Our paper "On-Demand Container Partitioning for Distributed Machine Learning" has been accepted at USENIX Annual Technical Conference (ATC) 2025 with a 15.7% acceptance rate (100 out of 634 submissions).
Led by Giovanni Bartolomeo in collaboration with Navidreza Asadi, Prof. Wolfgang Kellerer, Prof. Jörg Ott, and the SPEAR Lab team, this work addresses efficiently managing and distributing ML workloads across edge and cloud resources.
The approach enables dynamic partitioning of ML containers based on runtime requirements, on-demand resource allocation for training and inference, reduced data movement through intelligent placement, and improved resource utilization across heterogeneous infrastructures.
The paper achieved all three USENIX artifact badges (Available, Functional, Reproduced), demonstrating the robustness and reproducibility of the work. The techniques are being integrated into Oakestra, our open-source edge orchestration framework.
