Revisiting Bruck: Phase-Efficient All-to-All Communication in Reconfigurable Networks

Jul 16, 2026

All-to-all communication becomes a major performance bottleneck in distributed ML and HPC workloads because dense traffic stresses scale-up interconnects. Optical reconfigurable networks enable runtime topology changes that adapt to the active workload, but require global synchronization; communication pauses during reconfiguration introduce measurable delay. The paper revisits Bruck’s all-to-all strategy and shows how co-designing the communication pattern with the reconfiguration strategy enables phase-structured, sparse topology states. It introduces ReTri and reports large completion-time improvements under realistic reconfiguration overhead.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Problem: All-to-All as a Bottleneck

Optical Reconfigurable Networks and Reconfiguration Trade-offs

Motivation: Phase-Structured Collective Schedules

ReTri and Co-designed Topology Optimization

Revisiting Bruck: Phase-Efficient All-to-All Communication in Reconfigurable Networks Anton Juerss Weizenbaum Institute & TU Berlin Stefan Schmid TU Berlin & Weizenbaum Institute arXiv :2605 .26930v2 [ cs .DC] 9 Jul 2026 Abstract All-to-All communication is a key performance bottleneck for distributed machine learning (ML) and high-performance computing (HPC) workloads, where dense traffic increasingly stresses scale-up interconnects. While these ML and HPC workloadshave driven unprecedented infrastructure demand, optical reconfigurable networks (ORNs) offer a promising path forward as they can reconfigure the network at runtime. By adapting the physical topology to the active workload, they improve communication cost and bandwidth utilization. However, optical reconfigurable networks introduce a fundamental trade-off for collective communication: each reconfiguration requires global synchronization, during which communication is suspended for at a non-negligible delay. Additionally, their benefit is critically contingent on whether the collective consists of structured phases that can be served by sparse and reusable topology states. In this paper, we revisit Bruck’s All-to-All implementation and demonstrate the benefits of topology optimization in which both communication pattern and reconfiguration strategy are co-designed. We present ReTri, a bidirectional All-to-All schedule for ORNs based on the Trivance algorithm. ReTri uses balanced ternary block propagation to complete All-to-All in ⌈log3 􀀽⌉ phases. The reconfiguration strategy induced by ReTri’s pairwise bidirectional exchanges allows reconfiguration delays tobe amortized across multiple phases. Preliminary simulations show that ReTri improves completion time by up to 10× over Pairwise All-to-All, even for millisecond-scale reconfiguration delays, and improves reconfigurable Bruck by up to 2. 1× . 1 Introduction To meet the increasing compute and memory demands of deep learning workloads, including recommendation systems and Mixture of Expert models, modern distributed systems connect thousands of accelerators in hyperscale datacenters [12, 17, 21, 28] . In these large-scale ML systems, efficient communication across accelerators is crucial for training and inference performance, since activations, embeddings, and tokens must be synchronized [17] . These synchronizations are typically realized through All-to-All collective communication, where each accelerator sends distinct data to every other accelerator [27] . Their impact on end-toend performance is substantial, accounting for up to 55% of MoE end-to-end training time [17] . The dense communication pattern of All-to-All makes it a major performance bottleneck and increasingly stresses scale-up interconnects [15]: every accelerator exchanges distinct data, raising serious concerns about congestion and requiring high network bandwidth [17, 21] . The design of datacenter interconnect fabrics therefore plays a central role in the scalability and efficiency of largescale ML and HPC systems [7, 12, 21] . Whereas conventional, electrically switched networks are power-intensive and may lead to performance bottlenecks, optical reconfigurable networks (ORNs) have emerged as a promising alternative. ORNs establish bidirectional high-bandwidth optical links between endpoints, introducing the enhanced capability to adjust and optimize the physical topology [3, 4, 11] . However, direct optical connectivity incurs non-negligible reconfiguration delay [7, 10, 20] . Whether reconfiguration is beneficial depends on balancing its delay against the resulting reductions in path length, congestion, and transmission time. Determining when to reconfigure is itself non-trivial: per-phase reconfiguration may incur excessive overhead, whereas a static topology can forfeit substantial communication gains [2, 28] . This motivates collective schedules whose communication phases admit a small number of efficient and reu

FAQs

How does ReTri improve upon pairwise all-to-all scheduling in reconfigurable networks?

ReTri uses a bidirectional all-to-all schedule based on the Trivance algorithm, completing the communication in logarithmic phases and amortizing reconfiguration delays across multiple phases, leading to up to 10× faster completion time in preliminary simulations.

Jul 16, 2026

Upload by Olivia Brown