A two-layer collective communication framework that unlocks underutilized PCIe bandwidth and jointly harnesses dedicated interconnects and the PCIe bus.
- 💡 PCIe CCL accelerates collective communication on the PCIe path by making the CPU-rooted PCIe hierarchy explicit and treating NUMA-local host memory as an active routing point, powered by a CPU-driven DSL for topology-aware schedules and an auto-pipelined runtime built on GPU copy engines.
- 🔀 Multipath Orchestrator splits each collective across the dedicated link (via the vendor CCL) and the PCIe path (via PCIe CCL) on independent streams with a fork–join model, and a feedback-driven adaptive balancer retunes the split ratio at runtime to keep the two paths balanced under PCIe interference.
See each component's README for build and usage instructions:
- PCIe CCL — CUDA: PCIe CCL for NVIDIA CUDA.
- PCIe CCL — Ascend: PCIe CCL for Huawei Ascend CANN.
- Multipath Orchestrator: Multipath Orchestrator for Huawei Ascend CANN.
Your citations are greatly appreciated. 🥰
- ACM Reference
Yuezheng Liu, Menghao Zhang, Xuebin Song, Juner Shen, Chunming Hu, and Xudong Liu. 2026. Towards High-Throughput Collective Communication via Harnessing PCIe and Dedicated Interconnects. In Proceedings of the 2026 ACM SIGOPS Annual Technical Conference (ATC 2026).
- Bibtex
@inproceedings{hylink2026, author = {Liu, Yuezheng and Zhang, Menghao and Song, Xuebin and Shen, Juner and Hu, Chunming and Liu, Xudong}, title = {Towards High-Throughput Collective Communication via Harnessing PCIe and Dedicated Interconnects}, year = {2026}, booktitle = {Proceedings of the 2026 ACM SIGOPS Annual Technical Conference}, series = {ATC 2026} }
