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.
There was a problem hiding this comment.
Pull request overview
Adds a new nica2a preset to TransferBench to generate NIC-executor-based all-to-all traffic across GPUs with configurable device stride/grouping, intra/inter-group scope, same-rank filtering, and NIC plane partitioning—intended to stress scale-out networking independently of intra-node GPU communication.
Changes:
- Register new
nica2apreset in the preset map. - Implement
NicAllToAllPresetwith STRIDE/GROUP_SIZE orbit grouping, NIC plane filtering, and scope controls. - Document the new preset and its environment variables in the changelog.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.
| File | Description |
|---|---|
| src/client/Presets/Presets.hpp | Adds nica2a preset entry and includes the new preset header. |
| src/client/Presets/NicAllToAll.hpp | Implements the NIC all-to-all preset logic, transfer generation, and reporting table. |
| CHANGELOG.md | Documents the new preset and its configuration knobs. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
|
I'll test the new changes later today and tomorrow |
… limits - Add missing `bool bytesSpecified` param to NicAllToAllPreset to match PresetFunc typedef; this was causing a compile error - Fix per-NIC bandwidth attribution in RDMA-read mode: use dstRanks[i] as the row index (executor is on dstRank, not srcRank) and exeDevice.exeIndex as the NIC column (always the executor's NIC, consistent with NicPeerToPeer) - Replace std::numeric_limits<double>::min() with ::lowest() for nicMax and rankMax to match NicRings.hpp convention - Fix spacing alignment of NicAllToAllPreset entry in preset map
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
|
@pierreantoineH was encountering build error, so pushed a fix. |
Default GROUP_SIZE to orbitSize (M / gcd(STRIDE, M)) instead of M so the preset works out-of-the-box with non-trivial STRIDE values. Update CHANGELOG to document the stride-aware default.
|
if i understand correctly, |
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