Synthesizing Execution Traces with Graph-Based Generative Models
Author(s)
Shitole, Viraj
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The objective of this research is to investigate the use of graph-based generative models for synthesizing execution traces of machine learning workloads. Specifically, this work is guided by the following research question:
Can we develop a generative framework that produces synthetic execution traces which accurately mimic proprietary machine learning workloads, while preserving key performance characteristics such as latency and memory usage?
Execution traces are indispensable for benchmarking, system simulation, and hardware–software co-design; however, they also encode sensitive information that may reveal proprietary model architectures or training methodologies. Synthesizing realistic yet privacy-preserving traces thus presents a critical opportunity to enable secure data sharing and collaborative optimization at scale.
To address this challenge, we model execution traces as graphs and apply generative diffusion-based techniques to capture both structural and performance-related properties. The main contributions are threefold: (1) formalizing execution trace synthesis as a graph generation problem, (2) developing a generative framework that balances fidelity and privacy, and (3) performing comprehensive evaluations that quantify structural realism, performance relevance, and resistance to model-identification attacks.
By demonstrating that graph generative models can produce high-fidelity, privacy preserving execution traces, we advance the broader goal of secure, data-driven co-design for next-generation machine learning infrastructure.
Sponsor
Date
2025-12
Extent
Resource Type
Text
Resource Subtype
Thesis (Masters Degree)