Skip to content

Commit

Permalink
fix : Extend PAT-SIMPLE to small message sizes
Browse files Browse the repository at this point in the history 
Extend the region of PAT-SIMPLE to be used on smaller message size,
since that gives better performance on P5en.

Signed-off-by: Arun Karthik <akkart@amazon.com>
  • Loading branch information
@arunkarthik-akkart @rajachan
arunkarthik-akkart authored and rajachan committed Dec 10, 2024
1 parent 50c9e8d commit 75b31d0
Showing 1 changed file with 12 additions and 24 deletions.
36 changes: 12 additions & 24 deletions src/tuner/nccl_ofi_regions.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -495,10 +495,6 @@ static ncclResult_t region_init_internal_p5en(nccl_ofi_tuner_region_context_t *r
}
{
collType = ncclFuncAllGather;
nccl_ofi_tuner_point_t extended_ring_ll =
extend_region((nccl_ofi_tuner_point_t){65536, 2},
(nccl_ofi_tuner_point_t){65536, 64},
(nccl_ofi_tuner_point_t){TUNER_MAX_SIZE, TUNER_MAX_RANKS});
nccl_ofi_tuner_point_t extended_pat_simple =
extend_region((nccl_ofi_tuner_point_t){50331648, 64},
(nccl_ofi_tuner_point_t){117440512, 128},
Expand All @@ -509,21 +505,19 @@ static ncclResult_t region_init_internal_p5en(nccl_ofi_tuner_region_context_t *r
(nccl_ofi_tuner_point_t){TUNER_MAX_SIZE, TUNER_MAX_RANKS});

const nccl_ofi_tuner_region_t regions[] = {
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL,
.num_vertices = 4,
.vertices = {{0, 2}, {65536, 2}, {65536, 64}, extended_ring_ll}},
{.algorithm = NCCL_ALGO_PAT,
.protocol = NCCL_PROTO_SIMPLE,
.num_vertices = 8,
.vertices = {extended_ring_ll,
{65536, 64},
.num_vertices = 10,
.vertices = {{0, 2},
{65536, 2},
{1048576, 2},
{16777216, 32},
{50331648, 64},
{117440512, 128},
extended_pat_simple}},
extended_pat_simple,
{TUNER_MAX_SIZE, TUNER_MAX_RANKS},
{65536, TUNER_MAX_RANKS},
{0, TUNER_MAX_RANKS}}},
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL128,
.num_vertices = 9,
Expand Down Expand Up @@ -551,10 +545,6 @@ static ncclResult_t region_init_internal_p5en(nccl_ofi_tuner_region_context_t *r
}
{
collType = ncclFuncReduceScatter;
nccl_ofi_tuner_point_t extended_ring_ll =
extend_region((nccl_ofi_tuner_point_t){65536, 2},
(nccl_ofi_tuner_point_t){65536, 64},
(nccl_ofi_tuner_point_t){TUNER_MAX_SIZE, TUNER_MAX_RANKS});
nccl_ofi_tuner_point_t extended_pat_simple =
extend_region((nccl_ofi_tuner_point_t){50331648, 64},
(nccl_ofi_tuner_point_t){117440512, 128},
Expand All @@ -564,21 +554,19 @@ static ncclResult_t region_init_internal_p5en(nccl_ofi_tuner_region_context_t *r
(nccl_ofi_tuner_point_t){301989888, 128},
(nccl_ofi_tuner_point_t){TUNER_MAX_SIZE, TUNER_MAX_RANKS});
const nccl_ofi_tuner_region_t regions[] = {
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL,
.num_vertices = 4,
.vertices = {{0, 2}, {65536, 2}, {65536, 64}, extended_ring_ll}},
{.algorithm = NCCL_ALGO_PAT,
.protocol = NCCL_PROTO_SIMPLE,
.num_vertices = 8,
.vertices = {extended_ring_ll,
{65536, 64},
.num_vertices = 10,
.vertices = {{0, 2},
{65536, 2},
{1048576, 2},
{16777216, 32},
{50331648, 64},
{117440512, 128},
extended_pat_simple}},
extended_pat_simple,
{TUNER_MAX_SIZE, TUNER_MAX_RANKS},
{65536, TUNER_MAX_RANKS},
{0, TUNER_MAX_RANKS}}},
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL128,
.num_vertices = 9,
Expand Down

0 comments on commit 75b31d0

Please sign in to comment.