Skip to content

UCP/PROTO: Perf tuning - zcopy fast-completion latency is over-estimated #11016

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

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

UCP/PROTO: Perf tuning - zcopy fast-completion latency is over-estimated #11016

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
eca1d40
UCT/IB/BASE: perf tuning - increase bcopy overhead for grace cpus
roiedanino Nov 19, 2025
c3bc93c
CONFIG: adjusted IB bcopy overhead in ucx.conf
roiedanino Nov 23, 2025
3aa029e
UCP/PROTO: Fix zcopy latency estimation - fast completion is not wait…
roiedanino Nov 24, 2025
ad148bc
UCP/PROTO: added UCP_PROTO_COMMON_INIT_FLAG_RNDV flag, add ack latenc…
roiedanino Nov 27, 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
5 changes: 4 additions & 1 deletion src/ucp/proto/proto_common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,7 +76,10 @@ typedef enum {

/* Supports starting the request when its datatype iterator offset is > 0 */
UCP_PROTO_COMMON_INIT_FLAG_RESUME = UCS_BIT(10),
UCP_PROTO_COMMON_KEEP_MD_MAP = UCS_BIT(11)
UCP_PROTO_COMMON_KEEP_MD_MAP = UCS_BIT(11),

/* Protocol is rendezvous */
UCP_PROTO_COMMON_INIT_FLAG_RNDV = UCS_BIT(12),
} ucp_proto_common_init_flags_t;


Expand Down
8 changes: 4 additions & 4 deletions src/ucp/proto/proto_init.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,10 +169,10 @@ ucp_proto_init_add_tl_perf(const ucp_proto_common_init_params_t *params,
tl_perf->bandwidth;
}

/* Send time is representing request completion, which in case of zcopy
waits for ACK from remote side. */
if ((op_attr_mask & UCP_OP_ATTR_FLAG_FAST_CMPL) &&
(params->flags & UCP_PROTO_COMMON_INIT_FLAG_SEND_ZCOPY)) {
if (ucs_test_all_flags(params->flags, UCP_PROTO_COMMON_INIT_FLAG_SEND_ZCOPY | UCP_PROTO_COMMON_INIT_FLAG_RNDV)
&& (op_attr_mask & UCP_OP_ATTR_FLAG_FAST_CMPL)) {
/* Send time is representing request completion, which in case of zcopy
waits for ACK from remote side. */
perf_factors[UCP_PROTO_PERF_FACTOR_LATENCY].c += tl_perf->latency;
}

Expand Down
2 changes: 2 additions & 0 deletions src/ucp/rndv/rndv_am.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,6 +113,7 @@ static void ucp_rndv_am_bcopy_probe(const ucp_proto_init_params_t *init_params)
.super.memtype_op = UCT_EP_OP_GET_SHORT,
.super.flags = UCP_PROTO_COMMON_INIT_FLAG_CAP_SEG_SIZE |
UCP_PROTO_COMMON_INIT_FLAG_ERR_HANDLING |
UCP_PROTO_COMMON_INIT_FLAG_RNDV |
UCP_PROTO_COMMON_INIT_FLAG_RESUME,
.super.exclude_map = 0,
.super.reg_mem_info = ucp_mem_info_unknown,
Expand Down Expand Up @@ -188,6 +189,7 @@ static void ucp_rndv_am_zcopy_probe(const ucp_proto_init_params_t *init_params)
.super.memtype_op = UCT_EP_OP_LAST,
.super.flags = UCP_PROTO_COMMON_INIT_FLAG_SEND_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_CAP_SEG_SIZE |
UCP_PROTO_COMMON_INIT_FLAG_RNDV |
UCP_PROTO_COMMON_INIT_FLAG_ERR_HANDLING,
.super.exclude_map = 0,
.super.reg_mem_info = ucp_proto_common_select_param_mem_info(
Expand Down
3 changes: 2 additions & 1 deletion src/ucp/rndv/rndv_get.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,7 +42,8 @@ ucp_proto_rndv_get_common_probe(const ucp_proto_init_params_t *init_params,
.super.send_op = UCT_EP_OP_GET_ZCOPY,
.super.memtype_op = memtype_op,
.super.flags = flags | UCP_PROTO_COMMON_INIT_FLAG_RECV_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_REMOTE_ACCESS |
UCP_PROTO_COMMON_INIT_FLAG_REMOTE_ACCESS |
UCP_PROTO_COMMON_INIT_FLAG_RNDV |
UCP_PROTO_COMMON_INIT_FLAG_MIN_FRAG,
.super.exclude_map = 0,
.super.reg_mem_info = *reg_mem_info,
Expand Down
5 changes: 3 additions & 2 deletions src/ucp/rndv/rndv_put.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,8 +258,9 @@ ucp_proto_rndv_put_common_probe(const ucp_proto_init_params_t *init_params,
.super.send_op = UCT_EP_OP_PUT_ZCOPY,
.super.memtype_op = memtype_op,
.super.flags = flags | UCP_PROTO_COMMON_INIT_FLAG_RECV_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_REMOTE_ACCESS |
UCP_PROTO_COMMON_INIT_FLAG_MIN_FRAG,
UCP_PROTO_COMMON_INIT_FLAG_REMOTE_ACCESS |
UCP_PROTO_COMMON_INIT_FLAG_MIN_FRAG |
UCP_PROTO_COMMON_INIT_FLAG_RNDV,
.super.exclude_map = 0,
.super.reg_mem_info = *reg_mem_info,
.max_lanes = context->config.ext.max_rndv_lanes,
Expand Down
8 changes: 5 additions & 3 deletions src/ucp/rndv/rndv_rkey_ptr.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,10 +88,11 @@ ucp_proto_rndv_rkey_ptr_probe(const ucp_proto_init_params_t *init_params)
.super.hdr_size = 0,
.super.send_op = UCT_EP_OP_LAST,
.super.memtype_op = UCT_EP_OP_LAST,
.super.flags = UCP_PROTO_COMMON_INIT_FLAG_RKEY_PTR |
UCP_PROTO_COMMON_INIT_FLAG_RECV_ZCOPY |
.super.flags = UCP_PROTO_COMMON_INIT_FLAG_RKEY_PTR |
UCP_PROTO_COMMON_INIT_FLAG_RECV_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_REMOTE_ACCESS |
UCP_PROTO_COMMON_INIT_FLAG_SINGLE_FRAG,
UCP_PROTO_COMMON_INIT_FLAG_SINGLE_FRAG |
UCP_PROTO_COMMON_INIT_FLAG_RNDV,
.super.exclude_map = 0,
.super.reg_mem_info = ucp_mem_info_unknown,
.lane_type = UCP_LANE_TYPE_RKEY_PTR,
Expand Down Expand Up @@ -269,6 +270,7 @@ ucp_proto_rndv_rkey_ptr_mtype_probe(const ucp_proto_init_params_t *init_params)
.super.memtype_op = UCT_EP_OP_GET_ZCOPY,
.super.flags = UCP_PROTO_COMMON_INIT_FLAG_RKEY_PTR |
UCP_PROTO_COMMON_INIT_FLAG_RECV_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_RNDV |
UCP_PROTO_COMMON_INIT_FLAG_REMOTE_ACCESS,
.super.exclude_map = (rkey_ptr_lane == UCP_NULL_LANE) ?
0 : UCS_BIT(rkey_ptr_lane),
Expand Down
1 change: 1 addition & 0 deletions src/ucp/tag/offload/rndv.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,7 @@ ucp_tag_rndv_offload_proto_probe(const ucp_proto_init_params_t *init_params)
.super.memtype_op = UCT_EP_OP_LAST,
.super.flags = UCP_PROTO_COMMON_INIT_FLAG_SEND_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_RECV_ZCOPY |
UCP_PROTO_COMMON_INIT_FLAG_RNDV |
UCP_PROTO_COMMON_INIT_FLAG_SINGLE_FRAG,
.super.exclude_map = 0,
.super.reg_mem_info = ucp_proto_common_select_param_mem_info(
Expand Down
Loading