O-DAG framework optimizes complex O-RAN scheduling across edge and cloud

The O-RAN paradigm decomposes intelligent RAN control into interdependent AI/ML functions (traffic prediction, signal quality estimation, slice scheduling) that must execute across a dispersed continuum of far-edge (DU, RIC), near-edge (edge), and cloud resources under strict latency and bandwidth constraints. To address the lack of an integrated methodology, Yoonjae Hwang and Bhaskar Krishnamachari present O-DAG—a four-stage framework built on top of DagProfiler, a new open-source tool that instruments the O-RAN Slice Scheduler to extract per-task instruction counts and per-edge communication volumes. The framework encodes a parameterized three-tier network topology with realistic link bandwidths, extends the SAGA scheduling framework, and includes a custom DAG simulation module on the MintEDGE simulator.

Evaluating five scheduling algorithms (HEFT, MCT, MinMin, MaxMin, Duplex) across configurations spanning 5K–50K UEs, 2–20 cells, and 2–10 network slices, the authors found HEFT consistently achieves the lowest makespan, though scheduler rankings are workload-dependent. The SAGA–simulation gap serves as a diagnostic: negative gaps (up to -1.72%) identify compute-dominated configurations where HEFT overestimates conservatively, while a positive gap (+0.64%) at low slice counts exposes a communication-bound regime where bandwidth contention exceeds model assumptions. All artifacts, including DagProfiler, are released for reproducibility, enabling the community to profile and schedule complex O-RAN applications in real 5G edge-cloud deployments.