The MIT licence in LICENSE covers the code written for this
study. The parts below come from other projects and keep their own licences.
| Path | Origin | Licence |
|---|---|---|
Experiments/ibgda_bench.cu, the dpep_* helpers |
Ported from DeepEP csrc/kernels/legacy/{utils,ibgda_device}.cuh (commit dd758ca). DeepEP's ibgda_device.cuh is derived from NVSHMEM. |
DeepEP: MIT. The NVSHMEM-derived parts remain under the NVSHMEM licence and are not MIT-licensed. This software contains source code provided by NVIDIA Corporation. |
Experiments/ibgda_bench.cu, aq_put and aq_put_warp |
Adapted from NVSHMEM's ibgda_rma_thread |
NVSHMEM licence; not MIT-licensed |
Experiments/mlx5_defs.h, the DEVX_SET macro family |
Adapted from rdma-core providers/mlx5/mlx5dv.h |
GPL-2.0 or Linux-OpenIB (BSD) |
Experiments/mscclpp-constref.patch |
Patch to MSCCL++ | MIT |
Experiments/gb200/ucclep/*.patch |
Patches to UCCL | Apache-2.0 |
results/gb200/20260914-gdaki-investigation/builds/nccl-2.31.2-original-headers.tar.gz, results/gb200/**/header.diff |
Headers from NCCL 2.31.2 and diffs against them | Apache-2.0, BSD-3-Clause, or GPL-2.0/BSD-2-Clause/Apache-2.0, as marked in each file |
results/gb200/20260913-gdrcopy-install/bundle-LICENSE and the files it covers |
GDRCopy 2.5.2 install bundle | MIT |
The source.diff records under results/gb200/ contain the same
ibgda_bench.cu code and carry the same licences.
mini-gda (Experiments/mgda_*) is written for this study. It follows the
queue and doorbell layouts of rdma-core, NVSHMEM and DOCA GPUNetIO and cites
the source of each step in its comments.
Not included:
- NVIDIA's mlx5 PRM headers (
mlx5_ifc.h,mlx5_prm.h).make vendorinExperiments/copies them from an NVSHMEM source tree. - The source of NVSHMEM, NCCL, DeepEP, UCCL-EP, MSCCL++ and fabric-lib (pplx-garden). The benchmarks build against a separate checkout of each.
- Disassembly of NVSHMEM device code. The kernel-cost campaigns keep resource
usage (
*.res.txt) and the code-generation gates, but not the SASS of NVSHMEM builds.