Distributed Training Infrastructure for Hierarchical Cross-View Geo-Localization
Distributed Systems Theory and Analysis Course Project · Missouri S&T · Spring 2026
This repository implements and rigorously evaluates the distributed PyTorch DDP training pipeline for Coarse Retrieval Stage of the HCVGLoc framework -- a hierarchical cross-view geo-localization system for GPS-denied UAV navigation.
📄 Read project details — full project report (PDF)
The project treats the 4× NVIDIA RTX Ada 6000 training cluster as a distributed system:
| Metrics | What is Measured |
|---|---|
| Scalability | Throughput (img/s), speedup, parallel efficiency across 1/2/4 GPUs |
| Communication Overhead | AllReduce latency, compute-to-comm ratio, bandwidth sensitivity |
| Fault Tolerance | Mid-training failure recovery, checkpoint overhead vs. recovery time |
| Convergence Parity | Single-GPU vs DDP: loss curves, R@1/R@5, wall-clock time |
Stage 1 performs coarse satellite image retrieval: given a UAV query image, retrieve the top-K matching geo-referenced satellite tiles from a large-scale gallery.
Datasets: University-1652 · VIGOR · CVUSA · CVACT
Hardware: 4× NVIDIA RTX Ada 6000 (48 GB each), NCCL backend
Related repo: gxanesh/avgl_plus_plus
git clone https://github.com/gxanesh/distributed-HCVGLoc.git
cd distributed-HCVGLoc
bash scripts/setup_env.sh
# Single GPU baseline
bash scripts/run_single_gpu.sh
# 4-GPU DDP
bash scripts/run_ddp_4gpu.sh
# Scaling benchmark
python benchmarks/scaling_benchmark.py --gpus 1 2 4 --epochs 3
# Communication profiler
torchrun --nproc_per_node=4 benchmarks/communication_profiler.py
# Fault tolerance
python fault_tolerance/checkpoint_manager.py --benchmarkMIT — see LICENSE