Skip to content

🐋 orca

Orchestration for Research Cloud Access

Visit orcapod.io → | GitHub →

Overview

orca enables Kubernetes-to-AWS burst computing, seamlessly extending on-premises clusters to the cloud. Scale workloads from local infrastructure to AWS when additional capacity is needed without re-architecting applications or workflows.

ORCA = **O**rchestration for **R**esearch **C**loud **A**ccess

Key Features

Kubernetes Integration

  • Virtual Kubelet - Extends K8s to AWS seamlessly
  • Native Workloads - Run existing K8s workloads unchanged
  • Pod Scheduling - Transparent scheduling to cloud resources
  • API Compatibility - Standard Kubernetes API

Burst Computing

  • On-Demand Scaling - Scale to cloud when on-prem full
  • Cost-Efficient - Only pay for what you use, when you need it
  • Automatic Fallback - Return to on-prem when capacity available
  • Hybrid Workloads - Run some pods on-prem, some in AWS

Cloud Backend

  • AWS Batch - Managed batch computing
  • Spot Instances - Cost savings with spot
  • Multiple Architectures - x86, ARM Graviton support
  • EFA Networking - Low-latency MPI for HPC

Workload Support

  • Batch Jobs - Scientific computing, simulations
  • MPI Workloads - Multi-node parallel jobs
  • GPU Workloads - ML training, rendering
  • Containerized - Any container can burst to AWS

Use Cases

HPC Center Burst Capacity

Extend on-premises Slurm or Kubernetes cluster to AWS during peak demand periods.

Research Lab Scaling

Small on-premises cluster with ability to scale to hundreds of nodes in AWS for large experiments.

Cost Optimization

Maintain smaller on-premises infrastructure, burst to cloud for occasional large workloads.

Hybrid Workflows

Keep sensitive data on-premises while running compute-intensive tasks in AWS.

Architecture

graph TB
    Users[Researchers] -->|Submit Jobs| K8s[On-Prem Kubernetes]

    K8s -->|Local Capacity| OnPrem[On-Prem Nodes]
    K8s -->|Need More| VK[Virtual Kubelet<br/>ORCA]

    VK -->|Schedule| Batch[AWS Batch]
    Batch -->|Launch| EC2[EC2 Instances]

    OnPrem -->|Access| NFS[On-Prem Storage]
    EC2 -->|Mount| S3[S3 Storage]

    S3 -.->|Sync| NFS

    style VK fill:#e1e1ff
    style Batch fill:#e1f5ff
    style EC2 fill:#fff4e1

How It Works

1. Install ORCA Virtual Kubelet

ORCA runs as a virtual kubelet provider in your Kubernetes cluster.

# Deploy ORCA to your cluster
kubectl apply -f https://orcapod.io/deploy/latest.yaml

# Configure AWS credentials
kubectl create secret generic orca-aws \
  --from-literal=access-key=$AWS_ACCESS_KEY_ID \
  --from-literal=secret-key=$AWS_SECRET_ACCESS_KEY

2. Label Workloads for Burst

Add labels to pods that can burst to AWS.

apiVersion: batch/v1
kind: Job
metadata:
  name: my-simulation
spec:
  template:
    metadata:
      labels:
        orca.researchcomputing.io/burst: "enabled"
    spec:
      containers:
      - name: simulation
        image: my-simulation:latest
        resources:
          requests:
            cpu: "16"
            memory: "64Gi"

3. Automatic Scheduling

When on-premises capacity is full, ORCA automatically schedules pods to AWS.

# Submit workload
kubectl apply -f simulation-job.yaml

# Check scheduling
kubectl get pods -o wide
# NAME                         NODE
# simulation-1                 on-prem-node-1
# simulation-2                 on-prem-node-2
# simulation-3                 orca-aws-node
# simulation-4                 orca-aws-node

Configuration

Burst Policies

apiVersion: orca.researchcomputing.io/v1
kind: BurstPolicy
metadata:
  name: default-burst
spec:
  # When to burst
  trigger:
    onPremCapacity: 80%  # Burst when on-prem 80% full
    queueWaitTime: 5m     # Or after 5 min in queue

  # Where to burst
  target:
    provider: aws
    region: us-east-1
    instanceTypes:
      - c7a.16xlarge
      - c7i.16xlarge
      - c8g.16xlarge
    spot: true            # Use spot instances

  # Cost controls
  maxCost: 1000          # Max $1000/day in AWS
  maxNodes: 50           # Max 50 burst nodes

Storage Sync

apiVersion: orca.researchcomputing.io/v1
kind: StorageSync
metadata:
  name: research-data
spec:
  source:
    type: nfs
    path: /mnt/research
  destination:
    type: s3
    bucket: research-data-sync
  sync:
    direction: bidirectional
    schedule: "*/15 * * * *"  # Every 15 minutes

Integration with ResearchComputing

Hybrid Workflows

# 1. Development on-premises (lens or cloudworkspaces)
# 2. Small tests on on-prem K8s
# 3. Large production runs burst to AWS via ORCA
# 4. Results stored in S3 and synced back
# 5. Archive with cargoship

ORCA + ATOM

ORCA can target ATOM workloads for architecture-optimized HPC:

spec:
  target:
    provider: atom
    application: geos-chem
    architecture: c7a  # AMD Zen 4 optimized

Account Management

  • Provision AWS accounts via petri
  • Cost tracking per project
  • Budget alerts for burst spending

Getting Started

Prerequisites

  • Kubernetes cluster (1.24+)
  • AWS account
  • kubectl configured

Installation

# Install ORCA CLI
brew install orca

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

# Deploy to cluster
orca install \
  --aws-region us-east-1 \
  --aws-account-id 123456789012

# Verify installation
kubectl get nodes
# NAME                  STATUS   ROLE
# on-prem-node-1       Ready    worker
# on-prem-node-2       Ready    worker
# orca-virtual-node    Ready    virtual

Submit Burst Job

# Create burst-enabled job
cat <<EOF | kubectl apply -f -
apiVersion: batch/v1
kind: Job
metadata:
  name: molecular-dynamics
spec:
  parallelism: 16
  template:
    metadata:
      labels:
        orca.io/burst: "enabled"
        orca.io/priority: "high"
    spec:
      containers:
      - name: md-simulation
        image: my-org/md-sim:latest
        resources:
          requests:
            cpu: "4"
            memory: "16Gi"
      restartPolicy: Never
EOF

# Monitor job
orca job status molecular-dynamics

# Check costs
orca cost summary --job molecular-dynamics

Cost Management

Cost Estimation

# Estimate burst costs
orca estimate \
  --cpus 256 \
  --memory 1024 \
  --duration 4h \
  --spot

# Output:
# Instance type: c7a.16xlarge (spot)
# Instances needed: 4
# Duration: 4 hours
# Estimated cost: $12.80

Budget Controls

apiVersion: orca.researchcomputing.io/v1
kind: Budget
metadata:
  name: lab-budget
spec:
  limit: 5000  # $5000/month
  alerts:
    - threshold: 80%
      notify: lab-admin@university.edu
    - threshold: 100%
      action: pause-burst

Documentation

Technology Stack

  • Language: Go
  • Orchestration: Virtual Kubelet
  • Cloud Provider: AWS (Batch, ECS, EC2)
  • Storage Sync: S3 sync, FSx for Lustre
  • Monitoring: Prometheus, CloudWatch

Project Status

Current Version: v0.6.0 Status: Beta License: Apache 2.0

Contributing

Contributions welcome! See the contribution guide.

Support

← Back to Ecosystem