⚛️ atom¶

Automated Toolkit for Optimized Modeling

Cloud-native HPC made simple

Visit atomhpc.io → | GitHub →

Overview¶

atom is a cloud-native high-performance computing platform for running scientific applications with architecture-optimized containers and cost-efficient job scheduling. Built from the ground up for modern research computing without the complexity of traditional HPC cluster management.

The fundamental unit of cloud computing

Key Features¶

Architecture-Optimized¶

• AMD EPYC - Zen 2, Zen 3, Zen 4 optimized builds
• Intel Xeon - Cascade Lake, Ice Lake, Sapphire Rapids
• ARM Graviton - Graviton 2, Graviton 3, Graviton 4
• Processor-Specific - BLIS, MKL, ARM Performance Libraries

Application-Agnostic Platform¶

• Pluggable Applications - Add new apps without platform changes
• Layered Containers - Shared base images, fast app builds
• Version Independence - Platform and apps version separately
• Multi-Application - GEOS-Chem, Gaussian, WRF, VASP, ORCA

Cost-Optimized¶

• Spot Instances - Up to 90% savings vs on-demand
• Priority Queues - Balance cost and urgency
• Architecture Selection - Automatic best price/performance
• Cost Estimation - Know costs before running

Job Management¶

• AWS Batch - Managed job scheduling
• Single-Node & MPI - OpenMP and distributed workloads
• Retry Logic - Automatic spot interruption handling
• Monitoring - CloudWatch integration, job telemetry

Supported Applications¶

Architecture¶

graph TB
    User[Researcher] -->|Submit| CLI[atom CLI]
    CLI -->|Create Job| Batch[AWS Batch]

    Batch -->|Single-Node| EC2Single[EC2 Instance]
    Batch -->|Multi-Node| EC2Multi[ECS Cluster]

    EC2Single -->|Pull| ECR[Container Registry]
    EC2Multi -->|Pull| ECR

    ECR -->|Base Image| BaseAMD[AMD Zen4 Base]
    ECR -->|Base Image| BaseIntel[Intel SPR Base]
    ECR -->|Base Image| BaseARM[Graviton4 Base]

    EC2Single -->|Input| S3[S3 Data]
    EC2Single -->|Output| S3
    EC2Multi -->|Input| S3
    EC2Multi -->|Output| S3

    style CLI fill:#e1ffe1
    style Batch fill:#e1f5ff

Container Strategy¶

Layered Architecture¶

Base Images (maintained separately):
├── hpc-base-amd-zen4:latest
│   └── GCC, OpenMPI, AMD AOCL (BLIS + libFLAME)
├── hpc-base-intel-spr:latest
│   └── GCC, OpenMPI, Intel MKL
└── hpc-base-arm-graviton4:latest
    └── GCC, OpenMPI, ARM Performance Libraries

Application Layers (versioned independently):
├── geos-chem:14.4.3-c7a
│   └── FROM hpc-base-amd-zen4:latest
├── geos-chem:14.4.3-c7i
│   └── FROM hpc-base-intel-spr:latest
└── gaussian:16-c7a
    └── FROM hpc-base-amd-zen4:latest

Benefits: - Base images shared across applications - Faster application builds (no recompiling libraries) - Version applications independently from infrastructure - Security updates to base propagate to all apps

Application Types¶

Classic (Single-Node, OpenMP)¶

Applications that run on a single EC2 instance with shared memory parallelism.

Examples: GEOS-Chem Classic, Gaussian, ORCA, NWChem (single-node)

Instance Families: c7a (AMD Zen 4), c7i (Intel Sapphire Rapids), c8g (Graviton 4)

Best For: Small to medium workloads, memory-intensive applications

MPI (Multi-Node, Distributed)¶

Applications that run across multiple EC2 instances with message passing.

Examples: GEOS-Chem High Performance (GCHP), WRF, VASP, NWChem (MPI)

Networking: EFA (Elastic Fabric Adapter) for low latency

Best For: Large-scale simulations, domain decomposition workloads

Getting Started¶

Install CLI¶

# Install via Homebrew
brew install atom

# Or download from releases
curl -LO https://github.com/scttfrdmn/atom/releases/latest/download/atom-darwin-amd64
chmod +x atom-darwin-amd64
sudo mv atom-darwin-amd64 /usr/local/bin/atom

Deploy an Application¶

# Deploy GEOS-Chem application
atom app deploy geos-chem --env benchmark

# List deployed applications
atom app list

# View application details
atom app describe geos-chem

Submit a Job¶

# Submit GEOS-Chem Classic job
atom job submit geos-chem \
  --config benchmark.yaml \
  --architecture c7a \
  --input s3://my-bucket/input/ \
  --output s3://my-bucket/output/

# Submit GCHP (MPI) job
atom job submit gchp \
  --config c48.yaml \
  --nodes 6 \
  --architecture c7a \
  --input s3://my-bucket/input/ \
  --output s3://my-bucket/output/

# Monitor job
atom job status <job-id>

# View logs
atom job logs <job-id>

# Get job costs
atom job cost <job-id>

Use Cases¶

Atmospheric Modeling¶

Run GEOS-Chem Classic and GCHP simulations with architecture-optimized builds for 2-3x performance improvement.

Quantum Chemistry¶

Execute Gaussian and ORCA calculations with cost-efficient spot instances and automatic retry on interruption.

Weather Modeling¶

Deploy WRF simulations across multiple nodes with EFA networking for low-latency MPI communication.

Materials Science¶

Run VASP calculations on cost-optimized instances with AMD or Graviton processors.

Integration with ResearchComputing¶

Development Workflow¶

1. Develop and test in lens or cloudworkspaces
2. Submit production jobs to atom
3. Archive results with cargoship

Account Management¶

• Provision via petri
• Cost tracking per project
• Budget alerts

Hybrid Computing¶

• Use orca to burst from on-premises Kubernetes to atom workloads

Performance¶

Architecture-optimized containers deliver significant performance improvements:

Benchmarks vs generic x86_64 builds without processor-specific optimization

Cost Optimization¶

Spot Instances¶

• Up to 90% savings vs on-demand
• Automatic retry on interruption
• Checkpointing for long jobs

Priority Queues¶

• Urgent - On-demand instances, immediate start
• Standard - Spot with fallback, cost-balanced
• Economy - Spot only, maximum savings

Architecture Selection¶

atom automatically selects the most cost-effective architecture for your workload based on current spot prices.

Documentation¶

• Platform Architecture
• Adding Applications
• Container Build System
• Job Submission Guide
• Application: GEOS-Chem

Technology Stack¶

• Language: Go (CLI, platform), Bash (containers)
• Infrastructure: AWS Batch, ECS, CloudFormation/CDK
• Containers: Docker, Amazon ECR
• Networking: VPC, EFA (for MPI)
• Storage: S3, EBS, EFS (optional)
• Compilers: GCC, Intel oneAPI
• MPI: OpenMPI
• Math Libraries: BLIS, libFLAME, MKL, ARM PL

Project Status¶

Current Version: v0.1.0-alpha Platform API: v1.0.0-dev Status: 🚧 Active Development

The platform is being extracted from aws-geos-chem to create a reusable foundation for multiple scientific applications.

Target Release: Q1 2026

Contributing¶

Contributions welcome! See the contribution guide.

Interested in adding your application? Check out the Adding Applications Guide.

Support¶

• Documentation: atomhpc.io
• Issues: GitHub Issues
• Discussions: GitHub Discussions

← Back to Ecosystem