Skip to content

DOCKER_README.md

Latest commit

 

History

History
163 lines (126 loc) · 4.79 KB

DOCKER_README.md

File metadata and controls

163 lines (126 loc) · 4.79 KB

CoSMo PSS Lithops Docker Runtimes

This directory contains two production-ready Dockerfiles for running CoSMo PSS with Lithops.

Dockerfiles

1. Dockerfile.lithops.cpu (Recommended for Most Use Cases)

Best for:

  • AWS Lambda
  • Azure Functions
  • Google Cloud Functions
  • IBM Cloud Functions
  • Cost-sensitive deployments

Features:

  • CPU-only PyTorch (smaller image size: ~2GB)
  • Python 3.10 slim base
  • All required ML dependencies
  • Optimized for serverless FaaS platforms

Build:

docker build -f Dockerfile.lithops.cpu -t cosmo-pss-runtime-cpu .

Expected Performance:

  • ~100ms per page on Lambda (10GB memory)
  • 1.2M pages in 2-4 hours with 500-1000 workers
  • Cost: ~$2,400 for 1.2M pages

2. Dockerfile.lithops.gpu (High Performance)

Best for:

  • AWS Batch with GPU instances (g4dn.xlarge, g4dn.2xlarge)
  • Azure Container Instances with GPU
  • GCP Cloud Run with GPU
  • Performance-critical deployments

Features:

  • NVIDIA CUDA 11.8 + cuDNN 8
  • GPU-enabled PyTorch
  • Ubuntu 22.04 base
  • All required ML dependencies

Build:

docker build -f Dockerfile.lithops.gpu -t cosmo-pss-runtime-gpu .

Requirements:

  • NVIDIA Docker runtime
  • GPU-enabled compute instances

Expected Performance:

  • ~30ms per page on g4dn.xlarge (1x NVIDIA T4)
  • 1.2M pages in 1-2 hours with 100-200 workers
  • Cost: ~$520 for 1.2M pages

Quick Start

AWS Lambda (CPU)

# Build and push
docker build -f Dockerfile.lithops.cpu -t cosmo-pss-runtime-cpu .
aws ecr create-repository --repository-name cosmo-pss-runtime-cpu
docker tag cosmo-pss-runtime-cpu:latest 123456789012.dkr.ecr.us-east-1.amazonaws.com/cosmo-pss-runtime-cpu:latest
docker push 123456789012.dkr.ecr.us-east-1.amazonaws.com/cosmo-pss-runtime-cpu:latest

# Configure lithops
lithops runtime create cosmo-pss-runtime-cpu --backend aws_lambda

AWS Batch (GPU)

# Build and push
docker build -f Dockerfile.lithops.gpu -t cosmo-pss-runtime-gpu .
aws ecr create-repository --repository-name cosmo-pss-runtime-gpu
docker tag cosmo-pss-runtime-gpu:latest 123456789012.dkr.ecr.us-east-1.amazonaws.com/cosmo-pss-runtime-gpu:latest
docker push 123456789012.dkr.ecr.us-east-1.amazonaws.com/cosmo-pss-runtime-gpu:latest

# Configure lithops
lithops runtime create cosmo-pss-runtime-gpu --backend aws_batch

Image Sizes

Dockerfile Base Image Final Size Notes
CPU python:3.10-slim ~2.0 GB Optimized for Lambda
GPU nvidia/cuda:11.8 ~6.5 GB Includes CUDA toolkit

Verification

Test CPU Runtime

docker run --rm cosmo-pss-runtime-cpu python -c "import torch; print(f'PyTorch: {torch.__version__}'); print(f'CUDA available: {torch.cuda.is_available()}')"

Expected output:

PyTorch: 2.1.0+cpu
CUDA available: False

Test GPU Runtime

docker run --rm --gpus all cosmo-pss-runtime-gpu python -c "import torch; print(f'PyTorch: {torch.__version__}'); print(f'CUDA available: {torch.cuda.is_available()}'); print(f'CUDA version: {torch.version.cuda}')"

Expected output:

PyTorch: 2.1.0+cu118
CUDA available: True
CUDA version: 11.8

Cost Comparison (1.2 Million Pages)

Runtime Platform Workers Time Cost
CPU AWS Lambda 500-1000 2-4 hours ~$2,400 ✓ Most economical
GPU AWS Batch g4dn.xlarge 100-200 1-2 hours ~$520 ✓ Fastest
CPU Azure Functions 500-1000 2-4 hours ~$2,200
GPU Azure Container GPU 100-200 1-2 hours ~$480

Recommendations

  1. Start with CPU: Most cost-effective for initial testing and small-to-medium datasets
  2. Use GPU for:
    • Time-critical processing (need results in < 2 hours)
    • Very large datasets (> 5M pages) where time savings justify cost
    • When GPU instances are already available
  3. Production: Use CPU for regular batch processing, GPU for urgent re-processing

Troubleshooting

CPU Docker Build Fails

Error: Unable to find image 'python:3.10-slim'

Solution: Pull base image first: docker pull python:3.10-slim

GPU Docker Build Fails

Error: Unable to find image 'nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04'

Solution: Pull base image first: docker pull nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04

GPU Runtime Shows CUDA Unavailable

CUDA available: False

Solution: Ensure NVIDIA Docker runtime is installed and use --gpus all flag

Further Reading