Title

A Diffusion-Based Framework for Configurable and Realistic Storage Trace Generation

Alternative Title

구성 가능하고 현실적인 스토리지 트레이스 생성을 위한 확산 기반 프레임워크

DGIST Authors

Seohyun Kim ; Yeseong Kim

Advisor

김예성

Issued Date

2026

Awarded Date

2026-02-01

Type

Thesis

Description

Diffusion Models, Storage Systems, Block I/O Traces, Synthetic Data, System Benchmark

Table Of Contents

Ⅰ. Introduction 1
ⅠI. Background and Motivation 4
2.1 Lack of Real-world Traces 4
2.2 Limitations of Traditional Trace Generation Tools 5
2.3 Feasibility of Probabilistic Modeling 6
2.4 Limitations of Configuration-Based Simulation 8
2.5 Generative Models for Trace Generation 9
IⅠI. Overview of DiTTO 11
IV. Overview of STAMP 13
V. Detailed Method of STAMP 15
5.1 Data Pre-processing 15
5.1.1 Hierarchical Clustered Address Encoding 15
5.1.2 Uniform Feature Embedding 16
5.1.3 Hyper-configuration Labeling 17
5.2 Latent Representation Learning 18
5.3 CHIP : Contrastive Hyper-configuration and I/O trace Pre-training 22
5.4 Configuration-Guided Latent Denoising with U-Net 24
5.5 Inference 26
VI. Evaluation 28
6.1 Experimental Setup 28
6.2 Evaluation of DiTTO 29
6.3 Evaluation of STAMP 30
VII. Future Work 32
VIII. Conclusion 34

URI

https://scholar.dgist.ac.kr/handle/20.500.11750/59723
http://dgist.dcollection.net/common/orgView/200000942224

DOI

10.22677/THESIS.200000942224

Degree

Master

Department

Artificial Intelligence Major

Publisher

DGIST

Show Full Item Record