Doctoral Thesis: AI-Driven Modeling and Optimization of Networks

By: Chenning Li

Presenter’s Affiliation (CSAIL, RLE, LIDS, MTL, etc.): CSAIL

Thesis Supervisor(s): Mohammad Alizadeh

Details

• Date: Friday, June 12
• Time: 11:00 am - 12:30 pm
• Location: 32-G882

Additional Location Details:

Abstract: As data center networks scale, traditional network performance models face a critical speed–fidelity tradeoff: packet-level simulators are accurate because they capture detailed packet-level behavior, but are too slow to be practical at data center scale; analytical and flow-level models are fast, but can be inaccurate because they oversimplify queueing and congestion-control dynamics. This thesis shows that machine learning (ML) can create performance models that are both accurate and fast: accurate because they learn complex packet-level behavior from data, and fast because they operate at higher abstractions than packet-level simulation.

This thesis develops two learned performance models, m3 and m4, and uses m3 to support closed-loop network control in Polyphony. First, m3 is a learned aggregate-performance model for estimating data center tail-latency performance. It uses path-level decomposition and workload featurization to estimate aggregate flow completion times; on a 384-rack, 6,144-host fat-tree topology, it achieves over 1000x speedup over packet-level simulation with less than 10% error. Second, m4 is a learned flow-level network backend for distributed-application simulation: frontends issue per-transfer requests, and m4 returns online completion-time callbacks that can trigger subsequent dependent transfers. m4 represents the network as a stateful flow–link bipartite graph, updates only the affected subgraph on each flow-level event, and uses dense supervision from intermediate network signals to improve accuracy. When integrated with a distributed-application simulator as the network backend, m4 delivers 40x speedup over packet-level simulation while maintaining accuracy. Third, Polyphony demonstrates how m3 can support closed-loop network control. It monitors workloads, uses m3 for counterfactual prediction, and applies the best candidate configuration. On a physical testbed, Polyphony meets service level objectives within ten minutes, while a model-free baseline fails to converge.

Together, these systems show that learned models can make network modeling practical for large-scale design exploration, distributed-application simulation, and responsive network control.

Host

• Chenning Li