Paper Accepted to ACM SoCC 2026
Our paper titled “FluxReduce: Straggler-Resilient In-Network Aggregation for Distributed Machine Learning” has been accepted to the 17th Annual ACM Symposium on Cloud Computing (SoCC 2026), a flagship conference in cloud computing and data center systems, held in Singapore during November 18-20, 2026.
FluxReduce is an in-network aggregation system that detects and mitigates stragglers directly in the switch dataplane, filling a gap that prior programmable-switch aggregation work has overlooked. Since detection and mitigation are driven solely by the aggregation traffic, with no timers or external coordination, the design maps naturally onto pipelined, multi-terabit switch ASICs, and the Intel Tofino2 prototype is the first straggler-aware in-network aggregation system to run at line rate.
Here is the abstract of the paper:
In-network aggregation is a promising technique for accelerating gradient synchronization in distributed ML training by performing aggregation directly on programmable switches. However, the presence of stragglers can significantly undermine the gains and is largely overlooked in the literature. We present FluxReduce, an in-network aggregation system that detects and mitigates stragglers entirely within the network dataplane. Detection and mitigation are driven by the aggregation traffic itself, requiring no timers or external coordination, which makes the design a natural fit for pipelined, multi-terabit switch ASICs. Our prototype on an Intel Tofino2 switch is the first straggler-aware in-network aggregation system to run at line rate, and matches the AllReduce performance of straggler-oblivious in-network aggregation, offering 1.37x higher throughput than NCCL’s RDMA-based ring AllReduce. Across several real-world DNNs, our evaluation shows faster training in the presence of stragglers by 1.08-1.53x, and near-zero overhead in their absence, compared to straggler-oblivious training.