Queue Selection Strategy
- Start Small: Begin with test jobs to determine actual resource needs
- Match Resources to Requirements: Don't over-request cores or memory you won't use
- Consider Wait Times: Smaller resource requests often have shorter queue times
- Use Shared Queues: For jobs that don't need an entire node
- Verify Compatibility: Ensure your software works with the specific hardware
Pro Tip: Balance resource requests with wait times. Requesting exactly what you need minimizes both waste and delays.
Legacy Queues (login1/login2 access)
28-Core Intel Haswell Nodes
Established, reliable computing platform with AVX2 support and 128 GB memory per node.
| Queue | Duration | Max Nodes | Special Features | Best For |
|---|---|---|---|---|
| debug-28core | 1 hour max | 8 | Quick turnaround | Testing and debugging |
| short-28core | 1-4 hours | 12 | Fast access | Small to medium jobs |
| medium-28core | 4-12 hours | 24 | Min 8 nodes | Medium parallel jobs |
| long-28core | 8-48 hours | 8 | Extended runtime | Long-running simulations |
| extended-28core | 8 hours - 7 days | 2 | Maximum duration | Very long simulations |
| large-28core | 4-8 hours | 80 | Min 24 nodes | Large-scale parallel computing |
GPU Options (login1/login2)
| Queue | GPU Type | GPU Memory | Duration | Best For |
|---|---|---|---|---|
| gpu | 4x K80 | 24GB each | 1-8 hours | GPU computing, basic ML |
| gpu-long | 4x K80 | 24GB each | 8-48 hours | Long GPU computations |
| p100 | 2x Tesla P100 | 16GB each | 1-24 hours | Scientific computing |
| v100 | 2x V100 | 32GB each | 1-24 hours | Deep learning, AI research |
Modern Queues (milan1/milan2 access)
40-Core Intel Skylake Nodes
Modern computing platform with AVX512 support and 192 GB memory per node.
| Queue | Duration | Max Nodes | Shared Access | Best For |
|---|---|---|---|---|
| debug-40core | 1 hour max | 8 | No | Testing and debugging |
| short-40core | 1-4 hours | 8 | No | Standard HPC jobs |
| short-40core-shared | 1-4 hours | 4 | Yes | Efficient resource use for smaller jobs |
| medium-40core | 4-12 hours | 16 | No | Medium-duration parallel jobs |
| long-40core | 8-48 hours | 6 | No | Long simulations |
| long-40core-shared | 8-24 hours | 3 | Yes | Cost-effective long jobs |
| extended-40core | 8 hours - 7 days | 2 | No | Very long computations |
| extended-40core-shared | 8 hours - 3.5 days | 1 | Yes | Extended shared access |
| large-40core | 4-8 hours | 50 | No | Large-scale parallel jobs |
96-Core AMD EPYC Milan Nodes
High-density computing platform with 256 GB memory per node, ideal for massively parallel workloads.
| Queue | Duration | Max Nodes | Shared Access | Best For |
|---|---|---|---|---|
| short-96core | 1-4 hours | 8 | No | High-throughput computing |
| short-96core-shared | 1-4 hours | 4 | Yes | Parallel jobs with moderate resource needs |
| medium-96core | 4-12 hours | 16 | No | Parameter sweeps, Monte Carlo |
| long-96core | 8-48 hours | 6 | No | Long parallel computations |
| long-96core-shared | 8-24 hours | 3 | Yes | Efficient long-running jobs |
| extended-96core | 8 hours - 7 days | 2 | No | Very long parallel simulations |
| extended-96core-shared | 8 hours - 3.5 days | 1 | Yes | Extended shared parallel work |
| large-96core | 4-8 hours | 38 | No | Massive parallel computing |
High-Bandwidth Memory (HBM) Queues
Intel Sapphire Rapids with Revolutionary Memory Architecture
Cutting-edge nodes featuring 384 GB memory (256GB DDR5 + 128GB HBM) with AMX, AVX512, and Intel DL Boost capabilities.
| Queue | Duration | Max Nodes | Special Features | Best For |
|---|---|---|---|---|
| hbm-short-96core | 1-4 hours | 8 | High-bandwidth memory | Memory-intensive applications |
| hbm-medium-96core | 4-12 hours | 16 | Enhanced memory performance | Large dataset analysis |
| hbm-long-96core | 8-48 hours | 6 | 2-4x memory speed improvement | Memory-bound simulations |
| hbm-extended-96core | 8 hours - 7 days | 2 | Maximum duration with HBM | Long memory-intensive jobs |
| hbm-large-96core | 4-8 hours | 38 | Large-scale HBM computing | Massive memory-bound parallel jobs |
| hbm-1tb-long-96core | 8-48 hours | 1 | 1TB memory + 128GB HBM cache | Extremely large datasets in memory |
HBM Advantage: High-bandwidth memory provides 2-4x faster memory access for memory-bound applications, ideal for large-scale simulations and data analytics.
Modern GPU Queues
NVIDIA A100 GPU Nodes
State-of-the-art GPU computing with 4x A100 80GB GPUs and 64 Intel Ice Lake cores (256 GB memory).
| Queue | Duration | Max Nodes | Shared Access | Best For |
|---|---|---|---|---|
| a100 | 1-8 hours | 2 | Yes | AI/ML training, deep learning |
| a100-long | 8-48 hours | 1 | Yes | Extended GPU computations |
| a100-large | 1-8 hours | 4 | Yes | Large-scale GPU parallel computing |
GPU Usage Guidelines: Only use GPU queues for applications that can effectively utilize GPU acceleration. Verify software compatibility with CUDA and ensure proper GPU programming before submission.
Queue Selection Decision Tree
Choose the Right Queue for Your Work
| If You Need... | Consider... | Why |
|---|---|---|
| Quick testing/debugging | debug-* queues | Fastest turnaround, 1-hour limit |
| GPU acceleration | a100, gpu, p100, v100 | Specialized hardware for parallel computing |
| Large memory requirements | hbm-* queues, 96-core nodes | High memory capacity and bandwidth |
| Maximum parallel processing | 96-core AMD nodes | 96 cores per node for throughput computing |
| Cost-effective computing | shared queues | Multiple users per node, shorter wait times |
| Very long simulations | extended-* queues | Up to 7 days runtime |
| Many nodes working together | large-* queues | Access to 16+ nodes simultaneously |
| Proven, stable platform | 28-core Haswell nodes | Mature hardware for production workflows |
Resource Limits and Best Practices
System-Wide Limits
- Maximum simultaneous nodes: 32 (except in large queues)
- Maximum queued jobs per user: 100
- Memory reservation: Small amount reserved for OS (not available to applications)
Optimization Tips
Shared Queue Benefits: Use shared queues for jobs that don't need an entire node. This often results in shorter wait times and more efficient resource utilization.
- Test First: Start with debug or short queues to determine optimal resources
- Right-Size Requests: Don't request more cores or memory than your application can use
- Consider Architecture: Match your software's optimization to the CPU architecture (AVX2 vs AVX512)
- GPU Efficiency: Only use GPU queues for GPU-accelerated applications
- Memory Planning: Use HBM queues for memory-intensive workloads
- Duration Flexibility: If uncertain about runtime, start with longer queues and adjust based on actual performance
Remember: Effective queue selection balances your computational needs with system efficiency. The goal is to get your work done quickly while making the best use of shared resources.