# DualColl: Throughput-Optimal Bandwidth-Bound Collective Communications via Fractional Tree Packing

Preprint · Namir Ballan, Reuven Cohen

- Archive id: arxiv:2610.15704
- Published: 2026-10-11
- arXiv: https://arxiv.org/abs/2610.15704
- Citations: 0
- Source URL: https://arxiv.org/abs/2610.15704
- PDF: https://arxiv.org/pdf/2610.15704
- Provenance: arxiv

## Abstract

Large-scale distributed training increasingly makes collective communication bandwidth-bound. In modern GPU clusters, communication libraries such as NCCL implement collectives including AllReduce over heterogeneous interconnects such as NVLink, InfiniBand, and Ethernet, making efficient use of available network bandwidth a central systems challenge. ForestColl showed that collective communication on such fabrics can be formulated using fractional tree packing; its scalable polynomial-time construction is developed for AllGather, with AllReduce obtained by composing AllGather and ReduceScatter. We present DualColl, which targets bandwidth-bound AllReduce directly. Under the tree-based AllReduce model, DualColl applies a Garg--Könemann-style multiplicative-weights algorithm and, for any $0\<ε\<1$, computes a feasible schedule whose throughput is at least $(1-ε)^2$ of the optimal AllReduce throughput. Its key component is an exact minimum-cost collective-tree oracle that projects physical-link dual prices onto a logical graph of participating ranks, reducing each iteration to shortest-path computations followed by a minimum spanning tree. We further extend the framework to local equal-next-hop packet-sprayed routing and to networks without explicit per-tree bandwidth control. For the latter, we introduce a tree-coupled max-min bandwidth-sharing model and show that tree duplication approximates an explicit fractional allocation with a provable accuracy-complexity tradeoff. Finally, we extend the fractional-packing framework to additional collective operations and multiple concurrent collective groups.

_Abstract reproduced under CC0-1.0 (arXiv descriptive metadata)._

Topics: [Algorithms & Complexity Breakthroughs](https://space-kitty.com/topics/algorithms-complexity)

---
Space Kitty Archive · https://space-kitty.com · API: https://space-kitty.com/api/v1 · MCP: https://space-kitty.com/mcp · Premium: $0.20/request via x402 (Base Sepolia (testnet)). Metadata: arXiv (CC0), OpenAlex (CC0), Europe PMC (CC BY/CC0 works), ClinicalTrials.gov.