SPLIDT: Partitioned Decision Trees for Scalable Stateful Inference at Line Rate
summary
SPLIDT implements partitioned decision trees on switches/SmartNICs by cutting trees into per-subtree partitions, training each subtree on fixed-size packet windows, and using in-band recirculation to time-share TCAM/SRAMM and register resources. A Bayesian DSE jointly picks tree depth, partitioning, and per-subtree top-K features; training iterations take ~5 minutes and converge within ~150 iterations worst-case. Evaluation on real datasets shows up to 5x more stateful features, line-rate inference with ≤0.05% recirculation overhead, comparable time-to-detection, and scalability to millions of flows.
AI-written summary · about
chapters (10)
- 0:00SPLIDT introduction
- 2:25Traditional inference limits
- 4:52Partitioned inference recirculation
- 7:19Partitioned model design
- 9:26Bayesian model partitioning
- 11:56Training with flow windows
- 12:59Flow window sizes
- 13:43Recirculation bandwidth limits
- 14:25Evaluation results
- 16:51Summary and Q&A
related talks
18mLeo: Online ML-based Traffic Classification at Multi-Terabit Line Rate
14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
17mPrecise Data Center Traffic Engineering with Constrained Hardware Resources
18mLearning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane
35mUnderstanding, Simplicity, And Decision Trees
20mDeep distributed decision trees on Apache Spark
15mEnabling In-situ Programmability in Network Data Plane: From Architecture to Language