The offloading decision logic in offloading_algo.py needs a correctness review.
Areas to check:
- The evaluation formula:
initial_cost + (layer_data_size / avg_speed) + edge_computation_cost — verify it accurately models end-to-end latency
- Fallback behavior when
avg_speed is 0 (currently defaults to 1)
- Whether the valid offloading points filter correctly excludes skip-connection layers in all cases
- Consistency of EMA smoothing across device times (α=0.5), edge times (α=0.5), and per-layer inference times (α=0.2)
Goal: Confirm the algorithm reliably selects the optimal offloading layer under varying network and compute conditions, or identify where it may produce suboptimal decisions.
The offloading decision logic in
offloading_algo.pyneeds a correctness review.Areas to check:
initial_cost + (layer_data_size / avg_speed) + edge_computation_cost— verify it accurately models end-to-end latencyavg_speedis 0 (currently defaults to 1)Goal: Confirm the algorithm reliably selects the optimal offloading layer under varying network and compute conditions, or identify where it may produce suboptimal decisions.