Skip to content

TL/UCP: congestion avoidant onesided a2a #1096

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Oct 10, 2025
Merged

TL/UCP: congestion avoidant onesided a2a #1096

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
745238b
TL/UCP: move onesided a2a sync to barrier
wfaderhold21 Mar 17, 2025
7889a95
REVIEW: address feedback
wfaderhold21 Sep 30, 2025
8b37ad5
REVIEW: address feedback
wfaderhold21 Sep 30, 2025
1e5d690
REVIEW: address feedback
wfaderhold21 Oct 7, 2025
3865aab
REVIEW: clang-tidy
wfaderhold21 Oct 7, 2025
72e3ffe
REVIEW: address feedback
wfaderhold21 Oct 7, 2025
2e88c46
REVIEW: address feedback
wfaderhold21 Oct 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 0 additions & 47 deletions src/components/tl/ucp/alltoall/alltoall.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,50 +70,3 @@ ucc_status_t ucc_tl_ucp_alltoall_pairwise_init(ucc_base_coll_args_t *coll_args,
out:
return status;
}

ucc_status_t ucc_tl_ucp_alltoall_onesided_init(ucc_base_coll_args_t *coll_args,
ucc_base_team_t *team,
ucc_coll_task_t **task_h)
{
ucc_tl_ucp_team_t *tl_team = ucc_derived_of(team, ucc_tl_ucp_team_t);
ucc_tl_ucp_task_t *task;
ucc_status_t status;

ALLTOALL_TASK_CHECK(coll_args->args, tl_team);

if (!(coll_args->args.mask & UCC_COLL_ARGS_FIELD_GLOBAL_WORK_BUFFER)) {
tl_error(UCC_TL_TEAM_LIB(tl_team),
"global work buffer not provided nor associated with team");
status = UCC_ERR_NOT_SUPPORTED;
goto out;
}
if (coll_args->args.mask & UCC_COLL_ARGS_FIELD_FLAGS) {
if (!(coll_args->args.flags & UCC_COLL_ARGS_FLAG_MEM_MAPPED_BUFFERS)) {
tl_error(UCC_TL_TEAM_LIB(tl_team),
"non memory mapped buffers are not supported");
status = UCC_ERR_NOT_SUPPORTED;
goto out;
}
}
if (!(coll_args->args.mask & UCC_COLL_ARGS_FIELD_MEM_MAP_SRC_MEMH)) {
coll_args->args.src_memh.global_memh = NULL;
}
if (!(coll_args->args.mask & UCC_COLL_ARGS_FIELD_MEM_MAP_DST_MEMH)) {
coll_args->args.dst_memh.global_memh = NULL;
} else {
if (!(coll_args->args.flags & UCC_COLL_ARGS_FLAG_DST_MEMH_GLOBAL)) {
tl_error(UCC_TL_TEAM_LIB(tl_team),
"onesided alltoall requires global memory handles for dst buffers");
status = UCC_ERR_INVALID_PARAM;
goto out;
}
}

task = ucc_tl_ucp_init_task(coll_args, team);
*task_h = &task->super;
task->super.post = ucc_tl_ucp_alltoall_onesided_start;
task->super.progress = ucc_tl_ucp_alltoall_onesided_progress;
status = UCC_OK;
out:
return status;
}
319 changes: 281 additions & 38 deletions src/components/tl/ucp/alltoall/alltoall_onesided.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,55 +11,298 @@
#include "utils/ucc_math.h"
#include "tl_ucp_sendrecv.h"

void ucc_tl_ucp_alltoall_onesided_progress(ucc_coll_task_t *ctask);
#define CONGESTION_THRESHOLD 8

ucc_status_t ucc_tl_ucp_alltoall_onesided_start(ucc_coll_task_t *ctask)
/* Common helper function to check completion and handle polling */
static inline int alltoall_onesided_handle_completion(
ucc_tl_ucp_task_t *task, uint32_t *posted, uint32_t *completed,
uint32_t nreqs, int64_t npolls)
{
ucc_tl_ucp_task_t *task = ucc_derived_of(ctask, ucc_tl_ucp_task_t);
ucc_tl_ucp_team_t *team = TASK_TEAM(task);
ptrdiff_t src = (ptrdiff_t)TASK_ARGS(task).src.info.buffer;
ptrdiff_t dest = (ptrdiff_t)TASK_ARGS(task).dst.info.buffer;
size_t nelems = TASK_ARGS(task).src.info.count;
ucc_rank_t grank = UCC_TL_TEAM_RANK(team);
ucc_rank_t gsize = UCC_TL_TEAM_SIZE(team);
ucc_rank_t start = (grank + 1) % gsize;
long *pSync = TASK_ARGS(task).global_work_buffer;
ucc_mem_map_mem_h src_memh = TASK_ARGS(task).src_memh.local_memh;
ucc_mem_map_mem_h *dst_memh = TASK_ARGS(task).dst_memh.global_memh;
ucc_rank_t peer;

if (TASK_ARGS(task).flags & UCC_COLL_ARGS_FLAG_SRC_MEMH_GLOBAL) {
src_memh = TASK_ARGS(task).src_memh.global_memh[grank];
int64_t polls = 0;

if ((*posted - *completed) >= nreqs) {
while (polls < npolls) {
ucp_worker_progress(TASK_CTX(task)->worker.ucp_worker);
++polls;
if ((*posted - *completed) < nreqs) {
break;
}
}
if (polls >= npolls) {
return 0; /* Return 0 to indicate should return */
}
}
return 1; /* Return 1 to indicate should continue */
}

ucc_tl_ucp_task_reset(task, UCC_INPROGRESS);
/* TODO: change when support for library-based work buffers is complete */
/* Common helper function to wait for all operations to complete */
static inline void alltoall_onesided_wait_completion(ucc_tl_ucp_task_t *task,
int64_t npolls)
{
int64_t polls = 0;

if (!UCC_TL_UCP_TASK_ONESIDED_P2P_COMPLETE(task)) {
while (polls++ < npolls) {
ucp_worker_progress(TASK_CTX(task)->worker.ucp_worker);
if (UCC_TL_UCP_TASK_ONESIDED_P2P_COMPLETE(task)) {
task->super.status = UCC_OK;
return;
}
}
return;
}
task->super.status = UCC_OK;
}

ucc_status_t ucc_tl_ucp_alltoall_onesided_sched_start(ucc_coll_task_t *ctask)
{
return ucc_schedule_start(ctask);
}

ucc_status_t ucc_tl_ucp_alltoall_onesided_sched_finalize(ucc_coll_task_t *ctask)
{
ucc_schedule_t *schedule = ucc_derived_of(ctask, ucc_schedule_t);
ucc_status_t status;

status = ucc_schedule_finalize(ctask);
ucc_tl_ucp_put_schedule(schedule);
return status;
}

ucc_status_t ucc_tl_ucp_alltoall_onesided_finalize(ucc_coll_task_t *coll_task)
{
ucc_status_t status;

status = ucc_tl_ucp_coll_finalize(coll_task);
if (ucc_unlikely(UCC_OK != status)) {
tl_error(UCC_TASK_LIB(coll_task), "failed to finalize collective");
}
return status;
}

void ucc_tl_ucp_alltoall_onesided_get_progress(ucc_coll_task_t *ctask)
{
ucc_tl_ucp_task_t *task = ucc_derived_of(ctask, ucc_tl_ucp_task_t);
ucc_tl_ucp_team_t *team = TASK_TEAM(task);
ptrdiff_t src = (ptrdiff_t)TASK_ARGS(task).src.info.buffer;
ptrdiff_t dest = (ptrdiff_t)TASK_ARGS(task).dst.info.buffer;
ucc_rank_t grank = UCC_TL_TEAM_RANK(team);
ucc_rank_t gsize = UCC_TL_TEAM_SIZE(team);
uint32_t ntokens = task->alltoall_onesided.tokens;
int64_t npolls = task->alltoall_onesided.npolls;
/* To resolve remote virtual addresses, the dst_memh is the one that must
* have the rkey information. For this algorithm, we need to swap the
* src and dst handles to operate correctly */
ucc_mem_map_mem_h src_memh = TASK_ARGS(task).dst_memh.local_memh;
ucc_mem_map_mem_h *dst_memh = TASK_ARGS(task).src_memh.global_memh;
uint32_t *posted = &task->onesided.get_posted;
uint32_t *completed = &task->onesided.get_completed;
ucc_rank_t peer = (grank + *posted + 1) % gsize;
size_t nelems;

nelems = TASK_ARGS(task).src.info.count;
nelems = (nelems / gsize) * ucc_dt_size(TASK_ARGS(task).src.info.datatype);
dest = dest + grank * nelems;
for (peer = start; task->onesided.put_posted < gsize; peer = (peer + 1) % gsize) {
UCPCHECK_GOTO(ucc_tl_ucp_put_nb(
(void *)(src + peer * nelems), (void *)dest, nelems,
peer, src_memh, dst_memh, team, task),
task, out);
UCPCHECK_GOTO(ucc_tl_ucp_atomic_inc(pSync, peer, dst_memh, team),
task, out);

for (; *posted < gsize; peer = (peer + 1) % gsize) {
UCPCHECK_GOTO(ucc_tl_ucp_get_nb(PTR_OFFSET(dest, peer * nelems),
PTR_OFFSET(src, grank * nelems),
nelems, peer, src_memh, dst_memh, team,
task),
task, out);

if (!alltoall_onesided_handle_completion(task, posted, completed,
ntokens, npolls)) {
return;
}
}
return ucc_progress_queue_enqueue(UCC_TL_CORE_CTX(team)->pq, &task->super);

alltoall_onesided_wait_completion(task, npolls);
out:
return task->super.status;
return;
}

void ucc_tl_ucp_alltoall_onesided_progress(ucc_coll_task_t *ctask)
void ucc_tl_ucp_alltoall_onesided_put_progress(ucc_coll_task_t *ctask)
{
ucc_tl_ucp_task_t *task = ucc_derived_of(ctask, ucc_tl_ucp_task_t);
ucc_tl_ucp_team_t *team = TASK_TEAM(task);
ucc_rank_t gsize = UCC_TL_TEAM_SIZE(team);
long * pSync = TASK_ARGS(task).global_work_buffer;
ucc_tl_ucp_task_t *task = ucc_derived_of(ctask, ucc_tl_ucp_task_t);
ucc_tl_ucp_team_t *team = TASK_TEAM(task);
ptrdiff_t src = (ptrdiff_t)TASK_ARGS(task).src.info.buffer;
ptrdiff_t dest = (ptrdiff_t)TASK_ARGS(task).dst.info.buffer;
ucc_rank_t grank = UCC_TL_TEAM_RANK(team);
ucc_rank_t gsize = UCC_TL_TEAM_SIZE(team);
uint32_t ntokens = task->alltoall_onesided.tokens;
int64_t npolls = task->alltoall_onesided.npolls;
ucc_mem_map_mem_h src_memh = TASK_ARGS(task).src_memh.local_memh;
ucc_mem_map_mem_h *dst_memh = TASK_ARGS(task).dst_memh.global_memh;
uint32_t *posted = &task->onesided.put_posted;
uint32_t *completed = &task->onesided.put_completed;
ucc_rank_t peer = (grank + *posted + 1) % gsize;
size_t nelems;

if (ucc_tl_ucp_test_onesided(task, gsize) == UCC_INPROGRESS) {
return;
nelems = TASK_ARGS(task).src.info.count;
nelems = (nelems / gsize) * ucc_dt_size(TASK_ARGS(task).src.info.datatype);

for (; *posted < gsize; peer = (peer + 1) % gsize) {
UCPCHECK_GOTO(
ucc_tl_ucp_put_nb(PTR_OFFSET(src, peer * nelems),
PTR_OFFSET(dest, grank * nelems), nelems,
peer, src_memh, dst_memh, team, task),
task, out);
UCPCHECK_GOTO(ucc_tl_ucp_ep_flush(peer, team, task), task, out);

if (!alltoall_onesided_handle_completion(task, posted, completed,
ntokens, npolls)) {
return;
}
}

pSync[0] = 0;
task->super.status = UCC_OK;
alltoall_onesided_wait_completion(task, npolls);
out:
return;
}

ucc_status_t ucc_tl_ucp_alltoall_onesided_start(ucc_coll_task_t *ctask)
{
ucc_tl_ucp_task_t *task = ucc_derived_of(ctask, ucc_tl_ucp_task_t);
ucc_tl_ucp_team_t *team = TASK_TEAM(task);

ucc_tl_ucp_task_reset(task, UCC_INPROGRESS);
return ucc_progress_queue_enqueue(UCC_TL_CORE_CTX(team)->pq, &task->super);
}

ucc_status_t ucc_tl_ucp_alltoall_onesided_init(ucc_base_coll_args_t *coll_args,
ucc_base_team_t *team,
ucc_coll_task_t **task_h)
{
ucc_schedule_t *schedule = NULL;
ucc_tl_ucp_team_t *tl_team =
ucc_derived_of(team, ucc_tl_ucp_team_t);
ucc_base_coll_args_t barrier_coll_args = {
.team = team->params.team,
.args.coll_type = UCC_COLL_TYPE_BARRIER,
};
size_t perc_bw =
UCC_TL_UCP_TEAM_LIB(tl_team)->cfg.alltoall_onesided_percent_bw;
ucc_tl_ucp_alltoall_onesided_alg_t alg =
UCC_TL_UCP_TEAM_LIB(tl_team)->cfg.alltoall_onesided_alg;
ucc_tl_ucp_schedule_t *tl_schedule = NULL;
ucc_coll_task_t *barrier_task;
ucc_coll_task_t *a2a_task;
ucc_tl_ucp_task_t *task;
ucc_status_t status;
size_t nelems;
double rate;
size_t ratio;
ucp_ep_h ep;
ucp_ep_evaluate_perf_param_t param;
ucp_ep_evaluate_perf_attr_t attr;
int64_t npolls;
ucc_sbgp_t *sbgp;

ALLTOALL_TASK_CHECK(coll_args->args, tl_team);
if (!(coll_args->args.mask & UCC_COLL_ARGS_FIELD_FLAGS) ||
(coll_args->args.mask & UCC_COLL_ARGS_FIELD_FLAGS &&
(!(coll_args->args.flags &
UCC_COLL_ARGS_FLAG_MEM_MAPPED_BUFFERS)))) {
tl_error(UCC_TL_TEAM_LIB(tl_team),
"non memory mapped buffers are not supported");
status = UCC_ERR_NOT_SUPPORTED;
return status;
}

if (!(coll_args->args.mask & UCC_COLL_ARGS_FIELD_MEM_MAP_SRC_MEMH)) {
coll_args->args.src_memh.global_memh = NULL;
} else {
if (!(coll_args->args.flags & UCC_COLL_ARGS_FLAG_SRC_MEMH_GLOBAL)) {
tl_error(UCC_TL_TEAM_LIB(tl_team),
"onesided alltoall requires global memory handles for src buffers");
status = UCC_ERR_INVALID_PARAM;
return status;
}
}

if (!(coll_args->args.mask & UCC_COLL_ARGS_FIELD_MEM_MAP_DST_MEMH)) {
coll_args->args.dst_memh.global_memh = NULL;
} else {
if (!(coll_args->args.flags & UCC_COLL_ARGS_FLAG_DST_MEMH_GLOBAL)) {
tl_error(UCC_TL_TEAM_LIB(tl_team),
"onesided alltoall requires global memory handles for dst buffers");
status = UCC_ERR_INVALID_PARAM;
return status;
}
}
status = ucc_tl_ucp_get_schedule(tl_team, coll_args,
(ucc_tl_ucp_schedule_t **)&tl_schedule);
if (ucc_unlikely(UCC_OK != status)) {
return status;
}
schedule = &tl_schedule->super.super;
ucc_schedule_init(schedule, coll_args, team);
schedule->super.post = ucc_tl_ucp_alltoall_onesided_sched_start;
schedule->super.progress = NULL;
schedule->super.finalize = ucc_tl_ucp_alltoall_onesided_sched_finalize;

sbgp = ucc_topo_get_sbgp(tl_team->topo, UCC_SBGP_NODE);
if (sbgp->status == UCC_SBGP_NOT_EXISTS) {
status = UCC_ERR_NOT_SUPPORTED;
goto out;
}
task = ucc_tl_ucp_init_task(coll_args, team);
task->super.finalize = ucc_tl_ucp_alltoall_onesided_finalize;
a2a_task = &task->super;

status = ucc_tl_ucp_coll_init(&barrier_coll_args, team, &barrier_task);
if (status != UCC_OK) {
goto out;
}
if (perc_bw > 100) {
perc_bw = 100;
} else if (perc_bw == 0) {
perc_bw = 1;
}

nelems = TASK_ARGS(task).src.info.count;
nelems = nelems / UCC_TL_TEAM_SIZE(tl_team);
param.field_mask = UCP_EP_PERF_PARAM_FIELD_MESSAGE_SIZE;
attr.field_mask = UCP_EP_PERF_ATTR_FIELD_ESTIMATED_TIME;
param.message_size = nelems * ucc_dt_size(TASK_ARGS(task).src.info.datatype);;
ucc_tl_ucp_get_ep(
tl_team, (UCC_TL_TEAM_RANK(tl_team) + 1) % UCC_TL_TEAM_SIZE(tl_team),
&ep);
ucp_ep_evaluate_perf(ep, &param, &attr);

rate = (1 / attr.estimated_time) * (double)(perc_bw / 100.0);
ratio = (nelems > 0) ? nelems * sbgp->group_size : 1;
task->alltoall_onesided.tokens = rate / ratio;
if (task->alltoall_onesided.tokens < 1) {
task->alltoall_onesided.tokens = 1;
}
task->super.post = ucc_tl_ucp_alltoall_onesided_start;
npolls = task->n_polls;
if (alg == UCC_TL_UCP_ALLTOALL_ONESIDED_GET ||
(alg == UCC_TL_UCP_ALLTOALL_ONESIDED_AUTO &&
sbgp->group_size >= CONGESTION_THRESHOLD)) {
npolls = nelems * ucc_dt_size(TASK_ARGS(task).src.info.datatype);
if (npolls < task->n_polls) {
npolls = task->n_polls;
}
task->super.progress = ucc_tl_ucp_alltoall_onesided_get_progress;
} else {
task->super.progress = ucc_tl_ucp_alltoall_onesided_put_progress;
}
task->alltoall_onesided.npolls = npolls;

ucc_schedule_add_task(schedule, a2a_task);
ucc_task_subscribe_dep(&schedule->super, a2a_task,
UCC_EVENT_SCHEDULE_STARTED);

ucc_schedule_add_task(schedule, barrier_task);
ucc_task_subscribe_dep(a2a_task, barrier_task,
UCC_EVENT_COMPLETED);
*task_h = &schedule->super;
return status;
out:
if (tl_schedule) {
ucc_tl_ucp_put_schedule(&tl_schedule->super.super);
}
return status;
}
Loading