DNN partitioning: where to split?

Problem

A neural network is a chain of layers. Input: uplinkMbps rttMs inputKB n, then n layers name deviceMs edgeMs outputKB.

Split k (0 ≤ k ≤ n) runs layers 0 … k−1 on the device and k … n−1 on the edge. For k < n, the device sends the data entering layer k (the input for k = 0, else the output of layer k−1) — transfer time = KB × 8 / 1000 / Mbps seconds — then pays the RTT and the edge times. Split n runs everything on the device and sends nothing.

Print split 3 (after pool): send 30 KB, latency 62.0 ms for each k (all on edge for k = 0, all on device for k = n), then Best: split k, latency X ms (the smallest latency; ties → the smaller k).

Your solution

Write it here or in your IDE, then paste it. Compile and test it yourself before comparing. Your code stays in your browser — it is never sent to or stored on the server.

← OS13.2 Multi-user offloading: best-response dynamics OS13.4 Deadline- and cost-aware placement across device, edge and cloud →