LLM serving evolves fast
Every new serving mechanism means rewriting or forking a hand-built simulator.
simulator+synthesizer=Simthesizer
Describe a serving mechanism. A coding agent splices it into one shared simulator and validates its fidelity.
One dynamic graph
All serving work is one event graph. New behavior enters as nodes; the engine never changes.
An agent with guardrails
A coding agent implements each request and validates its fidelity.
The framework
From a feature request to a validated simulator
- 01
Specify behavior
task-design - 02
Map to simulator
sim-mapping - 03
Implement feature
sim-dev - 04
Validate fidelity
sim-validation
Corrections return to the stage that owns the error. How the workflow works →
The built-in simulator
A discrete-event simulator built from components
Inputs
Scenario TOML
Workload JSONL
Profile JSONL
Component tree
System
Scheduler
Router
ModelRunner
Model
Engine & cost models
Event graph
ComputeSimulator
NetworkSimulator
Output
Result JSONL
Engine trace
News
Updates
- Simthesizer is accepted to HPCA 2027.
Cite
Citation
@misc{simthesizer,
title = {Simthesizer: An Agent-Driven Simulation Framework for LLM Serving Systems},
author = {Kim, Wonung and Choi, Hyunmin and Kim, Minsu and Cho, Jaehong and Kim, Yeongwook and Park, Jongse},
year = {2026},
eprint = {2608.24650},
archivePrefix = {arXiv}
}