Source-linked AI summary
MINT: Modeling GenAI Impact on Network Traffic
Andrew Nguyen, Samson Kempiak, Agrim Gupta, Koushik Kar, Ish Kumar Jain
TL;DR
Packet-level simulators lack measurement-driven GenAI traffic models, despite GenAI traffic varying by modality and response behavior. MINT collects isolated client-side traces and models their burst patterns in ns-3, matching empirical burst timing within 2–25% normalized Wasserstein distance. The paper reports that MINT models single-shot text and image API interactions.
Problem
Existing network studies largely characterize GenAI traffic through coarse aggregate trends rather than packet-level models, limiting realistic evaluation of scheduling and capacity planning.
Source: §1 · p. 1Method
MINT isolates API traffic in network namespaces, collects client-side traces across modalities and providers, and models burst behavior for ns-3.
Source: §1 · p. 2Results
MINT’s ns-3 simulations reproduce empirical burst timing within 2–25% normalized Wasserstein distance, while a constant token-streaming baseline deviates by up to 491%.
Source: §4.3 · p. 6Takeaways & Limitations
MINT provides an open-source, measurement-driven GenAI network simulator model, with traffic patterns that vary by modality and differ from bulk transfer and video streaming.
Source: §5 · p. 8MINT currently models single-shot text and image API interactions.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Generative AI (GenAI) is becoming a mainstream network workload, yet packet-level simulators lack measure\-ment-driven GenAI traffic models. Currently researchers must approximate GenAI services using traditional sources such as file transfer and video streaming, limiting realistic network evaluation of scheduling and capacity planning. We present MINT, a measurement and modeling framework for GenAI network traffic. Using an isolated net\-work-namespace capture pipeline, we collect client-side traces from three LLM providers across four modalities, cloud and edge servers, and wired and wireless network access points. We find that GenAI modalities exhibit distinct upload/download asymmetry and burst structures that differ from traditional applications. MINT clusters and models these burst regimes and validate empirical burst timing distribution behavior in ns-3 with normalized Wasserstein distances of 2--25\%. Our results also reveal that realistic packet bursts have significantly more variability than constant token generator models. MINT open-sources the first measurement-driven GenAI traffic model for packet-level network simulation.
1 INTRODUCTION
MINT addresses the lack of packet-level GenAI traffic models with isolated measurements and a model validated in ns-3. Its measurements show modality-dependent traffic patterns, and its simulated burst timing matches empirical distributions.
1 INTRODUCTION
- GenAI traffic varies in upload/download asymmetry and packet-burst timing with modality, server processing, and response generation, unlike predictable video buffer-refill patterns.These dynamics complicate scheduling and capacity planning, while existing network studies largely use coarse aggregate trends.Source: §1 · p. 1
- MINT isolates API traffic in network namespaces and collects nearly 10,000 client-side traces across four modalities, three providers, cloud and edge servers, and Ethernet, Wi-Fi, and 5G.It models burst behavior at different timescales and implements the model in ns-3.Source: §1 · p. 2
- MINT separates packet serialization, sub-bursts, bursts, and initial server delay, and reports burst interarrival times consistent across Ethernet, Wi-Fi, and 5G.The authors attribute this burst structure to the application rather than the access network.Source: §1 · p. 2
- The downlink share varies with modality and application phase, ranging from 0% to 100%, rather than matching a single aggregate split.The cited macro-level reports give a 74%/26% downlink/uplink split.Source: §1 · p. 2
2 BACKGROUND AND RELATED WORK
The background contrasts established traffic models with GenAI’s server-delay and response-burst behavior. Prior GenAI studies measure token-level or network-level behavior, but do not provide a measurement-driven model of client-side packet bursts.
2.1 Background
- Video streaming is modeled as ON–OFF buffer refill and file transfer as bulk delivery, whereas GenAI calls need server delay plus token bursts or a bulk phase.Aggregate Poisson Pareto burst models discard per-application structure.Source: §2.1 · p. 2
2.2 Related Work
- Token-level studies characterize generation or streaming algorithms, but do not capture how providers convert tokens into packet bursts received by clients.Consequently, token metrics cannot drive a packet-level network model.Source: §2.2 · p. 2
- Network measurement studies cover goals including streaming, classification, and impact analysis, with prior work spanning multiple modalities and access networks.These studies include the first multi-modality measurements and first measurements across multiple access networks.Source: §2.2 · p. 2
- Prior systems do not combine prompt diversity, modality diversity, and access-network isolation or reproduce client packet bursts using measurement-driven timings and durations.Prior burstiness summaries often reduce behavior to aggregate statistics such as coefficient of variation.Source: §2.2 · p. 3
3 MINT DESIGN
MINT replays prompts against cloud and edge LLMs, captures each session’s traffic, and processes traces for burst analysis and ns-3 modeling. Its design combines multiple providers, modalities, capture paths, and methods for segmenting and cleaning packet traces.
3 MINT DESIGN
- MINT’s Data Engine, Data Extraction tool, and Data Processing stage replay prompts, isolate session captures, and prepare traces for burst analysis and ns-3 implementation.Captures use Ethernet, Wi-Fi, or a 5G modem; an edge-hosted LLM helps validate client measurements against ground truth.Source: §3 · p. 3
3.1 MINT Data Engine
- The data engine measures OpenAI, Anthropic, and Qwen across text, image, and code modalities, then focuses burst modeling on ChatGPT.Prompt datasets represent reproducible real-user inputs; multi-turn prompting and human think-time delays are outside the study’s scope.Source: §3.1 · p. 3
3.2 MINT Data Extraction Tool
- For cloud APIs, MINT captures traffic at a veth interface inside a dedicated Linux network namespace; browser comparisons use Selenium and dedicated Chrome profiles.The API focus reflects its single provider connection and separated interaction stages, unlike browser traffic with concurrent flows and additional exchanges.Source: §3.2 · p. 4
3.3 MINT Data Processing Stages
- Trace extraction records packet size, interarrival time, and direction, while one-shot conversations are segmented into handshake, prompt, processing, and response stages.The processing interval is measured as time-to-first-response, and the same stage structure was observed in encrypted cloud TLS traffic.Source: §3.3 · p. 4
- Cleaning ignores packets below 50 bytes for TTFR and replaces anomalous IATs with neighboring packet-serialization-scale timing after 11.5% of client packets were found distorted.This addresses small control packets and capture timing artifacts that would otherwise affect application-layer burst measurements.Source: §3.3 · p. 4
- Response packet-size and IAT distributions were repeatable across trials, and provider measurements were similar enough to support modeling on ChatGPT.Streaming varied slightly across providers; Qwen had longer TTFR than ChatGPT, with similar mean packet IAT but smaller variance.Source: §3.3 · p. 4
4 MINT NETWORK TRAFFIC ANALYSIS
MINT analyzes GenAI traffic across upload/download stages and burst timescales, then models those patterns in ns-3. Its measurements reveal modality-dependent behavior, and its simulated burst distributions closely match empirical traffic.
4.1 Upload/Download Asymmetry
- GenAI UL/DL stages vary by modality: prompts tend toward UL-heavy traffic, while response patterns differ from traditional downloads and across text, image, and code tasks.T2I responses have higher DL-heavy throughput; I2T uploads images; code generation includes exchanges and highly variable bursts.Source: §4.1 · pp. 4–5
4.2 Time-scale Burst Modeling Analysis
- AHC identifies packet-serialization, sub-burst, and burst regimes; throughput-based extraction handles Wi-Fi and 5G, where IAT clusters are less separable.The throughput method identifies the largest bursts but cannot recover sub-burst structure; its B-IAT distributions align with Ethernet AHC despite wireless IAT artifacts.Source: §4.2 · pp. 5–6
- GenAI exhibits five burst patterns, including near-constant image uploads, 10 ms median text-streaming B-IAT, delayed partial-image bursts, and varied code-generation exchanges.Non-streaming images arrive in bursts centered around 10 ms; code generation remains a qualitative comparison rather than a modeled workload.Source: §4.2 · p. 6
4.3 Modeling and ns-3 Simulation
- MINT fits burst variables to candidate distributions using AIC while penalizing complexity with BIC, then implements the model as reproducible ns-3 client/server applications.The model includes prompt size, time to first response, and response-burst structure; the implementation exposes distribution settings through JSON configuration.Source: §4.3 · pp. 6–7, Table 3 · p. 7
- 2.3–9% WS for B-IAT across modalities and under 5% for total response size, versus up to 491% for the token baseline; T2I rank-2 sub-bursts reach 24.3%.The reported exception is that the T2I streaming model’s rank-2 sub-bursts capture bimodality but not the full distribution.Source: Table 4 · p. 7, §4.3 · p. 7
5 CONCLUSION
MINT presents an open-source, measurement-driven GenAI traffic model whose ns-3 simulations reproduce measured burst distributions. Its current scope is single-shot text and image API interactions, with several other workloads left for future work.
5 CONCLUSION
- Across modalities, simulated burst distributions matched empirical measurements within 2–25% normalized Wasserstein distance, while constant-rate token streaming did not capture packet timing variability.MINT is presented as the first measurement-driven network-simulator model of GenAI traffic.Source: §5 · p. 8
- MINT currently models single-shot text and image API interactions, not browser/app traffic, sequential prompting with human think time, or automated agentic workloads.The authors infer application-stage boundaries from encrypted packet-level measurements and plan to validate them with decrypted measurements.Source: §5 · p. 8
- The authors plan to extend MINT to browser/app and agentic multi-turn workloads and to evaluate network optimization techniques under simulated GenAI workloads.Source: §5 · p. 8
6 APPENDIX
The appendix figures compare interarrival timing, response delay, and model behavior across GenAI and traditional applications.
6 APPENDIX
- Figure 9 presents browser-versus-API interarrival-time bursts.Source: Figure 9 · p. 10
- Figure 10 compares interarrival times for traditional and GenAI applications.Source: Figure 10 · p. 10
- Figure 11 shows the distribution of time to first response across repeated runs.Source: Figure 11 · p. 10
- Figure 12 compares models’ T2T-Stream response interarrival times and T2T-Nonstream time to first response.Source: Figure 12 · p. 10