pscom_err_t pscom_connect_socket_str(pscom_connection_t *connection, const char *socket_str)
{
int nodeid;
int portno;
char name[8];
int res;
res = pscom_parse_socket_ondemand_str(socket_str, &nodeid, &portno, &name);
if (res) goto err_parse;
connection->nodeid = nodeid;
connection->portno = portno;
if (!name[0]) {
return pscom_connect(connection, nodeid, portno);
} else {
return pscom_connect_ondemand(connection, nodeid, portno, name);
}
/* error code */
err_parse:
if (socket_str) {
DPRINT(1, "CONNECT (%s) failed : %s", socket_str, pscom_err_str(res));
} else {
DPRINT(1, "CONNECT (<null>) failed : %s", pscom_err_str(res));
}
return res;
}
void conn_error_server(pscom_connection_t *connection,
pscom_op_t operation, pscom_err_t error)
{
printf(RED "Error on connection from %s via %s : %s : %s" NORM "\n",
pscom_con_info_str(&connection->remote_con_info),
pscom_con_type_str(connection->type),
pscom_op_str(operation),
pscom_err_str(error));
}
static
int do_connect(pscom_socket_t *socket, int pg_rank, int dest, char *dest_addr)
{
pscom_connection_t *con;
pscom_err_t rc;
struct InitMsg init_msg;
/* printf("Connecting (rank %d to %d) (%s)\n", pg_rank, dest, dest_addr); */
con = pscom_open_connection(socket);
rc = pscom_connect_socket_str(con, dest_addr);
if (rc != PSCOM_SUCCESS) {
PRINTERROR("Connecting %s to %s (rank %d to %d) failed : %s",
pscom_listen_socket_str(socket),
dest_addr, pg_rank, dest, pscom_err_str(rc));
return -1; /* error */
}
grank2con_set(dest, con);
init_msg.from_rank = pg_rank;
pscom_send(con, NULL, 0, &init_msg, sizeof(init_msg));
return 0;
}
int MPID_Init(int *argc, char ***argv,
int threadlevel_requested, int *threadlevel_provided,
int *has_args, int *has_env)
{
int mpi_errno = MPI_SUCCESS;
int pg_rank, pg_size, pg_id_sz;
int appnum = -1;
/* int universe_size; */
int has_parent;
pscom_socket_t *socket;
pscom_err_t rc;
char *pg_id_name;
char *parent_port;
/* Call any and all MPID_Init type functions */
MPIR_Err_init();
MPIR_Datatype_init();
MPIR_Group_init();
mpid_debug_init();
assert(PSCOM_ANYPORT == -1); /* all codeplaces which depends on it are marked with: "assert(PSP_ANYPORT == -1);" */
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_INIT);
PMICALL(PMI_Init(&has_parent));
PMICALL(PMI_Get_rank(&pg_rank));
PMICALL(PMI_Get_size(&pg_size));
PMICALL(PMI_Get_appnum(&appnum));
*has_args = 1;
*has_env = 1;
/* without PMI_Get_universe_size() we see pmi error:
'[unset]: write_line error; fd=-1' in PMI_KVS_Get()! */
/* PMICALL(PMI_Get_universe_size(&universe_size)); */
if (pg_rank < 0) pg_rank = 0;
if (pg_size <= 0) pg_size = 1;
if (
#ifndef MPICH_IS_THREADED
1
#else
threadlevel_requested < MPI_THREAD_MULTIPLE
#endif
) {
rc = pscom_init(PSCOM_VERSION);
if (rc != PSCOM_SUCCESS) {
fprintf(stderr, "pscom_init(0x%04x) failed : %s\n",
PSCOM_VERSION,
pscom_err_str(rc));
exit(1);
}
} else {
rc = pscom_init_thread(PSCOM_VERSION);
if (rc != PSCOM_SUCCESS) {
fprintf(stderr, "pscom_init_thread(0x%04x) failed : %s\n",
PSCOM_VERSION,
pscom_err_str(rc));
exit(1);
}
}
/* Initialize the switches */
pscom_env_get_uint(&MPIDI_Process.env.enable_collectives, "PSP_COLLECTIVES");
#ifdef PSCOM_HAS_ON_DEMAND_CONNECTIONS
/* if (pg_size > 32) MPIDI_Process.env.enable_ondemand = 1; */
pscom_env_get_uint(&MPIDI_Process.env.enable_ondemand, "PSP_ONDEMAND");
#else
MPIDI_Process.env.enable_ondemand = 0;
#endif
/* enable_ondemand_spawn defaults to enable_ondemand */
MPIDI_Process.env.enable_ondemand_spawn = MPIDI_Process.env.enable_ondemand;
pscom_env_get_uint(&MPIDI_Process.env.enable_ondemand_spawn, "PSP_ONDEMAND_SPAWN");
/* take SMP-related locality information into account (e.g., for MPI_Win_allocate_shared) */
pscom_env_get_uint(&MPIDI_Process.env.enable_smp_awareness, "PSP_SMP_AWARENESS");
/* take MSA-related topology information into account */
pscom_env_get_uint(&MPIDI_Process.env.enable_msa_awareness, "PSP_MSA_AWARENESS");
if(MPIDI_Process.env.enable_msa_awareness) {
pscom_env_get_uint(&MPIDI_Process.msa_module_id, "PSP_MSA_MODULE_ID");
}
#ifdef MPID_PSP_TOPOLOGY_AWARE_COLLOPS
/* use hierarchy-aware collectives on SMP level */
pscom_env_get_uint(&MPIDI_Process.env.enable_smp_aware_collops, "PSP_SMP_AWARE_COLLOPS");
/* use hierarchy-aware collectives on MSA level (disables SMP-aware collops / FIX ME!) */
pscom_env_get_uint(&MPIDI_Process.env.enable_msa_aware_collops, "PSP_MSA_AWARE_COLLOPS");
if(MPIDI_Process.env.enable_msa_aware_collops) MPIDI_Process.env.enable_smp_aware_collops = 0;
#endif
#ifdef MPID_PSP_CREATE_HISTOGRAM
/* collect statistics information and print them at the end of a run */
pscom_env_get_uint(&MPIDI_Process.env.enable_histogram, "PSP_HISTOGRAM");
pscom_env_get_uint(&MPIDI_Process.histo.max_size, "PSP_HISTOGRAM_MAX");
pscom_env_get_uint(&MPIDI_Process.histo.min_size, "PSP_HISTOGRAM_MIN");
pscom_env_get_uint(&MPIDI_Process.histo.step_width, "PSP_HISTOGRAM_SHIFT");
#endif
/*
pscom_env_get_uint(&mpir_allgather_short_msg, "PSP_ALLGATHER_SHORT_MSG");
pscom_env_get_uint(&mpir_allgather_long_msg, "PSP_ALLGATHER_LONG_MSG");
pscom_env_get_uint(&mpir_allreduce_short_msg, "PSP_ALLREDUCE_SHORT_MSG");
pscom_env_get_uint(&mpir_alltoall_short_msg, "PSP_ALLTOALL_SHORT_MSG");
pscom_env_get_uint(&mpir_alltoall_medium_msg, "PSP_ALLTOALL_MEDIUM_MSG");
pscom_env_get_uint(&mpir_alltoall_throttle, "PSP_ALLTOALL_THROTTLE");
pscom_env_get_uint(&mpir_bcast_short_msg, "PSP_BCAST_SHORT_MSG");
pscom_env_get_uint(&mpir_bcast_long_msg, "PSP_BCAST_LONG_MSG");
pscom_env_get_uint(&mpir_bcast_min_procs, "PSP_BCAST_MIN_PROCS");
pscom_env_get_uint(&mpir_gather_short_msg, "PSP_GATHER_SHORT_MSG");
pscom_env_get_uint(&mpir_gather_vsmall_msg, "PSP_GATHER_VSMALL_MSG");
pscom_env_get_uint(&mpir_redscat_commutative_long_msg, "PSP_REDSCAT_COMMUTATIVE_LONG_MSG");
pscom_env_get_uint(&mpir_redscat_noncommutative_short_msg, "PSP_REDSCAT_NONCOMMUTATIVE_SHORT_MSG");
pscom_env_get_uint(&mpir_reduce_short_msg, "PSP_REDUCE_SHORT_MSG");
pscom_env_get_uint(&mpir_scatter_short_msg, "PSP_SCATTER_SHORT_MSG");
*/
socket = pscom_open_socket(0, 0);
if (!MPIDI_Process.env.enable_ondemand) {
socket->ops.con_accept = mpid_con_accept;
}
{
char name[10];
snprintf(name, sizeof(name), "r%07u", (unsigned)pg_rank);
pscom_socket_set_name(socket, name);
}
rc = pscom_listen(socket, PSCOM_ANYPORT);
if (rc != PSCOM_SUCCESS) { PRINTERROR("pscom_listen(PSCOM_ANYPORT)"); goto fn_fail; }
/* Note that if pmi is not availble, the value of MPI_APPNUM is not set */
/* if (appnum != -1) {*/
MPIR_Process.attrs.appnum = appnum;
/* }*/
#if 0
// see mpiimpl.h:
// typedef struct PreDefined_attrs {
// int appnum; /* Application number provided by mpiexec (MPI-2) */
// int host; /* host */
// int io; /* standard io allowed */
// int lastusedcode; /* last used error code (MPI-2) */
// int tag_ub; /* Maximum message tag */
// int universe; /* Universe size from mpiexec (MPI-2) */
// int wtime_is_global; /* Wtime is global over processes in COMM_WORLD */
// } PreDefined_attrs;
#endif
MPIR_Process.attrs.tag_ub = MPIDI_TAG_UB;
/* obtain the id of the process group */
PMICALL(PMI_KVS_Get_name_length_max(&pg_id_sz));
pg_id_name = MPL_malloc(pg_id_sz + 1, MPL_MEM_STRINGS);
if (!pg_id_name) { PRINTERROR("MPL_malloc()"); goto fn_fail; }
PMICALL(PMI_KVS_Get_my_name(pg_id_name, pg_id_sz));
/* safe */
/* MPIDI_Process.socket = socket; */
MPIDI_Process.my_pg_rank = pg_rank;
MPIDI_Process.my_pg_size = pg_size;
MPIDI_Process.pg_id_name = pg_id_name;
if (!MPIDI_Process.env.enable_ondemand) {
/* Create and establish all connections */
if (InitPortConnections(socket) != MPI_SUCCESS) goto fn_fail;
} else {
/* Create all connections as "on demand" connections. */
if (InitPscomConnections(socket) != MPI_SUCCESS) goto fn_fail;
}
#ifdef MPID_PSP_TOPOLOGY_AWARE_COLLOPS
{
int grank;
int my_node_id = -1;
int remote_node_id = -1;
int* node_id_table;
if(MPIDI_Process.env.enable_msa_awareness && MPIDI_Process.env.enable_msa_aware_collops) {
my_node_id = MPIDI_Process.msa_module_id;
assert(my_node_id > -1);
} else if(MPIDI_Process.env.enable_smp_awareness && MPIDI_Process.env.enable_smp_aware_collops) {
if (!MPIDI_Process.env.enable_ondemand) {
/* In the PSP_ONDEMAND=0 case, we can just check the pscom connection types: */
for (grank = 0; grank < pg_size; grank++) {
pscom_connection_t *con = grank2con_get(grank);
if( (con->type == PSCOM_CON_TYPE_SHM) || (pg_rank == grank) ) {
my_node_id = grank;
break;
}
}
} else {
/* In the PSP_ONDEMAND=1 case, we have to use a hash of the host name: */
my_node_id = MPID_PSP_get_host_hash();
if(my_node_id < 0) my_node_id *= -1;
}
assert(my_node_id > -1);
} else {
/* No hierarchy-awareness requested */
assert(my_node_id == -1);
}
if(my_node_id > -1) {
node_id_table = MPL_malloc(pg_size * sizeof(int), MPL_MEM_OBJECT);
if(pg_rank != 0) {
/* gather: */
pscom_connection_t *con = grank2con_get(0);
assert(con);
pscom_send(con, NULL, 0, &my_node_id, sizeof(int));
/* bcast: */
rc = pscom_recv_from(con, NULL, 0, node_id_table, pg_size*sizeof(int));
assert(rc == PSCOM_SUCCESS);
} else {
/* gather: */
node_id_table[0] = my_node_id;
for(grank=1; grank < pg_size; grank++) {
pscom_connection_t *con = grank2con_get(grank);
assert(con);
rc = pscom_recv_from(con, NULL, 0, &remote_node_id, sizeof(int));
assert(rc == PSCOM_SUCCESS);
node_id_table[grank] = remote_node_id;
}
/* bcast: */
for(grank=1; grank < pg_size; grank++) {
pscom_connection_t *con = grank2con_get(grank);
pscom_send(con, NULL, 0, node_id_table, pg_size*sizeof(int));
}
}
MPIDI_Process.node_id_table = node_id_table;
} else {
/* No hierarchy-awareness requested */
assert(MPIDI_Process.node_id_table == NULL);
}
}
#endif
/*
* Initialize the MPI_COMM_WORLD object
*/
{
MPIR_Comm * comm;
int grank;
MPIDI_PG_t * pg_ptr;
int pg_id_num;
MPIDI_VCRT_t * vcrt;
comm = MPIR_Process.comm_world;
comm->rank = pg_rank;
comm->remote_size = pg_size;
comm->local_size = pg_size;
comm->pscom_socket = socket;
vcrt = MPIDI_VCRT_Create(comm->remote_size);
assert(vcrt);
MPID_PSP_comm_set_vcrt(comm, vcrt);
MPIDI_PG_Convert_id(pg_id_name, &pg_id_num);
MPIDI_PG_Create(pg_size, pg_id_num, &pg_ptr);
assert(pg_ptr == MPIDI_Process.my_pg);
for (grank = 0; grank < pg_size; grank++) {
/* MPIR_CheckDisjointLpids() in mpi/comm/intercomm_create.c expect
lpid to be smaller than 4096!!!
Else you will see an "Fatal error in MPI_Intercomm_create"
*/
pscom_connection_t *con = grank2con_get(grank);
pg_ptr->vcr[grank] = MPIDI_VC_Create(pg_ptr, grank, con, grank);
comm->vcr[grank] = MPIDI_VC_Dup(pg_ptr->vcr[grank]);
}
mpi_errno = MPIR_Comm_commit(comm);
assert(mpi_errno == MPI_SUCCESS);
}
/*
* Initialize the MPI_COMM_SELF object
*/
{
MPIR_Comm * comm;
MPIDI_VCRT_t * vcrt;
comm = MPIR_Process.comm_self;
comm->rank = 0;
comm->remote_size = 1;
comm->local_size = 1;
comm->pscom_socket = socket;
vcrt = MPIDI_VCRT_Create(comm->remote_size);
assert(vcrt);
MPID_PSP_comm_set_vcrt(comm, vcrt);
comm->vcr[0] = MPIDI_VC_Dup(MPIR_Process.comm_world->vcr[pg_rank]);
mpi_errno = MPIR_Comm_commit(comm);
assert(mpi_errno == MPI_SUCCESS);
}
/* ToDo: move MPID_enable_receive_dispach to bg thread */
MPID_enable_receive_dispach(socket);
if (threadlevel_provided) {
*threadlevel_provided = (MPICH_THREAD_LEVEL < threadlevel_requested) ?
MPICH_THREAD_LEVEL : threadlevel_requested;
}
if (has_parent) {
MPIR_Comm * comm;
mpi_errno = MPID_PSP_GetParentPort(&parent_port);
assert(mpi_errno == MPI_SUCCESS);
/*
printf("%s:%u:%s Child with Parent: %s\n", __FILE__, __LINE__, __func__, parent_port);
*/
mpi_errno = MPID_Comm_connect(parent_port, NULL, 0,
MPIR_Process.comm_world, &comm);
if (mpi_errno != MPI_SUCCESS) {
fprintf(stderr, "MPI_Comm_connect(parent) failed!\n");
goto fn_fail;
}
assert(comm != NULL);
MPL_strncpy(comm->name, "MPI_COMM_PARENT", MPI_MAX_OBJECT_NAME);
MPIR_Process.comm_parent = comm;
}
MPID_PSP_shm_rma_init();
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_INIT);
return mpi_errno;
/* --- */
fn_fail:
/* A failing MPI_Init() did'nt call the MPI error handler, which
mostly calls abort(). This cause MPI_Init() to return the mpi_errno,
which nobody check, causing segfaultm double frees and so on. To
prevent strange error messages, we now call _exit(1) here.
*/
_exit(1);
}
static
int InitPscomConnections(pscom_socket_t *socket) {
char key[50];
unsigned long guard_pmi_key = MAGIC_PMI_KEY;
int i;
int mpi_errno = MPI_SUCCESS;
int pg_rank = MPIDI_Process.my_pg_rank;
int pg_size = MPIDI_Process.my_pg_size;
char *pg_id = MPIDI_Process.pg_id_name;
char *listen_socket;
char **psp_port = NULL;
/* Distribute my contact information */
snprintf(key, sizeof(key), "pscom%d", pg_rank);
listen_socket = MPL_strdup(pscom_listen_socket_ondemand_str(socket));
PMICALL(PMI_KVS_Put(pg_id, key, listen_socket));
#define IPSCOM_VERSION "pscom_v5.0"
i_version_set(pg_id, pg_rank, IPSCOM_VERSION);
PMICALL(PMI_KVS_Commit(pg_id));
PMICALL(PMI_Barrier());
i_version_check(pg_id, pg_rank, IPSCOM_VERSION);
init_grank_port_mapping();
/* Get portlist */
psp_port = MPL_malloc(pg_size * sizeof(*psp_port), MPL_MEM_OBJECT);
assert(psp_port);
for (i = 0; i < pg_size; i++) {
char val[100];
unsigned long guard_pmi_value = MAGIC_PMI_VALUE;
if (i != pg_rank) {
snprintf(key, sizeof(key), "pscom%d", i);
checked_PMI_KVS_Get(pg_id, key, val, sizeof(val));
/* simple_pmi.c has a bug.(fixed in mpich2-1.0.5)
Test for the bugfix: */
assert(guard_pmi_value == MAGIC_PMI_VALUE);
assert(guard_pmi_key == MAGIC_PMI_KEY);
} else {
/* myself: Dont use PMI_KVS_Get, because this fail
in the case of no pm (SINGLETON_INIT_BUT_NO_PM) */
strcpy(val, listen_socket);
}
psp_port[i] = MPL_strdup(val);
}
/* Create all connections */
for (i = 0; i < pg_size; i++) {
pscom_connection_t *con;
pscom_err_t rc;
const char *dest;
dest = psp_port[i];
con = pscom_open_connection(socket);
rc = pscom_connect_socket_str(con, dest);
if (rc != PSCOM_SUCCESS) {
PRINTERROR("Connecting %s to %s (rank %d to %d) failed : %s",
listen_socket, dest, pg_rank, i, pscom_err_str(rc));
goto fn_fail;
}
grank2con_set(i, con);
}
pscom_stop_listen(socket);
fn_exit:
if (psp_port) {
for (i = 0; i < pg_size; i++) {
MPL_free(psp_port[i]);
psp_port[i] = NULL;
}
MPL_free(psp_port);
}
MPL_free(listen_socket);
return mpi_errno;
/* --- */
fn_fail:
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
"InitPscomConnections", __LINE__, MPI_ERR_OTHER, "**connfailed", 0);
goto fn_exit;
}
void abort_on_error(const char *msg, pscom_err_t error)
{
if (!error) return;
printf(RED "%s : %s" NORM "\n", msg, pscom_err_str(error));
exit(1);
}