DNATree: Dynamic Tree Adaptation for Multi-Tenant In-Network Aggregation

Abstract

In-network aggregation (INA) reduces distributed-training traffic by aggregating gradient updates in programmable switches. However, in multi-tier, multi-tenant clusters, heterogeneous workloads and background traffic continually alter the aggregation state and link bandwidth available across aggregation trees. Existing systems either optimize resource use within a fixed tree or reconsider tree assignments only at job admission or minute-scale intervals, limiting their ability to exploit changing resource availability within and across aggregation trees. We present DNATree, a multi-tenant INA service that adapts at millisecond timescales by combining flexible aggregation (i.e., Exploitation) with dynamic tree adaptation (i.e., Exploration). Exploitation progressively uses available switch memory and link bandwidth along the installed tree without fixed per-job allocations. This decouples fine-grained resource allocation from tree selection, enabling Exploration to identify and validate candidate trees using job-local observations. Safe tree transitions steer new work toward better-resourced trees without disrupting in-flight aggregation. On a Tofino~1 testbed, DNATree achieves up to 74.3% higher system goodput over fixed-partition progressive aggregation under constrained aggregation capacity. Packet-level simulations show a 25% average improvement over the best-performing aggregation-tree adaptation baseline across fat-tree and leaf-spine datacenter topologies.