TL/MLX5: various optimizations #1012
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
MamziB
reviewed
Sep 25, 2024
| ucc_tl_mlx5_alltoall_t *a2a = team->a2a; | ||
| int seq_index = task->alltoall.seq_index; | ||
| int npolls = UCC_TL_MLX5_TEAM_CTX(team)->cfg.npolls; | ||
| int radix = UCC_TL_MLX5_TEAM_LIB(team)->cfg.fanin_kn_radix; |
MamziB
reviewed
Sep 25, 2024
| } | ||
| *dist *= radix; |
MamziB
reviewed
Sep 25, 2024
| @@ -242,7 +282,13 @@ static ucc_status_t ucc_tl_mlx5_fanout_start(ucc_coll_task_t *coll_task) | |||
|
|
|||
| tl_debug(UCC_TASK_LIB(task), "fanout start"); | |||
| /* start task if completion event received */ | |||
| UCC_TL_MLX5_PROFILE_REQUEST_EVENT(task, "mlx5_alltoall_fanout_start", 0); | |||
| if (team->a2a->node.sbgp->group_rank == team->a2a->node.asr_rank) { | |||
There was a problem hiding this comment.
can we remove this if/else when PROFILING is not requested? this is to improve the performance when profiling is not needed
MamziB
reviewed
Sep 25, 2024
| ucc_tl_mlx5_alltoall_t *a2a = team->a2a; | ||
| int node_size = a2a->node.sbgp->group_size; | ||
| int net_size = a2a->net.sbgp->group_size; | ||
| int op_msgsize = node_size * a2a->max_msg_size * UCC_TL_TEAM_SIZE(team) * |
MamziB
reviewed
Sep 25, 2024
| int block_msgsize = block_h * block_w * task->alltoall.msg_size; | ||
| ucc_status_t status = UCC_OK; | ||
| int node_grid_w = node_size / block_w; | ||
| int node_nbr_blocks = (node_size * node_size) / (block_h * block_w); |
MamziB
reviewed
Sep 25, 2024
| if (!send_to_self && | ||
| task->alltoall.op->blocks_sent[cyc_rank] < node_nbr_blocks) { | ||
| dm = ucc_tl_mlx5_a2a_wait_for_dm_chunk(task); | ||
| if (status != UCC_OK) { |
There was a problem hiding this comment.
which status you are checking?
MamziB
reviewed
Sep 25, 2024
| @@ -28,6 +28,19 @@ static ucc_config_field_t ucc_tl_mlx5_lib_config_table[] = { | |||
| ucc_offsetof(ucc_tl_mlx5_lib_config_t, dm_buf_num), | |||
| UCC_CONFIG_TYPE_ULUNITS}, | |||
|
|
|||
| {"FORCE_REGULAR", "y", | |||
There was a problem hiding this comment.
can you encapsulate all these three env vars into one with 3 options
MamziB
reviewed
Sep 25, 2024
| c->offset = (ptrdiff_t)team->dm_offset; | ||
| team->dm_offset = PTR_OFFSET(team->dm_offset, | ||
| UCC_TL_MLX5_TEAM_LIB(team)->cfg.dm_buf_size); | ||
| c->addr = (uintptr_t)PTR_OFFSET( |
|
Can one of the admins verify this patch? |
TL/MLX5: add npolls cfg for FANIN TL/MLX5: knomial fanin TL/MLX5: add prints and profile events TL/MLX5: remove debug prints
tiny bit more robust print blocks dimensions fully working configurable batch_size, serialization, and pollings
clean and working TL/MLX5: add more config for block dimensions force longer by default
lintrunner cleaning
samnordmann
force-pushed
the
tl/mlx5/pr/main_clean_history
branch
from
32d0718 to
43dd1d7
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What
This PR contains various optimizations for TL/MLX5/a2a. In order of importance/relevance:
We might want to merge this PR as is, or to divide it into several smaller ones. But this branch is at least a pointer for a working version, that can be used as is for performance experimentation.
TODO:
One important optimization that is yet to be implemented is to support using several NICs. So far, our algorithm only uses one NIC.
cc @lappazos @x41lakazam