtlNIFini(MPIDI_nem_ptl_ni);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlnifini", "**ptlnifini %s", MPID_nem_ptl_strerror(ret));
/* set higher limits if they are determined to be too low */
if (desired.max_unexpected_headers < UNEXPECTED_HDR_COUNT && getenv("PTL_LIM_MAX_UNEXPECTED_HEADERS") == NULL)
desired.max_unexpected_headers = UNEXPECTED_HDR_COUNT;
if (desired.max_list_size < LIST_SIZE && getenv("PTL_LIM_MAX_LIST_SIZE") == NULL)
desired.max_list_size = LIST_SIZE;
if (desired.max_entries < ENTRY_COUNT && getenv("PTL_LIM_MAX_ENTRIES") == NULL)
desired.max_entries = ENTRY_COUNT;
/* do the real init */
ret = PtlNIInit(PTL_IFACE_DEFAULT, PTL_NI_MATCHING | PTL_NI_PHYSICAL,
PTL_PID_ANY, &desired, &MPIDI_nem_ptl_ni_limits, &MPIDI_nem_ptl_ni);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
/* allocate EQs for each portal */
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_get_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_control_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_rpt_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate a separate EQ for origin events. with this, we can implement rate-limit operations
to prevent a locally triggered flow control even */
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_origin_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for matching messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for large messages where receiver does a get */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_get_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_get_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_control_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_control_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_rpt_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_rpt_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_get_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_rpt_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_control_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* create an MD that covers all of memory */
md.start = 0;
md.length = (ptl_size_t)-1;
md.options = 0x0;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &MPIDI_nem_ptl_global_md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
/* currently, rportlas only works with a single NI and EQ */
ret = MPID_nem_ptl_rptl_init(MPIDI_Process.my_pg->size, ORIGIN_EVENTS, get_target_info);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
/* allow rportal to manage the primary portal and retransmit if needed */
ret = MPID_nem_ptl_rptl_ptinit(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_origin_eq, MPIDI_nem_ptl_pt, MPIDI_nem_ptl_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allow rportal to manage the get and control portals, but we
* don't expect retransmission to be needed on the get portal, so
* we pass PTL_PT_ANY as the dummy portal. unfortunately, portals
* does not have an "invalid" PT constant, which would have been
* more appropriate to pass over here. */
ret = MPID_nem_ptl_rptl_ptinit(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_origin_eq, MPIDI_nem_ptl_get_pt, MPIDI_nem_ptl_get_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
ret = MPID_nem_ptl_rptl_ptinit(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_origin_eq, MPIDI_nem_ptl_control_pt, MPIDI_nem_ptl_control_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* create business card */
mpi_errno = get_business_card(pg_rank, bc_val_p, val_max_sz_p);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* init other modules */
mpi_errno = MPID_nem_ptl_poll_init();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_ptl_nm_init();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_PTL_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int
btl_portals4_init_interface(void)
{
mca_btl_portals4_module_t *portals4_btl;
unsigned int ret, interface;
ptl_md_t md;
ptl_me_t me;
// The initialisation of EQ, PT and ME must be done after the SetMap !
for (interface=0; interface<mca_btl_portals4_component.num_btls; interface++) {
portals4_btl = mca_btl_portals4_component.btls[interface];
/* create event queue */
ret = PtlEQAlloc(portals4_btl->portals_ni_h,
mca_btl_portals4_component.recv_queue_size,
&portals4_btl->recv_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
mca_btl_portals4_component.eqs_h[interface] = portals4_btl->recv_eq_h;
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlEQAlloc (recv_eq=%d) OK for NI %d\n", portals4_btl->recv_eq_h, interface));
/* Create recv_idx portal table entry */
ret = PtlPTAlloc(portals4_btl->portals_ni_h,
PTL_PT_ONLY_TRUNCATE,
portals4_btl->recv_eq_h,
REQ_BTL_TABLE_ID,
&portals4_btl->recv_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlPTAlloc (recv_idx) OK for NI %d recv_idx=%d", interface, portals4_btl->recv_idx));
if (portals4_btl->recv_idx != REQ_BTL_TABLE_ID) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlPTAlloc did not allocate the requested PT: %d",
__FILE__, __LINE__, portals4_btl->recv_idx);
goto error;
}
/* bind zero-length md for sending acks */
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(portals4_btl->portals_ni_h,
&md,
&portals4_btl->zero_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMDBind failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlMDBind (zero-length md=%d) OK for NI %d", portals4_btl->zero_md_h, interface));
/* Bind MD across all memory */
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = 0;
md.eq_handle = portals4_btl->recv_eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(portals4_btl->portals_ni_h,
&md,
&portals4_btl->send_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMDBind failed for NI %d: %d\n",
__FILE__, __LINE__, interface, ret);
goto error;
}
/* Handle long overflows */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE;
if (mca_btl_portals4_component.use_logical) {
me.match_id.rank = PTL_RANK_ANY;
} else {
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
}
me.match_bits = BTL_PORTALS4_LONG_MSG;
me.ignore_bits = BTL_PORTALS4_CONTEXT_MASK |
BTL_PORTALS4_SOURCE_MASK |
BTL_PORTALS4_TAG_MASK;
ret = PtlMEAppend(portals4_btl->portals_ni_h,
portals4_btl->recv_idx,
&me,
PTL_OVERFLOW_LIST,
NULL,
&portals4_btl->long_overflow_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output, "PtlMEAppend (overflow list) OK for NI %d", interface));
}
ret = mca_btl_portals4_recv_enable(portals4_btl);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: Initialization of recv buffer failed: %d",
__FILE__, __LINE__, ret);
goto error;
}
return OPAL_SUCCESS;
error:
opal_output_verbose(1, opal_btl_base_framework.framework_output, "Error in btl_portals4_init_interface");
for (interface=0; interface<mca_btl_portals4_component.num_btls; interface++) {
portals4_btl = mca_btl_portals4_component.btls[interface];
if (NULL != portals4_btl) mca_btl_portals4_free_module(portals4_btl);
}
mca_btl_portals4_component.num_btls = 0;
if (NULL != mca_btl_portals4_component.btls) free(mca_btl_portals4_component.btls);
if (NULL != mca_btl_portals4_component.eqs_h) free(mca_btl_portals4_component.eqs_h);
mca_btl_portals4_component.btls = NULL;
mca_btl_portals4_component.eqs_h = NULL;
return OPAL_ERROR;
}
int
shmem_transport_startup(void)
{
int ret, i;
ptl_process_t *desired = NULL;
ptl_md_t md;
ptl_le_t le;
ptl_uid_t uid = PTL_UID_ANY;
ptl_process_t my_id;
#ifdef USE_ON_NODE_COMMS
int num_on_node = 0;
#endif
#ifdef ENABLE_REMOTE_VIRTUAL_ADDRESSING
/* Make sure the heap and data bases are actually symmetric */
{
int peer;
uint64_t bases[2];
peer = (shmem_internal_my_pe + 1) % shmem_internal_num_pes;
ret = shmem_runtime_get(peer, "portals4-bases", bases, sizeof(uint64_t) * 2);
if (0 != ret) {
fprintf(stderr, "[%03d] ERROR: runtime_put failed: %d\n",
shmem_internal_my_pe, ret);
return ret;
}
if ((uintptr_t) shmem_internal_heap_base != bases[0]) {
fprintf(stderr, "[%03d] ERROR: heap base address does not match with rank %03d and virtual addressing is enabled\n",
shmem_internal_my_pe, peer);
return -1;
}
if ((uintptr_t) shmem_internal_data_base != bases[1]) {
fprintf(stderr, "[%03d] ERROR: data base address does not match with rank %03d and virtual addressing is enabled\n",
shmem_internal_my_pe, peer);
return -1;
}
}
#endif
desired = malloc(sizeof(ptl_process_t) * shmem_internal_num_pes);
if (NULL == desired) {
ret = 1;
goto cleanup;
}
ret = PtlGetPhysId(shmem_transport_portals4_ni_h, &my_id);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlGetPhysId failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
for (i = 0 ; i < shmem_internal_num_pes; ++i) {
ret = shmem_runtime_get(i, "portals4-procid",
&desired[i], sizeof(ptl_process_t));
if (0 != ret) {
fprintf(stderr, "[%03d] ERROR: runtime_get failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#ifdef USE_ON_NODE_COMMS
/* update the connectivity map... */
if (desired[i].phys.nid == my_id.phys.nid) {
SHMEM_SET_RANK_SAME_NODE(i, num_on_node++);
if (num_on_node > 255) {
fprintf(stderr, "[%03d] ERROR: Too many local ranks.\n",
shmem_internal_my_pe);
goto cleanup;
}
}
#endif
}
ret = PtlSetMap(shmem_transport_portals4_ni_h,
shmem_internal_num_pes,
desired);
if (PTL_OK != ret && PTL_IGNORED != ret) {
fprintf(stderr, "[%03d] ERROR: PtlSetMap failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = PtlGetUid(shmem_transport_portals4_ni_h, &uid);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlGetUid failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
shmem_transport_portals4_max_volatile_size = ni_limits.max_volatile_size;
shmem_transport_portals4_max_atomic_size = ni_limits.max_atomic_size;
shmem_transport_portals4_max_fetch_atomic_size = ni_limits.max_fetch_atomic_size;
shmem_transport_portals4_max_msg_size = ni_limits.max_msg_size;
if (shmem_transport_portals4_max_volatile_size < sizeof(long double complex)) {
fprintf(stderr, "[%03d] ERROR: Max volatile size found to be %lu, too small to continue\n",
shmem_internal_my_pe, (unsigned long) shmem_transport_portals4_max_volatile_size);
goto cleanup;
}
if (shmem_transport_portals4_max_atomic_size < sizeof(long double complex)) {
fprintf(stderr, "[%03d] ERROR: Max atomic size found to be %lu, too small to continue\n",
shmem_internal_my_pe, (unsigned long) shmem_transport_portals4_max_atomic_size);
goto cleanup;
}
if (shmem_transport_portals4_max_fetch_atomic_size < sizeof(long double complex)) {
fprintf(stderr, "[%03d] ERROR: Max fetch atomic size found to be %lu, too small to continue\n",
shmem_internal_my_pe, (unsigned long) shmem_transport_portals4_max_fetch_atomic_size);
goto cleanup;
}
/* create portal table entries */
ret = PtlEQAlloc(shmem_transport_portals4_ni_h,
shmem_transport_portals4_event_slots,
&shmem_transport_portals4_eq_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlEQAlloc failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#ifdef ENABLE_REMOTE_VIRTUAL_ADDRESSING
ret = PtlPTAlloc(shmem_transport_portals4_ni_h,
0,
shmem_transport_portals4_eq_h,
shmem_transport_portals4_pt,
&all_pt);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlPTAlloc of table entry failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#else
ret = PtlPTAlloc(shmem_transport_portals4_ni_h,
0,
shmem_transport_portals4_eq_h,
shmem_transport_portals4_data_pt,
&data_pt);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlPTAlloc of data table failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = PtlPTAlloc(shmem_transport_portals4_ni_h,
0,
shmem_transport_portals4_eq_h,
shmem_transport_portals4_heap_pt,
&heap_pt);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlPTAlloc of heap table failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#endif
#ifndef ENABLE_HARD_POLLING
/* target ct */
ret = PtlCTAlloc(shmem_transport_portals4_ni_h, &shmem_transport_portals4_target_ct_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlCTAlloc of target ct failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
le.ct_handle = shmem_transport_portals4_target_ct_h;
#endif
le.uid = uid;
le.options = PTL_LE_OP_PUT | PTL_LE_OP_GET |
PTL_LE_EVENT_LINK_DISABLE |
PTL_LE_EVENT_SUCCESS_DISABLE;
#if !defined(ENABLE_HARD_POLLING)
le.options |= PTL_LE_EVENT_CT_COMM;
#endif
#ifdef ENABLE_REMOTE_VIRTUAL_ADDRESSING
le.start = NULL;
le.length = PTL_SIZE_MAX;
ret = PtlLEAppend(shmem_transport_portals4_ni_h,
shmem_transport_portals4_pt,
&le,
PTL_PRIORITY_LIST,
NULL,
&shmem_transport_portals4_le_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlLEAppend of all memory failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#else
/* Open LE to heap section */
le.start = shmem_internal_heap_base;
le.length = shmem_internal_heap_length;
ret = PtlLEAppend(shmem_transport_portals4_ni_h,
shmem_transport_portals4_heap_pt,
&le,
PTL_PRIORITY_LIST,
NULL,
&shmem_transport_portals4_heap_le_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlLEAppend of heap section failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
/* Open LE to data section */
le.start = shmem_internal_data_base;
le.length = shmem_internal_data_length;
ret = PtlLEAppend(shmem_transport_portals4_ni_h,
shmem_transport_portals4_data_pt,
&le,
PTL_PRIORITY_LIST,
NULL,
&shmem_transport_portals4_data_le_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlLEAppend of data section failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#endif
/* Open MD to all memory */
ret = PtlCTAlloc(shmem_transport_portals4_ni_h, &shmem_transport_portals4_put_ct_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlCTAlloc of put ct failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = PtlCTAlloc(shmem_transport_portals4_ni_h, &shmem_transport_portals4_get_ct_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlCTAlloc of get ct failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_ACK;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_put_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_put_event_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of put MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_ACK |
PTL_MD_EVENT_SUCCESS_DISABLE |
PTL_MD_VOLATILE;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_put_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_put_volatile_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of put MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_ACK |
PTL_MD_EVENT_SUCCESS_DISABLE;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_put_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_put_cntr_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of put cntr MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_REPLY |
PTL_MD_EVENT_SUCCESS_DISABLE;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_get_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_get_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of get MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = 0;
cleanup:
if (NULL != desired) free(desired);
return ret;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_handle;
ptl_process_t *procs;
int rank;
ptl_pt_index_t pt_index, signal_pt_index;
HANDLE_T value_e_handle, signal_e_handle;
int num_procs;
ptl_handle_eq_t eq_handle;
ptl_handle_ct_t ct_handle;
ptl_handle_md_t md_handle;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
if (num_procs < 2) {
fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
return 77;
}
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, NULL, NULL, &ni_handle));
procs = libtest_get_mapping(ni_handle);
CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));
if (0 == rank) {
ENTRY_T value_e;
/* create data ME */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * ITERS / 2, &eq_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
&pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = PTL_CT_NONE;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &value_e_handle));
/* create signal ME */
CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 0, PTL_EQ_NONE, 6,
&signal_pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = ct_handle;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS | PTL_LE_EVENT_SUCCESS_DISABLE | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 6, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e_handle));
} else {
ptl_md_t md;
/* 16 extra just in case... */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, ITERS * 2 + 16, &eq_handle));
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = eq_handle;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
}
libtest_barrier();
if (0 == rank) {
ptl_ct_event_t ct;
ptl_event_t ev;
int ret, count = 0, saw_dropped = 0, saw_flowctl = 0;
/* wait for signal counts */
CHECK_RETURNVAL(PtlCTWait(ct_handle, num_procs - 1, &ct));
if (ct.success != num_procs - 1 || ct.failure != 0) {
return 1;
}
/* wait for event entries */
while (count < ITERS * (num_procs - 1)) {
ret = PtlEQWait(eq_handle, &ev);
if (PTL_OK == ret) {
;
} else if (PTL_EQ_DROPPED == ret) {
saw_dropped++;
if (ev.type == PTL_EVENT_PT_DISABLED){
saw_flowctl++;
CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
}
break;
} else {
fprintf(stderr, "0: Unexpected return code from EQWait: %d\n", ret);
return 1;
}
if (ev.type == PTL_EVENT_PT_DISABLED) {
CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
saw_flowctl++;
} else {
count++;
}
}
fprintf(stderr, "0: Saw %d dropped, %d flowctl\n", saw_dropped, saw_flowctl);
if (saw_flowctl == 0) {
return 1;
}
} else {
ptl_process_t target;
ptl_event_t ev;
int ret, count = 0, fails = 0;
int i;
int *fail_seen;
fail_seen = malloc(sizeof(int) * ITERS);
if (NULL == fail_seen) {
fprintf(stderr, "%d: malloc failed\n", rank);
return 1;
}
memset(fail_seen, 0, sizeof(int) * ITERS);
target.rank = 0;
for (i = 0 ; i < ITERS ; ++i) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
(void*)(size_t)i,
0));
usleep(100);
}
while (count < ITERS) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_EQ_EMPTY == ret) {
continue;
} else if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
count++;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
int iter = (size_t) ev.user_ptr;
if (fail_seen[iter]++ > 0) {
fprintf(stderr, "%d: Double report of PT_DISABLED for "
"iteration %d\n", rank, iter);
return 1;
}
count++;
fails++;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
fprintf(stderr, "%d: Saw %d of %d events as fails\n", rank, fails, count);
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_NO_ACK_REQ,
target,
6,
0,
0,
NULL,
0));
/* wait for the send event on the last put */
CHECK_RETURNVAL(PtlEQWait(eq_handle, &ev));
while (fails > 0) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
NULL,
0));
while (1) {
ret = PtlEQWait(eq_handle, &ev);
if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQWait returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
fails--;
break;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
break;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
}
}
libtest_barrier();
if (0 == rank) {
CHECK_RETURNVAL(UNLINK(signal_e_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
CHECK_RETURNVAL(PtlCTFree(ct_handle));
CHECK_RETURNVAL(UNLINK(value_e_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
} else {
CHECK_RETURNVAL(PtlMDRelease(md_handle));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
}
CHECK_RETURNVAL(PtlNIFini(ni_handle));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
void test_prepostME(int cache_size,
int *cache_buf,
ptl_handle_ni_t ni,
int npeers,
int nmsgs,
int nbytes,
int niters)
{
int i, j, k;
double tmp, total = 0;
ptl_handle_md_t send_md_handle;
ptl_md_t send_md;
ptl_process_t dest;
ptl_size_t offset;
ptl_pt_index_t index;
ptl_handle_eq_t recv_eq_handle;
ptl_handle_me_t me_handles[npeers * nmsgs];
ptl_event_t event;
ptl_assert(PtlEQAlloc(ni, nmsgs * npeers + 1,
&send_md.eq_handle), PTL_OK);
send_md.start = send_buf;
send_md.length = SEND_BUF_SIZE;
send_md.options = PTL_MD_UNORDERED;
send_md.ct_handle = PTL_CT_NONE;
ptl_assert(PtlMDBind(ni, &send_md, &send_md_handle), PTL_OK);
ptl_assert(PtlEQAlloc(ni, nmsgs * npeers + 1, &recv_eq_handle), PTL_OK);
ptl_assert(PtlPTAlloc(ni, 0, recv_eq_handle, TestSameDirectionIndex,
&index), PTL_OK);
ptl_assert(TestSameDirectionIndex, index);
tmp = timer();
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
ptl_process_t src;
src.rank = recv_peers[j];
postME(ni, index, recv_buf + (nbytes * (k + j * nmsgs)),
nbytes, src, magic_tag, &me_handles[k + j * nmsgs]);
}
}
total += (timer() - tmp);
for (i = 0; i < niters - 1; ++i) {
cache_invalidate(cache_size, cache_buf);
libtest_Barrier();
tmp = timer();
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
offset = (nbytes * (k + j * nmsgs));
dest.rank = send_peers[npeers - j - 1],
ptl_assert(libtest_Put_offset(send_md_handle, offset, nbytes,
dest, index, magic_tag, offset), PTL_OK);
}
}
/* wait for sends */
for (j = 0; j < npeers * nmsgs; ++j) {
ptl_assert(PtlEQWait(send_md.eq_handle, &event), PTL_OK);
ptl_assert(event.type, PTL_EVENT_SEND);
}
/* wait for receives */
for (j = 0; j < npeers * nmsgs; j++) {
PtlEQWait(recv_eq_handle, &event);
}
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
ptl_process_t src;
src.rank = recv_peers[j];
postME(ni, index, recv_buf + (nbytes * (k + j * nmsgs)),
nbytes, src, magic_tag, &me_handles[k + j * nmsgs]);
}
}
total += (timer() - tmp);
}
libtest_Barrier();
tmp = timer();
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
offset = (nbytes * (k + j * nmsgs));
dest.rank = send_peers[npeers - j - 1],
ptl_assert(libtest_Put_offset(send_md_handle, offset, nbytes, dest,
index, magic_tag, offset), PTL_OK);
}
}
/* wait for sends */
for (j = 0; j < npeers * nmsgs; ++j) {
ptl_assert(PtlEQWait(send_md.eq_handle, &event), PTL_OK);
ptl_assert(event.type, PTL_EVENT_SEND);
}
/* wait for receives */
for (j = 0; j < npeers * nmsgs; j++) {
PtlEQWait(recv_eq_handle, &event);
}
total += (timer() - tmp);
ptl_assert(PtlEQFree(send_md.eq_handle), PTL_OK);
ptl_assert(PtlMDRelease(send_md_handle), PTL_OK);
ptl_assert(PtlEQFree(recv_eq_handle), PTL_OK);
ptl_assert(PtlPTFree(ni, index), PTL_OK);
tmp = libtest_AllreduceDouble(total, PTL_SUM);
display_result("pre-post", (niters * npeers * nmsgs * 2) / (tmp / world_size));
}
int
ompi_mtl_portals4_flowctl_init(void)
{
ptl_me_t me;
int ret;
ompi_mtl_portals4.flowctl.flowctl_active = false;
OBJ_CONSTRUCT(&ompi_mtl_portals4.flowctl.pending_sends, opal_list_t);
OBJ_CONSTRUCT(&ompi_mtl_portals4.flowctl.pending_fl, opal_free_list_t);
opal_free_list_init(&ompi_mtl_portals4.flowctl.pending_fl,
sizeof(ompi_mtl_portals4_pending_request_t),
opal_cache_line_size,
OBJ_CLASS(ompi_mtl_portals4_pending_request_t),
0, 0, 1, -1, 1, NULL, 0, NULL, NULL, NULL);
ompi_mtl_portals4.flowctl.max_send_slots = (ompi_mtl_portals4.send_queue_size - 3) / 3;
ompi_mtl_portals4.flowctl.send_slots = ompi_mtl_portals4.flowctl.max_send_slots;
ompi_mtl_portals4.flowctl.alert_req.type = portals4_req_flowctl;
ompi_mtl_portals4.flowctl.alert_req.event_callback = flowctl_alert_callback;
ompi_mtl_portals4.flowctl.fanout_req.type = portals4_req_flowctl;
ompi_mtl_portals4.flowctl.fanout_req.event_callback = flowctl_fanout_callback;
ompi_mtl_portals4.flowctl.epoch_counter = -1;
ret = PtlPTAlloc(ompi_mtl_portals4.ni_h,
PTL_PT_ONLY_TRUNCATE,
ompi_mtl_portals4.send_eq_h,
REQ_FLOWCTL_TABLE_ID,
&ompi_mtl_portals4.flowctl_idx);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.trigger_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* everyone creates the trigger ME, even if the root may be the
only to use it */
me.start = NULL;
me.length = 0;
me.min_free = 0;
me.uid = ompi_mtl_portals4.uid;
if (ompi_mtl_portals4.use_logical) {
me.match_id.rank = PTL_RANK_ANY;
} else {
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
}
me.ignore_bits = 0;
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.trigger_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_TRIGGER;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&ompi_mtl_portals4.flowctl.trigger_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Alert CT/ME for broadcasting out alert when root receives a
trigger */
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.alert_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.alert_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_ALERT;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
&ompi_mtl_portals4.flowctl.alert_req,
&ompi_mtl_portals4.flowctl.alert_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Fanin CT/ME for receiving fan-in for restart */
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.fanin_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.fanin_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_FANIN;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&ompi_mtl_portals4.flowctl.fanin_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Fan-out CT/ME for sending restart messages after fan-in */
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.fanout_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.fanout_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_FANOUT;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
&ompi_mtl_portals4.flowctl.fanout_req,
&ompi_mtl_portals4.flowctl.fanout_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ompi_mtl_portals4.flowctl.num_children = 0;
gettimeofday(&ompi_mtl_portals4.flowctl.tv, NULL);
ompi_mtl_portals4.flowctl.backoff_count = 0;
ret = OMPI_SUCCESS;
error:
return ret;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_h;
ptl_pt_index_t pt_index;
uint64_t *buf;
ENTRY_T entry;
HANDLE_T entry_h;
ptl_md_t md;
ptl_handle_md_t md_h;
int rank;
int num_procs;
int ret;
ptl_process_t *procs;
ptl_handle_eq_t eq_h;
ptl_event_t ev;
ptl_hdr_data_t rcvd = 0;
ptl_hdr_data_t goal = 0;
ptl_hdr_data_t hdr_data = 1;
ptl_size_t offset = sizeof(uint64_t);
uint32_t distance;
int sends = 0;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
/* This test only succeeds if we have more than one rank */
if (num_procs < 2) return 77;
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, NULL, NULL, &ni_h));
procs = libtest_get_mapping(ni_h);
CHECK_RETURNVAL(PtlSetMap(ni_h, num_procs, procs));
CHECK_RETURNVAL(PtlEQAlloc(ni_h, 1024, &eq_h));
CHECK_RETURNVAL(PtlPTAlloc(ni_h, 0, eq_h, 0, &pt_index));
assert(pt_index == 0);
buf = malloc(sizeof(uint64_t) * num_procs);
assert(NULL != buf);
md.start = buf;
md.length = sizeof(uint64_t) * num_procs;
md.options = PTL_MD_UNORDERED;
md.eq_handle = eq_h;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_h, &md, &md_h));
entry.start = buf;
entry.length = sizeof(uint64_t) * num_procs;
entry.ct_handle = PTL_CT_NONE;
entry.uid = PTL_UID_ANY;
entry.options = OPTIONS;
#if MATCHING == 1
entry.match_id.rank = PTL_RANK_ANY;
entry.match_bits = 0;
entry.ignore_bits = 0;
entry.min_free = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_h, pt_index, &entry,
PTL_PRIORITY_LIST, NULL, &entry_h));
/* ensure ME is linked before the barrier */
CHECK_RETURNVAL(PtlEQWait(eq_h, &ev));
assert( ev.type == PTL_EVENT_LINK );
libtest_barrier();
/* Bruck's Concatenation Algorithm */
memcpy(buf, &rank, sizeof(uint64_t));
for (distance = 1; distance < num_procs; distance *= 2) {
ptl_size_t to_xfer;
int peer;
ptl_process_t proc;
if (rank >= distance) {
peer = rank - distance;
} else {
peer = rank + (num_procs - distance);
}
to_xfer = sizeof(uint64_t) * MIN(distance, num_procs - distance);
proc.rank = peer;
CHECK_RETURNVAL(PtlPut(md_h,
0,
to_xfer,
PTL_NO_ACK_REQ,
proc,
0,
0,
offset,
NULL,
hdr_data));
sends += 1;
/* wait for completion of the proper receive, and keep count
of uncompleted sends. "rcvd" is an accumulator to deal
with out-of-order receives, which are IDed by the
hdr_data */
goal |= hdr_data;
while ((rcvd & goal) != goal) {
ret = PtlEQWait(eq_h, &ev);
switch (ret) {
case PTL_OK:
if (ev.type == PTL_EVENT_SEND) {
sends -= 1;
} else {
rcvd |= ev.hdr_data;
assert(ev.type == PTL_EVENT_PUT);
assert(ev.rlength == ev.mlength);
assert((ev.rlength == to_xfer) || (ev.hdr_data != hdr_data));
}
break;
default:
fprintf(stderr, "PtlEQWait failure: %d\n", ret);
abort();
}
}
hdr_data <<= 1;
offset += to_xfer;
}
/* wait for any SEND_END events not yet seen */
while (sends) {
ret = PtlEQWait(eq_h, &ev);
switch (ret) {
case PTL_OK:
assert( ev.type == PTL_EVENT_SEND );
sends -= 1;
break;
default:
fprintf(stderr, "PtlEQWait failure: %d\n", ret);
abort();
}
}
CHECK_RETURNVAL(UNLINK(entry_h));
CHECK_RETURNVAL(PtlMDRelease(md_h));
free(buf);
libtest_barrier();
/* cleanup */
CHECK_RETURNVAL(PtlPTFree(ni_h, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_h));
CHECK_RETURNVAL(PtlNIFini(ni_h));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
void
Init(ArgStruct *p, int* pargc, char*** pargv)
{
int rc;
ptl_pt_index_t pt_handle;
/* Initialize Portals and get some runtime info */
rc= PtlInit();
LIBTEST_CHECK(rc, "PtlInit");
libtest_init();
_my_rank= libtest_get_rank();
_nprocs= libtest_get_size();
if (_nprocs < 2) {
if (_my_rank == 0) {
fprintf(stderr, "Need at least two processes!\n", _my_rank);
}
exit(-2);
}
/*
** We need an ni to do barriers and allreduces on.
** It needs to be a non-matching ni.
*/
rc= PtlNIInit(PTL_IFACE_DEFAULT, PTL_NI_NO_MATCHING | PTL_NI_LOGICAL, PTL_PID_ANY, NULL, NULL, &ni_logical);
LIBTEST_CHECK(rc, "PtlNIInit");
rc= PtlSetMap(ni_logical, _nprocs, libtest_get_mapping(ni_logical));
LIBTEST_CHECK(rc, "PtlSetMap");
/* Initialize the barrier in the P4support library. */
libtest_BarrierInit(ni_logical, _my_rank, _nprocs);
/* Allocate a Portal Table Index entry for data transmission */
PtlPTAlloc(ni_logical, 0, PTL_EQ_NONE, PTL_XMIT_INDEX, &pt_handle);
/* Allocate a Portal Table Index entry to receive an int */
PtlPTAlloc(ni_logical, 0, PTL_EQ_NONE, PTL_SEND_INT_INDEX, &pt_handle);
/* Allocate a Portal Table Index entry to receive a double */
PtlPTAlloc(ni_logical, 0, PTL_EQ_NONE, PTL_SEND_DOUBLE_INDEX, &pt_handle);
/* Set up the MD to send a single int */
send_int_ct_handle= PTL_INVALID_HANDLE;
libtest_CreateMDCT(ni_logical, &send_int, sizeof(int), &send_int_md_handle, &send_int_ct_handle);
/* Set up the MD to send a single double */
send_double_ct_handle= PTL_INVALID_HANDLE;
libtest_CreateMDCT(ni_logical, &send_double, sizeof(double), &send_double_md_handle, &send_double_ct_handle);
/* Create a persistent LE to receive a single int */
recv_int_ct_handle= PTL_INVALID_HANDLE;
libtest_CreateLECT(ni_logical, PTL_SEND_INT_INDEX, &recv_int, sizeof(int), &recv_int_le_handle, &recv_int_ct_handle);
/* Create a persistent LE to receive a single double */
recv_double_ct_handle= PTL_INVALID_HANDLE;
libtest_CreateLECT(ni_logical, PTL_SEND_DOUBLE_INDEX, &recv_double, sizeof(double), &recv_double_le_handle, &recv_double_ct_handle);
/*
** Initialize the benchmark data ct handles. Once allocated we'll
** reuse them, instead of reallocating them each time in
** AfterAlignmentInit()
*/
send_ct_handle= PTL_INVALID_HANDLE;
recv_ct_handle= PTL_INVALID_HANDLE;
md_handle= PTL_INVALID_HANDLE;
md_size= -1;
md_buf= NULL;
le_handle= PTL_INVALID_HANDLE;
le_size= -1;
le_buf= NULL;
libtest_barrier();
} /* end of Init() */
static int
portals4_init_interface(void)
{
unsigned int ret;
ptl_md_t md;
ptl_me_t me;
/* create event queues */
ret = PtlEQAlloc(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.send_queue_size,
&ompi_mtl_portals4.send_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ret = PtlEQAlloc(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_queue_size,
&ompi_mtl_portals4.recv_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Create send and long message (read) portal table entries */
ret = PtlPTAlloc(ompi_mtl_portals4.ni_h,
PTL_PT_ONLY_USE_ONCE |
PTL_PT_ONLY_TRUNCATE |
PTL_PT_FLOWCTRL,
ompi_mtl_portals4.recv_eq_h,
REQ_RECV_TABLE_ID,
&ompi_mtl_portals4.recv_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
if (ompi_mtl_portals4.recv_idx != REQ_RECV_TABLE_ID) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc did not allocate the requested PT: %d\n",
__FILE__, __LINE__, ompi_mtl_portals4.recv_idx);
goto error;
}
ret = PtlPTAlloc(ompi_mtl_portals4.ni_h,
PTL_PT_ONLY_USE_ONCE |
PTL_PT_ONLY_TRUNCATE,
ompi_mtl_portals4.send_eq_h,
REQ_READ_TABLE_ID,
&ompi_mtl_portals4.read_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
if (ompi_mtl_portals4.read_idx != REQ_READ_TABLE_ID) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc did not allocate the requested PT: %d\n",
__FILE__, __LINE__, ompi_mtl_portals4.read_idx);
goto error;
}
/* bind zero-length md for sending acks */
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&ompi_mtl_portals4.zero_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Bind MD across all memory */
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = 0;
md.eq_handle = ompi_mtl_portals4.send_eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&ompi_mtl_portals4.send_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Handle long overflows */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = ompi_mtl_portals4.uid;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE;
if (ompi_mtl_portals4.use_logical) {
me.match_id.rank = PTL_RANK_ANY;
} else {
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
}
me.match_bits = MTL_PORTALS4_LONG_MSG;
me.ignore_bits = MTL_PORTALS4_CONTEXT_MASK |
MTL_PORTALS4_SOURCE_MASK |
MTL_PORTALS4_TAG_MASK;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_idx,
&me,
PTL_OVERFLOW_LIST,
NULL,
&ompi_mtl_portals4.long_overflow_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* attach short unex recv blocks */
ret = ompi_mtl_portals4_recv_short_init();
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: short receive block initialization failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ompi_mtl_portals4.opcount = 0;
#if OPAL_ENABLE_DEBUG
ompi_mtl_portals4.recv_opcount = 0;
#endif
#if OMPI_MTL_PORTALS4_FLOW_CONTROL
ret = ompi_mtl_portals4_flowctl_init();
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: ompi_mtl_portals4_flowctl_init failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
#endif
return OMPI_SUCCESS;
error:
if (!PtlHandleIsEqual(ompi_mtl_portals4.long_overflow_me_h, PTL_INVALID_HANDLE)) {
PtlMEUnlink(ompi_mtl_portals4.long_overflow_me_h);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.zero_md_h, PTL_INVALID_HANDLE)) {
PtlMDRelease(ompi_mtl_portals4.zero_md_h);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.send_md_h, PTL_INVALID_HANDLE)) {
PtlMDRelease(ompi_mtl_portals4.send_md_h);
}
if (ompi_mtl_portals4.read_idx != (ptl_pt_index_t) ~0UL) {
PtlPTFree(ompi_mtl_portals4.ni_h, ompi_mtl_portals4.read_idx);
}
if (ompi_mtl_portals4.recv_idx != (ptl_pt_index_t) ~0UL) {
PtlPTFree(ompi_mtl_portals4.ni_h, ompi_mtl_portals4.recv_idx);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.send_eq_h, PTL_INVALID_HANDLE)) {
PtlEQFree(ompi_mtl_portals4.send_eq_h);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.recv_eq_h, PTL_INVALID_HANDLE)) {
PtlEQFree(ompi_mtl_portals4.recv_eq_h);
}
return OMPI_ERROR;
}