MXM_VERNO_MAJOR,
MXM_VERNO_MINOR,
(cur_ver >> MXM_MAJOR_BIT) & 0xff, (cur_ver >> MXM_MINOR_BIT) & 0xff));
}
_mxm_obj.compiletime_version = MXM_VERNO_STRING;
#if MXM_API >= MXM_VERSION(3,0)
_mxm_obj.runtime_version = MPL_strdup(mxm_get_version_string());
#else
_mxm_obj.runtime_version = MPL_malloc(sizeof(MXM_VERNO_STRING) + 10, MPL_MEM_STRINGS);
snprintf(_mxm_obj.runtime_version, (sizeof(MXM_VERNO_STRING) + 9),
"%ld.%ld", (cur_ver >> MXM_MAJOR_BIT) & 0xff, (cur_ver >> MXM_MINOR_BIT) & 0xff);
#endif
_mxm_obj.conf.bulk_connect =
cur_ver < MXM_VERSION(3, 2) ? 0 : MPIR_CVAR_NEMESIS_MXM_BULK_CONNECT;
_mxm_obj.conf.bulk_disconnect =
cur_ver < MXM_VERSION(3, 2) ? 0 : MPIR_CVAR_NEMESIS_MXM_BULK_DISCONNECT;
if (cur_ver < MXM_VERSION(3, 2) &&
(_mxm_obj.conf.bulk_connect || _mxm_obj.conf.bulk_disconnect)) {
_mxm_obj.conf.bulk_connect = 0;
_mxm_obj.conf.bulk_disconnect = 0;
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_CHANNEL, VERBOSE,
(MPL_DBG_FDEST,
"WARNING: MPICH runs with %s version of MXM that is less than 3.2, "
"so bulk connect/disconnect cannot work properly and will be turn off.",
_mxm_obj.runtime_version));
}
fn_exit:
int ompi_mtl_mxm_add_procs(struct mca_mtl_base_module_t *mtl, size_t nprocs,
struct ompi_proc_t** procs, /*const*/
struct mca_mtl_base_endpoint_t **mtl_peer_data)
{
#if MXM_API < MXM_VERSION(2,0)
ompi_mtl_mxm_ep_conn_info_t *ep_info;
mxm_conn_req_t *conn_reqs;
int timeout;
#endif
void *ep_address;
size_t ep_address_len;
mxm_error_t err;
size_t i;
int rc;
assert(mtl == &ompi_mtl_mxm.super);
#if MXM_API < MXM_VERSION(2,0)
/* Allocate connection requests */
conn_reqs = calloc(nprocs, sizeof(mxm_conn_req_t));
ep_info = calloc(nprocs, sizeof(ompi_mtl_mxm_ep_conn_info_t));
if (NULL == conn_reqs || NULL == ep_info) {
rc = OMPI_ERR_OUT_OF_RESOURCE;
goto bail;
}
#endif
/* Get the EP connection requests for all the processes from modex */
for (i = 0; i < nprocs; ++i) {
rc = ompi_mtl_mxm_recv_ep_address(procs[i], &ep_address, &ep_address_len);
if (rc != OMPI_SUCCESS) {
goto bail;
}
#if MXM_API < MXM_VERSION(2,0)
if (ep_address_len != sizeof(ep_info[i])) {
MXM_ERROR("Invalid endpoint address length");
rc = OMPI_ERROR;
goto bail;
}
memcpy(&ep_info[i], ep_address, ep_address_len);
conn_reqs[i].ptl_addr[MXM_PTL_SELF] = (struct sockaddr *)&(ep_info[i].ptl_addr[MXM_PTL_SELF]);
conn_reqs[i].ptl_addr[MXM_PTL_SHM] = (struct sockaddr *)&(ep_info[i].ptl_addr[MXM_PTL_SHM]);
conn_reqs[i].ptl_addr[MXM_PTL_RDMA] = (struct sockaddr *)&(ep_info[i].ptl_addr[MXM_PTL_RDMA]);
#else
mtl_peer_data[i] = (mca_mtl_mxm_endpoint_t *) OBJ_NEW(mca_mtl_mxm_endpoint_t);
mtl_peer_data[i]->mtl_mxm_module = &ompi_mtl_mxm;
err = mxm_ep_connect(ompi_mtl_mxm.ep, ep_address, &mtl_peer_data[i]->mxm_conn);
if (err != MXM_OK) {
MXM_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
rc = OMPI_ERROR;
goto bail;
}
#endif
free(ep_address);
}
#if MXM_API < MXM_VERSION(2,0)
/* Connect to remote peers */
timeout = (mxm_get_version() < MXM_VERSION(1,5)) ? 1000 : -1;
err = mxm_ep_connect(ompi_mtl_mxm.ep, conn_reqs, nprocs, timeout);
if (MXM_OK != err) {
MXM_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
for (i = 0; i < nprocs; ++i) {
if (MXM_OK != conn_reqs[i].error) {
MXM_ERROR("MXM EP connect to %s error: %s\n", procs[i]->proc_hostname,
mxm_error_string(conn_reqs[i].error));
}
}
rc = OMPI_ERROR;
goto bail;
}
/* Save returned connections */
for (i = 0; i < nprocs; ++i) {
mtl_peer_data[i] = (mca_mtl_mxm_endpoint_t *) OBJ_NEW(mca_mtl_mxm_endpoint_t);
mtl_peer_data[i]->mtl_mxm_module = &ompi_mtl_mxm;
mtl_peer_data[i]->mxm_conn = conn_reqs[i].conn;
}
#endif
rc = OMPI_SUCCESS;
bail:
#if MXM_API < MXM_VERSION(2,0)
free(conn_reqs);
free(ep_info);
#endif
return rc;
}
static int ompi_mtl_mxm_component_register(void)
{
mca_base_component_t*c;
c = &mca_mtl_mxm_component.super.mtl_version;
ompi_mtl_mxm.verbose = 0;
(void) mca_base_component_var_register(c, "verbose",
"Verbose level of the MXM component",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_LOCAL,
&ompi_mtl_mxm.verbose);
#if MXM_API > MXM_VERSION(2,0)
ompi_mtl_mxm.mxm_np = 0;
#else
ompi_mtl_mxm.mxm_np = 128;
#endif
(void) mca_base_component_var_register(c, "np",
"[integer] Minimal number of MPI processes in a single job "
"required to activate the MXM transport",
MCA_BASE_VAR_TYPE_INT, NULL,0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mtl_mxm.mxm_np);
param_priority = 30;
(void) mca_base_component_var_register (c,
"priority", "Priority of the MXM MTL component",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
¶m_priority);
#if MXM_API >= MXM_VERSION(3,1)
{
unsigned long cur_ver = mxm_get_version();
ompi_mtl_mxm.bulk_connect = 0;
if (cur_ver < MXM_VERSION(3,2)) {
ompi_mtl_mxm.bulk_disconnect = 0;
} else {
ompi_mtl_mxm.bulk_disconnect = 1;
}
(void) mca_base_component_var_register(c, "bulk_connect",
"[integer] use bulk connect",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mtl_mxm.bulk_connect);
(void) mca_base_component_var_register(c, "bulk_disconnect",
"[integer] use bulk disconnect",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mtl_mxm.bulk_disconnect);
if (cur_ver < MXM_VERSION(3,2) &&
(ompi_mtl_mxm.bulk_connect || ompi_mtl_mxm.bulk_disconnect)) {
ompi_mtl_mxm.bulk_connect = 0;
ompi_mtl_mxm.bulk_disconnect = 0;
MXM_VERBOSE(1, "WARNING: OMPI runs with %s version of MXM that is less than 3.2, "
"so bulk connect/disconnect cannot work properly and will be turn off.",
ompi_mtl_mxm.runtime_version);
}
}
#endif
return OMPI_SUCCESS;
}
int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
{
spml_ikrit_mxm_ep_conn_info_t *ep_info = NULL;
spml_ikrit_mxm_ep_conn_info_t *ep_hw_rdma_info = NULL;
spml_ikrit_mxm_ep_conn_info_t my_ep_info = {{0}};
#if MXM_API < MXM_VERSION(2,0)
mxm_conn_req_t *conn_reqs;
int timeout;
#else
size_t mxm_addr_len = MXM_MAX_ADDR_LEN;
#endif
mxm_error_t err;
size_t i, n;
int rc = OSHMEM_ERROR;
ompi_proc_t *proc_self;
int my_rank = oshmem_my_proc_id();
OBJ_CONSTRUCT(&mca_spml_ikrit.active_peers, opal_list_t);
/* Allocate connection requests */
#if MXM_API < MXM_VERSION(2,0)
conn_reqs = malloc(nprocs * sizeof(mxm_conn_req_t));
if (NULL == conn_reqs) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
memset(conn_reqs, 0x0, sizeof(mxm_conn_req_t));
#endif
ep_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
if (mca_spml_ikrit.hw_rdma_channel) {
ep_hw_rdma_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_hw_rdma_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
}
mca_spml_ikrit.mxm_peers = (mxm_peer_t **) malloc(nprocs
* sizeof(*(mca_spml_ikrit.mxm_peers)));
if (NULL == mca_spml_ikrit.mxm_peers) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
#if MXM_API < MXM_VERSION(2,0)
if (OSHMEM_SUCCESS
!= spml_ikrit_get_ep_address(&my_ep_info, MXM_PTL_SELF)) {
rc = OSHMEM_ERROR;
goto bail;
}
if (OSHMEM_SUCCESS
!= spml_ikrit_get_ep_address(&my_ep_info, MXM_PTL_RDMA)) {
rc = OSHMEM_ERROR;
goto bail;
}
#else
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_get_address(mca_spml_ikrit.mxm_hw_rdma_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_hw_rdma_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
}
err = mxm_ep_get_address(mca_spml_ikrit.mxm_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
#endif
oshmem_shmem_allgather(&my_ep_info, ep_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
opal_progress_register(spml_ikrit_progress);
/* Get the EP connection requests for all the processes from modex */
for (n = 0; n < nprocs; ++n) {
/* mxm 2.0 keeps its connections on a list. Make sure
* that list have different order on every rank */
i = (my_rank + n) % nprocs;
mca_spml_ikrit.mxm_peers[i] = OBJ_NEW(mxm_peer_t);
if (NULL == mca_spml_ikrit.mxm_peers[i]) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
mca_spml_ikrit.mxm_peers[i]->pe = i;
#if MXM_API < MXM_VERSION(2,0)
conn_reqs[i].ptl_addr[MXM_PTL_SELF] =
(struct sockaddr *) &ep_info[i].addr.ptl_addr[MXM_PTL_SELF];
conn_reqs[i].ptl_addr[MXM_PTL_SHM] = NULL;
conn_reqs[i].ptl_addr[MXM_PTL_RDMA] =
(struct sockaddr *) &ep_info[i].addr.ptl_addr[MXM_PTL_RDMA];
#else
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i]->mxm_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
if (OSHMEM_SUCCESS != create_ptl_idx(i))
goto bail;
mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i]->mxm_conn, mca_spml_ikrit.mxm_peers[i]);
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i]->mxm_hw_rdma_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
} else {
mca_spml_ikrit.mxm_peers[i]->mxm_hw_rdma_conn = mca_spml_ikrit.mxm_peers[i]->mxm_conn;
}
#endif
}
#if MXM_API < MXM_VERSION(2,0)
/* Connect to remote peers */
if (mxm_get_version() < MXM_VERSION(1,5)) {
timeout = 1000;
} else {
timeout = -1;
}
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, conn_reqs, nprocs, timeout);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
for (i = 0; i < nprocs; ++i) {
if (MXM_OK != conn_reqs[i].error) {
SPML_ERROR("MXM EP connect to %s error: %s\n",
procs[i]->proc_hostname, mxm_error_string(conn_reqs[i].error));
}
}
rc = OSHMEM_ERR_CONNECTION_FAILED;
goto bail;
}
/* Save returned connections */
for (i = 0; i < nprocs; ++i) {
mca_spml_ikrit.mxm_peers[i]->mxm_conn = conn_reqs[i].conn;
if (OSHMEM_SUCCESS != create_ptl_idx(i)) {
rc = OSHMEM_ERR_CONNECTION_FAILED;
goto bail;
}
mxm_conn_ctx_set(conn_reqs[i].conn, mca_spml_ikrit.mxm_peers[i]);
}
if (conn_reqs)
free(conn_reqs);
#endif
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
#if MXM_API >= MXM_VERSION(2,0)
if (mca_spml_ikrit.bulk_connect) {
/* Need a barrier to ensure remote peers already created connection */
oshmem_shmem_barrier();
mxm_ep_wireup(mca_spml_ikrit.mxm_ep);
}
#endif
proc_self = oshmem_proc_group_find(oshmem_group_all, my_rank);
/* identify local processes and change transport to SHM */
for (i = 0; i < nprocs; i++) {
if (procs[i]->super.proc_name.jobid != proc_self->super.proc_name.jobid ||
!OPAL_PROC_ON_LOCAL_NODE(procs[i]->super.proc_flags)) {
continue;
}
if (procs[i] == proc_self)
continue;
/* use zcopy for put/get via sysv shared memory */
OSHMEM_PROC_DATA(procs[i])->transport_ids[0] = MXM_PTL_SHM;
OSHMEM_PROC_DATA(procs[i])->transport_ids[1] = MXM_PTL_RDMA;
OSHMEM_PROC_DATA(procs[i])->num_transports = 2;
}
SPML_VERBOSE(50, "*** ADDED PROCS ***");
return OSHMEM_SUCCESS;
bail:
#if MXM_API < MXM_VERSION(2,0)
if (conn_reqs)
free(conn_reqs);
#endif
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
SPML_ERROR("add procs FAILED rc=%d", rc);
return rc;
}
"WARNING: MPICH was compiled with MXM version %d.%d but version %ld.%ld detected.",
MXM_VERNO_MAJOR,
MXM_VERNO_MINOR,
(cur_ver >> MXM_MAJOR_BIT) & 0xff, (cur_ver >> MXM_MINOR_BIT) & 0xff));
}
_mxm_obj.compiletime_version = MXM_VERNO_STRING;
#if MXM_API >= MXM_VERSION(3,0)
_mxm_obj.runtime_version = MPIU_Strdup(mxm_get_version_string());
#else
_mxm_obj.runtime_version = MPIU_Malloc(sizeof(MXM_VERNO_STRING) + 10);
snprintf(_mxm_obj.runtime_version, (sizeof(MXM_VERNO_STRING) + 9),
"%ld.%ld", (cur_ver >> MXM_MAJOR_BIT) & 0xff, (cur_ver >> MXM_MINOR_BIT) & 0xff);
#endif
_mxm_obj.conf.bulk_connect = cur_ver < MXM_VERSION(3, 2) ? 0 : MPIR_CVAR_NEMESIS_MXM_BULK_CONNECT;
_mxm_obj.conf.bulk_disconnect = cur_ver < MXM_VERSION(3, 2) ? 0 : MPIR_CVAR_NEMESIS_MXM_BULK_DISCONNECT;
if (cur_ver < MXM_VERSION(3, 2) &&
(_mxm_obj.conf.bulk_connect || _mxm_obj.conf.bulk_disconnect)) {
_mxm_obj.conf.bulk_connect = 0;
_mxm_obj.conf.bulk_disconnect = 0;
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE,
(MPIU_DBG_FDEST,
"WARNING: MPICH runs with %s version of MXM that is less than 3.2, "
"so bulk connect/disconnect cannot work properly and will be turn off.",
_mxm_obj.runtime_version));
}
fn_exit:
return mpi_errno;
