Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

m4: A Learned Flow-level Network Simulator

Chenning Li, Anton A. Zabreyko, Arash Nasr-Esfahany, Kevin Zhao, Prateesh Goyal, Mohammad Alizadeh, Thomas Anderson

TL;DR

Flow-level simulators scale to data-center networks but sacrifice accuracy by omitting packet-level effects. m4 tackles this gap with a learned flow-level model that decomposes state transitions into spatial (Graph Neural Network) and temporal (GRU) components, augmented by dense per-event supervision signals such as remaining flow size and queue length, to capture congestion control and queuing dynamics. Trained on ns-3 data and evaluated on large-scale fat-tree topologies, m4 achieves up to 104× speedups over packet-level simulation and reduces per-flow FCT slowdown errors by about 45.3% (mean) and 53.0% (p90) relative to traditional flow-level models, while generalizing across workloads and CC schemes. This approach enables accurate, scalable, closed-loop network simulations suitable for design-space exploration and application-level performance forecasting, with practical impact on data-center network design and ML training/inference workloads.

Abstract

Flow-level simulation is widely used to model large-scale data center networks due to its scalability. Unlike packet-level simulators that model individual packets, flow-level simulators abstract traffic as continuous flows with dynamically assigned transmission rates. While this abstraction enables orders-of-magnitude speedup, it is inaccurate by omitting critical packet-level effects such as queuing, congestion control, and retransmissions. We present m4, an accurate and scalable flow-level simulator that uses machine learning to learn the dynamics of the network of interest. At the core of m4 lies a novel ML architecture that decomposes state transition computations into distinct spatial and temporal components, each represented by a suitable neural network. To efficiently learn the underlying flow-level dynamics, m4 adds dense supervision signals by predicting intermediate network metrics such as remaining flow size and queue length during training. m4 achieves a speedup of up to 104$\times$ over packet-level simulation. Relative to a traditional flow-level simulation, m4 reduces per-flow estimation errors by 45.3% (mean) and 53.0% (p90). For closed-loop applications, m4 accurately predicts network throughput under various congestion control schemes and workloads.

m4: A Learned Flow-level Network Simulator

TL;DR

Flow-level simulators scale to data-center networks but sacrifice accuracy by omitting packet-level effects. m4 tackles this gap with a learned flow-level model that decomposes state transitions into spatial (Graph Neural Network) and temporal (GRU) components, augmented by dense per-event supervision signals such as remaining flow size and queue length, to capture congestion control and queuing dynamics. Trained on ns-3 data and evaluated on large-scale fat-tree topologies, m4 achieves up to 104× speedups over packet-level simulation and reduces per-flow FCT slowdown errors by about 45.3% (mean) and 53.0% (p90) relative to traditional flow-level models, while generalizing across workloads and CC schemes. This approach enables accurate, scalable, closed-loop network simulations suitable for design-space exploration and application-level performance forecasting, with practical impact on data-center network design and ML training/inference workloads.

Abstract

Flow-level simulation is widely used to model large-scale data center networks due to its scalability. Unlike packet-level simulators that model individual packets, flow-level simulators abstract traffic as continuous flows with dynamically assigned transmission rates. While this abstraction enables orders-of-magnitude speedup, it is inaccurate by omitting critical packet-level effects such as queuing, congestion control, and retransmissions. We present m4, an accurate and scalable flow-level simulator that uses machine learning to learn the dynamics of the network of interest. At the core of m4 lies a novel ML architecture that decomposes state transition computations into distinct spatial and temporal components, each represented by a suitable neural network. To efficiently learn the underlying flow-level dynamics, m4 adds dense supervision signals by predicting intermediate network metrics such as remaining flow size and queue length during training. m4 achieves a speedup of up to 104 over packet-level simulation. Relative to a traditional flow-level simulation, m4 reduces per-flow estimation errors by 45.3% (mean) and 53.0% (p90). For closed-loop applications, m4 accurately predicts network throughput under various congestion control schemes and workloads.

Paper Structure

This paper contains 23 sections, 6 equations, 13 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Why flow-level simulation?
  4. Learning flow-level models from data
  5. Key Insight
  6. System Design of m4
  7. High-Level Overview
  8. m4's Neural Network Architecture
  9. Handling network states
  10. Tracking network states
  11. Querying flows and links
  12. Adding Dense Supervision Signals
  13. Using GNN on the Network Snapshot
  14. What m4 Does and Does Not Do
  15. Implementation
  16. ...and 8 more sections

Figures (13)

  • Figure 1: m4 mimics the computational structure of flowSim but replaces its components with learnable modules.
  • Figure 2: m4’s workflow: Inputs (yellow boxes), outputs (red boxes), intermediate components (white boxes).
  • Figure 3: m4 adds "dense" supervision during training by querying intermediate network states for "remaining size" and "queue length". Dashed boxes represent subsequent simulations triggered by new flow-level events.
  • Figure 4: m4 converts (a) a network snapshot in time to a (b) bipartite graph and uses GNN to capture spatial dynamics.
  • Figure 5: m4's implementation
  • ...and 8 more figures