static ucs_status_t ucp_wireup_add_am_lane(ucp_ep_h ep, unsigned address_count,
const ucp_address_entry_t *address_list,
ucp_wireup_lane_desc_t *lane_descs,
ucp_lane_index_t *num_lanes_p)
{
ucp_wireup_criteria_t criteria;
ucp_rsc_index_t rsc_index;
ucp_lane_index_t lane;
ucs_status_t status;
unsigned addr_index;
double score;
int need_am;
/* Check if we need active messages, for wireup */
if (!(ucp_ep_get_context_features(ep) & UCP_FEATURE_TAG)) {
need_am = 0;
for (lane = 0; lane < *num_lanes_p; ++lane) {
need_am = need_am || ucp_worker_is_tl_p2p(ep->worker,
lane_descs[lane].rsc_index);
}
if (!need_am) {
return UCS_OK;
}
}
/* Select one lane for active messages */
criteria.title = "active messages";
criteria.local_md_flags = 0;
criteria.remote_md_flags = 0;
criteria.remote_iface_flags = UCT_IFACE_FLAG_AM_BCOPY |
UCT_IFACE_FLAG_AM_CB_SYNC;
criteria.local_iface_flags = UCT_IFACE_FLAG_AM_BCOPY;
criteria.calc_score = ucp_wireup_am_score_func;
if (ucs_test_all_flags(ucp_ep_get_context_features(ep), UCP_FEATURE_TAG |
UCP_FEATURE_WAKEUP)) {
criteria.remote_iface_flags |= UCT_IFACE_FLAG_WAKEUP;
}
status = ucp_wireup_select_transport(ep, address_list, address_count, &criteria,
-1, 1, &rsc_index, &addr_index, &score);
if (status != UCS_OK) {
return status;
}
ucp_wireup_add_lane_desc(lane_descs, num_lanes_p, rsc_index, addr_index,
address_list[addr_index].md_index, score,
UCP_WIREUP_LANE_USAGE_AM);
return UCS_OK;
}
static void ucp_wireup_ep_remote_connected(ucp_ep_h ep)
{
ucp_worker_h worker = ep->worker;
ucp_rsc_index_t rsc_index;
ucp_lane_index_t lane;
ucs_trace("ep %p: remote connected", ep);
for (lane = 0; lane < ucp_ep_num_lanes(ep); ++lane) {
rsc_index = ucp_ep_get_rsc_index(ep, lane);
if (ucp_worker_is_tl_p2p(worker, rsc_index)) {
ucp_stub_ep_remote_connected(ep->uct_eps[lane]);
}
}
}
static ucs_status_t ucp_wireup_connect_local(ucp_ep_h ep, const uint8_t *tli,
unsigned address_count,
const ucp_address_entry_t *address_list)
{
ucp_worker_h worker = ep->worker;
const ucp_address_entry_t *address;
ucp_rsc_index_t rsc_index;
ucp_lane_index_t lane, amo_index;
ucs_status_t status;
ucp_md_map_t UCS_V_UNUSED md_map;
ucs_trace("ep %p: connect local transports", ep);
for (lane = 0; lane < ucp_ep_num_lanes(ep); ++lane) {
rsc_index = ucp_ep_get_rsc_index(ep, lane);
if (!ucp_worker_is_tl_p2p(worker, rsc_index)) {
continue;
}
address = &address_list[tli[lane]];
ucs_assert(address->tl_addr_len > 0);
/* Check that if the lane is used for RMA/AMO, destination md index matches */
md_map = ucp_lane_map_get_lane(ucp_ep_config(ep)->key.rma_lane_map, lane);
ucs_assertv((md_map == 0) || (md_map == UCS_BIT(address->md_index)),
"lane=%d ai=%d md_map=0x%x md_index=%d", lane, tli[lane],
md_map, address->md_index);
amo_index = ucp_ep_get_amo_lane_index(&ucp_ep_config(ep)->key, lane);
if (amo_index != UCP_NULL_LANE) {
md_map = ucp_lane_map_get_lane(ucp_ep_config(ep)->key.amo_lane_map,
amo_index);
ucs_assertv((md_map == 0) || (md_map == UCS_BIT(address->md_index)),
"lane=%d ai=%d md_map=0x%x md_index=%d", lane, tli[lane],
md_map, address->md_index);
}
status = uct_ep_connect_to_ep(ep->uct_eps[lane], address->dev_addr,
address->ep_addr);
if (status != UCS_OK) {
return status;
}
}
return UCS_OK;
}
static ucp_lane_index_t
ucp_wireup_select_wireup_msg_lane(ucp_worker_h worker,
const ucp_address_entry_t *address_list,
const ucp_wireup_lane_desc_t *lane_descs,
ucp_lane_index_t num_lanes)
{
ucp_context_h context = worker->context;
ucp_lane_index_t p2p_lane = UCP_NULL_RESOURCE;
uct_tl_resource_desc_t *resource;
ucp_rsc_index_t rsc_index;
ucp_lane_index_t lane;
unsigned addr_index;
for (lane = 0; lane < num_lanes; ++lane) {
rsc_index = lane_descs[lane].rsc_index;
addr_index = lane_descs[lane].addr_index;
resource = &context->tl_rscs[rsc_index].tl_rsc;
/* if the current lane satisfies the wireup criteria, choose it for wireup.
* if it doesn't take a lane with a p2p transport */
if (ucp_wireup_check_flags(resource,
worker->iface_attrs[rsc_index].cap.flags,
ucp_wireup_aux_criteria.local_iface_flags,
ucp_wireup_aux_criteria.title,
ucp_wireup_iface_flags, NULL, 0) &&
ucp_wireup_check_flags(resource,
address_list[addr_index].iface_attr.cap_flags,
ucp_wireup_aux_criteria.remote_iface_flags,
ucp_wireup_aux_criteria.title,
ucp_wireup_iface_flags, NULL, 0))
{
return lane;
} else if (ucp_worker_is_tl_p2p(worker, rsc_index)) {
p2p_lane = lane;
}
}
return p2p_lane;
}
ucs_status_t ucp_wireup_send_request(ucp_ep_h ep)
{
ucp_worker_h worker = ep->worker;
ucp_rsc_index_t rsc_tli[UCP_MAX_LANES];
ucp_rsc_index_t rsc_index;
uint64_t tl_bitmap = 0;
ucp_lane_index_t lane;
ucs_status_t status;
if (ep->flags & UCP_EP_FLAG_CONNECT_REQ_SENT) {
return UCS_OK;
}
ucs_assert_always(!ucp_ep_is_stub(ep));
for (lane = 0; lane < UCP_MAX_LANES; ++lane) {
if (lane < ucp_ep_num_lanes(ep)) {
rsc_index = ucp_ep_get_rsc_index(ep, lane);
rsc_tli[lane] = ucp_worker_is_tl_p2p(worker, rsc_index) ? rsc_index :
UCP_NULL_RESOURCE;
tl_bitmap |= UCS_BIT(rsc_index);
} else {
rsc_tli[lane] = UCP_NULL_RESOURCE;
}
}
/* TODO make sure such lane would exist */
rsc_index = ucp_stub_ep_get_aux_rsc_index(
ep->uct_eps[ucp_ep_get_wireup_msg_lane(ep)]);
if (rsc_index != UCP_NULL_RESOURCE) {
tl_bitmap |= UCS_BIT(rsc_index);
}
ucs_debug("ep %p: send wireup request (flags=0x%x)", ep, ep->flags);
status = ucp_wireup_msg_send(ep, UCP_WIREUP_MSG_REQUEST, tl_bitmap, rsc_tli);
ep->flags |= UCP_EP_FLAG_CONNECT_REQ_SENT;
return status;
}
ucs_status_t ucp_wireup_init_lanes(ucp_ep_h ep, unsigned address_count,
const ucp_address_entry_t *address_list,
uint8_t *addr_indices)
{
ucp_worker_h worker = ep->worker;
ucp_ep_config_key_t key;
uint16_t new_cfg_index;
ucp_lane_index_t lane;
ucs_status_t status;
uint8_t conn_flag;
char str[32];
ucs_trace("ep %p: initialize lanes", ep);
status = ucp_wireup_select_lanes(ep, address_count, address_list,
addr_indices, &key);
if (status != UCS_OK) {
goto err;
}
key.reachable_md_map |= ucp_ep_config(ep)->key.reachable_md_map;
new_cfg_index = ucp_worker_get_ep_config(worker, &key);
if ((ep->cfg_index == new_cfg_index)) {
return UCS_OK; /* No change */
}
if ((ep->cfg_index != 0) && !ucp_ep_is_stub(ep)) {
/*
* TODO handle a case where we have to change lanes and reconfigure the ep:
*
* - if we already have uct ep connected to an address - move it to the new lane index
* - if we don't yet have connection to an address - create it
* - if an existing lane is not connected anymore - delete it (possibly)
* - if the configuration has changed - replay all pending operations on all lanes -
* need that every pending callback would return, in case of failure, the number
* of lane it wants to be queued on.
*/
ucs_debug("cannot reconfigure ep %p from [%d] to [%d]", ep, ep->cfg_index,
new_cfg_index);
ucp_wireup_print_config(worker->context, &ucp_ep_config(ep)->key, "old", NULL);
ucp_wireup_print_config(worker->context, &key, "new", NULL);
ucs_fatal("endpoint reconfiguration not supported yet");
}
ep->cfg_index = new_cfg_index;
ep->am_lane = key.am_lane;
snprintf(str, sizeof(str), "ep %p", ep);
ucp_wireup_print_config(worker->context, &ucp_ep_config(ep)->key, str,
addr_indices);
ucs_trace("ep %p: connect lanes", ep);
/* establish connections on all underlying endpoints */
conn_flag = UCP_EP_FLAG_LOCAL_CONNECTED;
for (lane = 0; lane < ucp_ep_num_lanes(ep); ++lane) {
status = ucp_wireup_connect_lane(ep, lane, address_count, address_list,
addr_indices[lane]);
if (status != UCS_OK) {
goto err;
}
if (ucp_worker_is_tl_p2p(worker, ucp_ep_get_rsc_index(ep, lane))) {
conn_flag = 0; /* If we have a p2p transport, we're not connected */
}
}
ep->flags |= conn_flag;
return UCS_OK;
err:
for (lane = 0; lane < ucp_ep_num_lanes(ep); ++lane) {
if (ep->uct_eps[lane] != NULL) {
uct_ep_destroy(ep->uct_eps[lane]);
ep->uct_eps[lane] = NULL;
}
}
return status;
}
static void ucp_wireup_process_request(ucp_worker_h worker, const ucp_wireup_msg_t *msg,
uint64_t uuid, const char *peer_name,
unsigned address_count,
const ucp_address_entry_t *address_list)
{
ucp_ep_h ep = ucp_worker_ep_find(worker, uuid);
ucp_rsc_index_t rsc_tli[UCP_MAX_LANES];
uint8_t addr_indices[UCP_MAX_LANES];
ucp_lane_index_t lane, remote_lane;
ucp_rsc_index_t rsc_index;
ucs_status_t status;
uint64_t tl_bitmap = 0;
ucs_trace("ep %p: got wireup request from %s", ep, peer_name);
/* Create a new endpoint if does not exist */
if (ep == NULL) {
status = ucp_ep_new(worker, uuid, peer_name, "remote-request", &ep);
if (status != UCS_OK) {
return;
}
}
/* Initialize lanes (possible destroy existing lanes) */
status = ucp_wireup_init_lanes(ep, address_count, address_list, addr_indices);
if (status != UCS_OK) {
return;
}
/* Connect p2p addresses to remote endpoint */
if (!(ep->flags & UCP_EP_FLAG_LOCAL_CONNECTED)) {
status = ucp_wireup_connect_local(ep, addr_indices, address_count,
address_list);
if (status != UCS_OK) {
return;
}
ep->flags |= UCP_EP_FLAG_LOCAL_CONNECTED;
/* Construct the list that tells the remote side with which address we
* have connected to each of its lanes.
*/
memset(rsc_tli, -1, sizeof(rsc_tli));
for (lane = 0; lane < ucp_ep_num_lanes(ep); ++lane) {
rsc_index = ucp_ep_get_rsc_index(ep, lane);
for (remote_lane = 0; remote_lane < UCP_MAX_LANES; ++remote_lane) {
/* If 'lane' has connected to 'remote_lane' ... */
if (addr_indices[lane] == msg->tli[remote_lane]) {
ucs_assert(ucp_worker_is_tl_p2p(worker, rsc_index));
rsc_tli[remote_lane] = rsc_index;
tl_bitmap |= UCS_BIT(rsc_index);
}
}
}
ucs_trace("ep %p: sending wireup reply", ep);
status = ucp_wireup_msg_send(ep, UCP_WIREUP_MSG_REPLY, tl_bitmap, rsc_tli);
if (status != UCS_OK) {
return;
}
}
}
void ucp_ep_config_init(ucp_worker_h worker, ucp_ep_config_t *config)
{
ucp_context_h context = worker->context;
ucp_ep_rma_config_t *rma_config;
uct_iface_attr_t *iface_attr;
ucp_rsc_index_t rsc_index;
uct_md_attr_t *md_attr;
ucp_lane_index_t lane;
double zcopy_thresh, numerator, denumerator;
size_t rndv_thresh;
/* Default settings */
config->zcopy_thresh = SIZE_MAX;
config->sync_zcopy_thresh = -1;
config->bcopy_thresh = context->config.ext.bcopy_thresh;
config->rndv_thresh = SIZE_MAX;
config->sync_rndv_thresh = SIZE_MAX;
config->max_rndv_get_zcopy = SIZE_MAX;
config->p2p_lanes = 0;
/* Collect p2p lanes */
for (lane = 0; lane < config->key.num_lanes; ++lane) {
rsc_index = config->key.lanes[lane];
if ((rsc_index != UCP_NULL_RESOURCE) &&
ucp_worker_is_tl_p2p(worker, rsc_index))
{
config->p2p_lanes |= UCS_BIT(lane);
}
}
/* Configuration for active messages */
if (config->key.am_lane != UCP_NULL_LANE) {
lane = config->key.am_lane;
rsc_index = config->key.lanes[lane];
if (rsc_index != UCP_NULL_RESOURCE) {
iface_attr = &worker->iface_attrs[rsc_index];
md_attr = &context->md_attrs[context->tl_rscs[rsc_index].md_index];
if (iface_attr->cap.flags & UCT_IFACE_FLAG_AM_SHORT) {
config->max_eager_short = iface_attr->cap.am.max_short -
sizeof(ucp_eager_hdr_t);
config->max_am_short = iface_attr->cap.am.max_short -
sizeof(uint64_t);
}
if (iface_attr->cap.flags & UCT_IFACE_FLAG_AM_BCOPY) {
config->max_am_bcopy = iface_attr->cap.am.max_bcopy;
}
if ((iface_attr->cap.flags & UCT_IFACE_FLAG_AM_ZCOPY) &&
(md_attr->cap.flags & UCT_MD_FLAG_REG))
{
config->max_am_zcopy = iface_attr->cap.am.max_zcopy;
if (context->config.ext.zcopy_thresh == UCS_CONFIG_MEMUNITS_AUTO) {
/* auto */
zcopy_thresh = md_attr->reg_cost.overhead / (
(1.0 / context->config.ext.bcopy_bw) -
(1.0 / iface_attr->bandwidth) -
md_attr->reg_cost.growth);
if (zcopy_thresh < 0) {
config->zcopy_thresh = SIZE_MAX;
config->sync_zcopy_thresh = -1;
} else {
config->zcopy_thresh = zcopy_thresh;
config->sync_zcopy_thresh = zcopy_thresh;
}
} else {
config->zcopy_thresh = context->config.ext.zcopy_thresh;
config->sync_zcopy_thresh = context->config.ext.zcopy_thresh;
}
}
} else {
config->max_am_bcopy = UCP_MIN_BCOPY; /* Stub endpoint */
}
}
/* Configuration for remote memory access */
for (lane = 0; lane < config->key.num_lanes; ++lane) {
if (ucp_lane_map_get_lane(config->key.rma_lane_map, lane) == 0) {
continue;
}
rma_config = &config->rma[lane];
rsc_index = config->key.lanes[lane];
iface_attr = &worker->iface_attrs[rsc_index];
if (rsc_index != UCP_NULL_RESOURCE) {
if (iface_attr->cap.flags & UCT_IFACE_FLAG_PUT_SHORT) {
rma_config->max_put_short = iface_attr->cap.put.max_short;
}
if (iface_attr->cap.flags & UCT_IFACE_FLAG_PUT_BCOPY) {
rma_config->max_put_bcopy = iface_attr->cap.put.max_bcopy;
}
if (iface_attr->cap.flags & UCT_IFACE_FLAG_GET_BCOPY) {
rma_config->max_get_bcopy = iface_attr->cap.get.max_bcopy;
}
} else {
rma_config->max_put_bcopy = UCP_MIN_BCOPY; /* Stub endpoint */
}
}
/* Configuration for Rendezvous data */
if (config->key.rndv_lane != UCP_NULL_LANE) {
lane = config->key.rndv_lane;
rsc_index = config->key.lanes[lane];
if (rsc_index != UCP_NULL_RESOURCE) {
iface_attr = &worker->iface_attrs[rsc_index];
md_attr = &context->md_attrs[context->tl_rscs[rsc_index].md_index];
ucs_assert_always(iface_attr->cap.flags & UCT_IFACE_FLAG_GET_ZCOPY);
if (context->config.ext.rndv_thresh == UCS_CONFIG_MEMUNITS_AUTO) {
/* auto */
/* Make UCX calculate the rndv threshold on its own.
* We do it by finding the message size at which rndv and eager_zcopy get
* the same latency. Starting this message size (rndv_thresh), rndv's latency
* would be lower and the reached bandwidth would be higher.
* The latency function for eager_zcopy is:
* [ reg_cost.overhead + size * md_attr->reg_cost.growth +
* max(size/bw , size/bcopy_bw) + overhead ]
* The latency function for Rendezvous is:
* [ reg_cost.overhead + size * md_attr->reg_cost.growth + latency + overhead +
* reg_cost.overhead + size * md_attr->reg_cost.growth + overhead + latency +
* size/bw + latency + overhead + latency ]
* Isolating the 'size' yields the rndv_thresh.
* The used latency functions for eager_zcopy and rndv are also specified in
* the UCX wiki */
numerator = ((2 * iface_attr->overhead) + (4 * iface_attr->latency) +
md_attr->reg_cost.overhead);
denumerator = (ucs_max((1.0 / iface_attr->bandwidth),(1.0 / context->config.ext.bcopy_bw)) -
md_attr->reg_cost.growth - (1.0 / iface_attr->bandwidth));
if (denumerator > 0) {
rndv_thresh = numerator / denumerator;
ucs_trace("rendezvous threshold is %zu ( = %f / %f)",
rndv_thresh, numerator, denumerator);
} else {
rndv_thresh = context->config.ext.rndv_thresh_fallback;
ucs_trace("rendezvous threshold is %zu", rndv_thresh);
}
/* for the 'auto' mode in the rndv_threshold, we enforce the usage of rndv
* to a value that can be set by the user.
* to disable rndv, need to set a high value for the rndv_threshold
* (without the 'auto' mode)*/
config->rndv_thresh = rndv_thresh;
config->sync_rndv_thresh = rndv_thresh;
config->max_rndv_get_zcopy = iface_attr->cap.get.max_zcopy;
} else {
config->rndv_thresh = context->config.ext.rndv_thresh;
config->sync_rndv_thresh = context->config.ext.rndv_thresh;
config->max_rndv_get_zcopy = iface_attr->cap.get.max_zcopy;
ucs_trace("rendezvous threshold is %zu", config->rndv_thresh);
}
} else {
ucs_debug("rendezvous protocol is not supported ");
}
}
}
