CCCL: Node-Spanning GPU Collectives with CXL Memory Pooling
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866915985519280128 |
|---|---|
| author | Xu, Dong Meng, Han Chen, Xinyu Zhu, Dengcheng Tang, Wei Liu, Fei Xie, Liguang Xiang, Wu Shi, Rui Li, Yue Hu, Henry Zhang, Hui Jiang, Jianping Li, Dong |
| author_facet | Xu, Dong Meng, Han Chen, Xinyu Zhu, Dengcheng Tang, Wei Liu, Fei Xie, Liguang Xiang, Wu Shi, Rui Li, Yue Hu, Henry Zhang, Hui Jiang, Jianping Li, Dong |
| contents | Large language models (LLMs) training or inference across multiple nodes introduces significant pressure on GPU memory and interconnect bandwidth. The Compute Express Link (CXL) shared memory pool offers a scalable solution by enabling memory sharing across nodes, reducing over-provisioning and improving resource utilization. We propose \name, a collective communication library, leveraging the CXL shared memory pool to support cross-node GPU operations without relying on traditional RDMA-based networking. Our design addresses the challenges on synchronization, data interleaving, and communication parallelization faced by using the CXL shared memory pool for collective communications. Evaluating on multiple nodes with a TITAN-II CXL switch and six Micron CZ120 memory cards, we show that \name achieves highly efficient collective operations across hosts, demonstrating CXL's potential for scalable, memory-centric GPU communication. Our evaluation demonstrates that \name achieves average performance improvements of 1.34$\times$ for AllGather, 1.84$\times$ for Broadcast, 1.94$\times$ for Gather, and 1.04$\times$ for Scatter, compared to the original RDMA-based implementation over 200 Gbps InfiniBand. \textcolor{dong}{In addition, the evaluation with a case of LLM training shows 1.11$\times$ speedup compared with the InfiniBand while saving production cost by $2.75\times$ in hardware.} |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_22457 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | CCCL: Node-Spanning GPU Collectives with CXL Memory Pooling Xu, Dong Meng, Han Chen, Xinyu Zhu, Dengcheng Tang, Wei Liu, Fei Xie, Liguang Xiang, Wu Shi, Rui Li, Yue Hu, Henry Zhang, Hui Jiang, Jianping Li, Dong Distributed, Parallel, and Cluster Computing Emerging Technologies Large language models (LLMs) training or inference across multiple nodes introduces significant pressure on GPU memory and interconnect bandwidth. The Compute Express Link (CXL) shared memory pool offers a scalable solution by enabling memory sharing across nodes, reducing over-provisioning and improving resource utilization. We propose \name, a collective communication library, leveraging the CXL shared memory pool to support cross-node GPU operations without relying on traditional RDMA-based networking. Our design addresses the challenges on synchronization, data interleaving, and communication parallelization faced by using the CXL shared memory pool for collective communications. Evaluating on multiple nodes with a TITAN-II CXL switch and six Micron CZ120 memory cards, we show that \name achieves highly efficient collective operations across hosts, demonstrating CXL's potential for scalable, memory-centric GPU communication. Our evaluation demonstrates that \name achieves average performance improvements of 1.34$\times$ for AllGather, 1.84$\times$ for Broadcast, 1.94$\times$ for Gather, and 1.04$\times$ for Scatter, compared to the original RDMA-based implementation over 200 Gbps InfiniBand. \textcolor{dong}{In addition, the evaluation with a case of LLM training shows 1.11$\times$ speedup compared with the InfiniBand while saving production cost by $2.75\times$ in hardware.} |
| title | CCCL: Node-Spanning GPU Collectives with CXL Memory Pooling |
| topic | Distributed, Parallel, and Cluster Computing Emerging Technologies |
| url | https://arxiv.org/abs/2602.22457 |