TranSys: Explaining Deep Learning-Based Networked Systems
TranSys is an integrated explainer to provide post-hoc explanations for different types of Deep Learning (DL)-based networked systems. In the current stage, we provide the explanation methods and implementations for three DL-based networked systems:
- Pensieve (
explain_pensieve) is an adaptive video streaming algorithm based on deep reinforcement learning.
- AuTO (
explain_auto) is an on-switch traffic scheduler in datacenters (under refactoring).
- RouteNet (
explain_routenet) is an SDN traffic optimizer to find routes for all src-dst pairs.
We further provide several use cases of TranSys:
- We troubleshoot the DNN in Pensieve and improve the average quality of experience (QoE) by up to 3% over DNN policies with only decision trees (
- With decision trees generated by TranSys, we lightweightify Pensieve and AuTO and achieve shorter decision-making latency by 27x on average and lower resource consumption by up to 156x (
- We also provide an efficient way to compare the latency of several paths in traffic optimization based on the explanations provided by TranSys (
The running scripts for explanation methods and use cases could be found in respective directories. Currently we are still working on documentating and refactoring the repository. Other codes will be released very soon.
For any questions, please post an issue or send an email to email@example.com.