Skip to content
This repository was archived by the owner on Aug 13, 2025. It is now read-only.

maltempi/msc_dissertation

Repository files navigation

GPUZIP: A Checkpointing Optimization Library for Adjoint-Mode Applications

This repository contains the MSc dissertation by Thiago Maltempi titled "GPUZIP: A checkpointing optimization library for adjoint-mode applications," completed at the Institute of Computing, UNICAMP, Brazil.

Abstract

Adjoint-mode applications, such as seismic Reverse Time Migration (RTM), require significant computational resources, particularly when running on GPUs. The checkpoint-recompute trade-off is a critical aspect of these applications, where optimizing the storage and retrieval of intermediate states during computation can significantly improve performance.

GPUZIP addresses this challenge by combining two key techniques:

  1. Prefetching: Proactively schedules asynchronous data transfers from host memory to GPU memory, reducing wait times.
  2. Compression: Applies lossy compression techniques (cuZFP, NVIDIA Bitcomp) to reduce the data transfer volume.

Together, these approaches substantially accelerate checkpointing operations in GPU-based adjoint computations, with demonstrated speedups of up to 8.9× in real-world seismic imaging applications.

Key Features

  • Compatible with multiple checkpointing algorithms (Revolve, zCut, Uniform)
  • NVIDIA CUDA-based implementation for GPU acceleration
  • Asynchronous prefetching mechanism to hide data transfer latency
  • Integration with lossy compression libraries
  • Application-specific optimizations for seismic RTM workflows

Contents

This repository includes:

  • Full dissertation text in LaTeX format (ic-tese-v3.tex)
  • Final PDF version (output/Thiago_Maltempi__Dissertação_Mestrado__FINALFINALFINAL.pdf)
  • Defense Presentation PDF and PPT version (Portuguese only) (output/PRESENTATION__ThiagoMaltempi__Defesa__Mestrado.pdf, output/PRESENTATION__ThiagoMaltempi__Defesa__Mestrado.pptx)
  • Bibliography (ic-tese-v3.bib)
  • Figures and diagrams illustrating the architecture and results

Key Findings

  • The prefetching mechanism alone achieves up to 1.7× speedup in checkpoint retrieval
  • Compression techniques alone achieve up to 5.4× speedup
  • The combined GPUZIP approach achieves up to 8.9× speedup in overall application performance
  • Minimal impact on result quality when using carefully tuned lossy compression parameters

Publications

This research has been presented in several venues:

  • Maltempi, T., Rigo, S., Pereira, M., Yviquel, H., Leite, G., Lee, O., Costa, J., & Araujo, G. (2025). Checkpointing fine-tuning for accelerating seismic applications in GPUs. The International Journal of High Performance Computing Applications. DOI: 10.1177/10943420251340794

  • Maltempi, T., Rigo, S., Pereira, M., Yviquel, H., Costa, J., & Araujo, G. (2024). [Euro-Par Conference Publication]

  • Maltempi, T., Rigo, S., Pereira, M., Yviquel, H., Costa, J., Souza, A., & Araujo, G. (2023). Introducing Prefetching and Data Compression to Accelerate Checkpointing for Inverse Seismic Problems. Poster presented at Super Computing 2023 (SC'23)

About

MSC dissertation assets

Resources

Stars

Watchers

Forks

Languages