workflows

ec2-gha Demos

This directory contains the reusable workflow and demo workflows for ec2-gha, demonstrating various capabilities.

For documentation about the main workflow, runner.yml, see the main README.

• demos – run all demo workflows
• Core demos
  • dbg-minimal – configurable debugging instance
  • gpu-minimal – nvidia-smi "hello world"
  • cpu-sweep – OS/architecture matrix
  • gpu-sweep – GPU instance types with PyTorch
• Parallelization
  • instances-mtx – multiple instances for parallel jobs
  • runners-mtx – multiple runners on single instance
  • jobs-split – different job types on separate instances
• Stress testing
  • test-disk-full – disk-full scenario testing
• Real-world example: Mamba installation testing

demos – run all demo workflows

Useful regression test, demonstrates and verifies features.

Core demos

dbg-minimal – configurable debugging instance

• workflow_dispatch with customizable parameters (instance type, AMI, timeouts)
• Also callable via workflow_call (used by cpu-sweep)
• Extended debug mode for troubleshooting
• Instance type: t3.large (default), configurable
• Use case: Interactive debugging and testing

gpu-minimal – nvidia-smi "hello world"

• Instance type: g4dn.xlarge

cpu-sweep – OS/architecture matrix

• Tests 12 combinations across operating systems and architectures
• OS: Ubuntu 22.04/24.04, Debian 11/12, AL2, AL2023
• Architectures: x86 (t3.*) and ARM (t4g.*)
• Calls dbg-minimal for each combination
• Use case: Cross-platform compatibility testing

gpu-sweep – GPU instance types with PyTorch

• Tests different GPU instance families
• Instance types: g4dn.xlarge, g5.xlarge, g6.xlarge, g5g.xlarge (ARM64 + GPU)
• Uses Deep Learning OSS PyTorch 2.5.1 AMIs
• Activates conda environment and runs PyTorch CUDA tests
• Use case: GPU compatibility and performance testing

Parallelization

instances-mtx – multiple instances for parallel jobs

• Creates configurable number of instances (default: 3)
• Uses matrix strategy to run jobs in parallel
• Each job runs on its own EC2 instance
• Instance type: t3.medium
• Use case: Parallel test execution, distributed builds

runners-mtx – multiple runners on single instance

• Configurable runners per instance (default: 3)
• All runners share the same instance resources
• Demonstrates resource-efficient parallel execution
• Instance type: t3.xlarge (larger instance for multiple runners)
• Use case: Shared environment testing, resource optimization

jobs-split – different job types on separate instances

• Launches 2 instances
• Build job runs on first instance
• Test job runs on second instance
• Demonstrates targeted job placement
• Instance type: t3.medium
• Use case: Pipeline with dedicated instances per stage

Stress testing

test-disk-full – disk-full scenario testing

• Tests runner behavior when disk space is exhausted
• Configurable parameters:
  • disk_size: Root disk size (0=AMI default, +N=AMI+N GB, e.g., +2)
  • fill_strategy: How to fill disk (gradual, immediate, or during-tests)
  • debug: Debug mode (false, true/trace, or number for trace+sleep)
  • max_instance_lifetime: Maximum lifetime before forced shutdown (default: 15 minutes)
• Features tested:
  • Heartbeat mechanism for detecting stuck jobs
  • Stale job file detection and cleanup
  • Worker/Listener process monitoring
  • Robust shutdown with multiple fallback methods
• Instance type: t3.medium (default)
• Use case: Verifying robustness in resource-constrained environments

Real-world example: Mamba installation testing

• Tests different versions of mamba_ssm package on GPU instances
• Customizes instance_name: "$repo/$name==${{ inputs.mamba_version }} (#$run)"
  • Results in descriptive names like "mamba/install==2.2.5 (#123)"
  • Makes it easy to identify which version is being tested on each instance
• Uses pre-installed PyTorch from DLAMI conda environment
• Use case: Package compatibility testing across versions