DeepQueueNet in Practice: Quickly Achieve High‑Precision Network Simulation

DeepQueueNet (DQN) is an innovative network performance evaluation tool that combines deep learning with discrete‑event simulation to provide scalable, packet‑level accurate estimates; the underlying paper was accepted at ACM SIGCOMM 2022.

Preparation & Model Training

Hardware: at least one GPU (tested on Tesla P100‑12 GB) and ~200 GB disk space (pre‑processed data ≈180 GB).

Dataset: download from Dropbox, set up the Conda environment with conda env create -f env.yml, unzip data, run Dataset.py for preprocessing.

Training: use the PyTorch version, execute python train1000.py (set max_epoch=20); data processing takes ~1 h, model training ~15 h.

The trained device model’s probability density and cumulative distribution functions on the test set closely match the true distribution, demonstrating strong fitting ability.

Network Simulation

We configure a fattree16 topology by defining a two‑dimensional array T in fattree_model.py, where T(i,j) specifies the port connecting switch i to switch j (‑1 indicates no link). Example: T(0,4)=0 means switch 0 connects to switch 4 via port 0.

Multi‑GPU simulation is supported; device ranges are assigned per GPU (e.g., four GPUs each handling a partition of the topology as shown in the diagram).

Simulation packets are represented as an 8‑tuple (sequence number, send timestamp, byte size, priority, source, destination port, destination, path); the eighth element etime is used only for accuracy verification.

Using the provided validation dataset (30 s of traffic generated by three random processes), we run the simulation with the trained model. The predicted delays are virtually indistinguishable from the ground‑truth delays.

Performance on a Tesla P100‑12 GB GPU:

1 GPU: 5 min 12 s

2 GPU: 2 min 45 s

4 GPU: 1 min 27 s

References: https://www.dropbox.com/s/q56sx4hxe93n4g5/DeepQueueNet-dataset.zip?dl=0 (training and simulation dataset) and https://github.com/HUAWEI-Theory-Lab/deepqueuenet/tree/pytorch (official repository).