Optimizing Cloud-to-GPU Throughput for Deep Learning With Earth Observation Data

TL;DR

This paper enhances GPU utilization and data loading efficiency for deep learning on Earth observation data by optimizing GeoTIFF streaming from cloud storage, achieving significant throughput improvements and maintaining model accuracy.

Contribution

It introduces a systematic benchmarking and Bayesian optimization approach to improve GeoTIFF data loading from cloud storage, enabling high GPU utilization and faster training.

Findings

Remote data loading throughput increased by 20x

Local throughput increased by 4x

Model accuracy maintained with 85-95% GPU utilization

Abstract

Training deep learning models on petabyte-scale Earth observation (EO) data requires separating compute resources from data storage. However, standard PyTorch data loaders cannot keep modern GPUs utilized when streaming GeoTIFF files directly from cloud storage. In this work, we benchmark GeoTIFF loading throughput from both cloud object storage and local SSD, systematically testing different loader configurations and data parameters. We focus on tile-aligned reads and worker thread pools, using Bayesian optimization to find optimal settings for each storage type. Our optimized configurations increase remote data loading throughput by 20x and local throughput by 4x compared to default settings. On three public EO benchmarks, models trained with optimized remote loading achieve the same accuracy as local training within identical time budgets. We improve validation IoU by 6-15% and maintain 85-95% GPU utilization versus 0-30% with standard configurations. Code is publicly available at https://github.com/microsoft/pytorch-cloud-geotiff-optimization