Open
Conversation
…ulti-node): stride/device grouping and NIC planes. - `STRIDE` — Step size for stride orbits on rank-major devices (`gcd` with total device count); no traffic between different orbits. - `GROUP_SIZE` — Devices per subgroup inside each stride orbit (natural rank-major order); must divide orbit size (default: all devices per rank × GPUs). - `NIC_A2A_SCOPE` — `intra`: transfers only within the same device subgroup; `inter`: only between different subgroups (same stride orbit only). - `NIC_A2A_NO_SAME_RANK` — When non-zero, omit transfers where source and destination are the same rank. - `NUM_NIC_PLANES` — Split NIC endpoints into this many disjoint planes (rank-major index modulo planes); traffic only between NICs in the same plane.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Description
Adding nica2a preset (NIC all-to-all over GPUs via NIC executors, multi-node): stride/device grouping and NIC planes.
STRIDE— Step size for stride orbits on rank-major devices (gcdwith total device count); no traffic between different orbits.GROUP_SIZE— Devices per subgroup inside each stride orbit (natural rank-major order); must divide orbit size (default: all devices per rank × GPUs).NIC_A2A_SCOPE—intra: transfers only within the same device subgroup;inter: only between different subgroups (same stride orbit only).NIC_A2A_NO_SAME_RANK— When non-zero, omit transfers where source and destination are the same rank.NUM_NIC_PLANES— Split NIC endpoints into this many disjoint planes (rank-major index modulo planes); traffic only between NICs in the same plane.Motivation
Need to stress scale-out network independently from intra node communication with Alltoall pattern
Technical Details
The pattern is defined to use all NICs, can be used in the case of several NICs per GPU .
The stride startegy differ from the one defined in PodAlltoall . Here natural order inside each stride orbit is used. Syntax close to the one of PodAlltoall.
The scope allows to considers running inside the same groups but the need was also to run only between groups.
Also an option in terms of number of planes have been added (can also be used in rail topology) to be able to restrict a2a pattern to have only peers connected through scale-out network.
Test Plan
Tested up to 6 nodes varying different variables.
Test Result
Expected patterns used in each case