Please note: This master’s thesis presentation will take place online.
Qi Fan (Eric) Yan, Master’s candidate
David R. Cheriton School of Computer Science
Supervisors: Professor Raouf Boutaba
In-network aggregation (INA) reduces distributed-training traffic by aggregating gradient updates in programmable switches. However, in conventional multi-tier, multi-tenant clusters, changing heterogeneous workloads and background traffic continually alter the switch memory and link bandwidth available across aggregation trees. Existing systems either optimize resource use within a fixed tree or reconsider tree assignments only at admission or minute-scale intervals, causing them to miss fine-grained opportunities in both installed and candidate 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 rapidly identify and validate candidate trees using job-local observations. Strong tree safety guarantees enable seamless steering toward better-resourced trees, increasing Exploitation effectiveness. Combining exploration and exploitation, DNATree achieves 30% average goodput improvement over the state-of-art competitors in multi-tenant data center environments while remaining feasible on commodity programmable switches.
To attend this master's thesis presentation you can join virtually through Teams.