Accurately Reproducing Application–Cross-Traffic Interactions Under Dynamic Conditions

Establish rigorous methods to accurately account for and precisely reproduce the complex closed-loop interactions between application-generated traffic and competing cross-traffic under dynamic network conditions, so that experimental emulation and data generation faithfully reflect real-world protocol and application behaviors at bottleneck links.

Background

The paper introduces NetReplica to generate realistic and controllable network datasets for training ML models that generalize across domains. A central challenge to realism is capturing the closed-loop dynamics produced by the mutual interactions between target application traffic and cross-traffic at bottleneck links. Record–replay and model-based approaches often simplify these dynamics, leading to unrealistic behaviors that hinder domain adaptation.

Within the discussion of cross-traffic replay, the authors explicitly acknowledge that accurately accounting for and reproducing these interactions is a longstanding unresolved problem. NetReplica adopts a pragmatic hybrid approach—using endogenous packet traces for cross-traffic while not letting them react to exogenous target traffic during replay—to balance fidelity with controllability. This workaround underscores the open problem of developing replay or modeling techniques that preserve true reactive cross-traffic behavior without sacrificing practical control.