REPACSS Architecture¶
This document provides a comprehensive overview of the REPACSS (REmotely-managed Power Aware Computing Systems and Services) high-performance computing (HPC) infrastructure at Texas Tech University. The system is designed to support computationally intensive and data-driven research and is powered by renewable energy sources. REPACSS supports a wide range of scientific workloads, including simulations, artificial intelligence (AI), and large-scale data analytics.
Access Requirements¶
To utilize REPACSS resources, users must meet the following prerequisites:
- A valid TTU eRaider account (for current users)
- Access to the GlobalProtect VPN
- Basic familiarity with HPC concepts
- Competence with Linux command-line environments
Compute Resources¶
REPACSS consists of a heterogeneous cluster of compute resources optimized for both CPU-intensive and GPU-accelerated workloads.
CPU Compute Nodes¶
- Total Nodes: 110
- Processor: Dual AMD EPYC 9754
- Cores per Node: 256
- Memory per Node: 1.5 TB DDR5
- Storage per Node: 1.92 TB NVMe SSD
GPU Compute Nodes¶
- Total Nodes: 8
- Processor: Dual Intel Xeon Gold 6448Y
- Cores per Node: 64
- Memory per Node: 512 GB
- GPUs per Node: 4 × NVIDIA H100 NVL (94 GB HBM per GPU)
- Storage per Node: 1.92 TB SSD
Login Nodes¶
- Total Nodes: 3
- Processor: Dual AMD EPYC 9254
- Cores per Node: 48
- Memory per Node: 256 GB
- Storage per Node: 1.92 TB NVMe SSD
Storage Infrastructure¶
REPACSS includes nine storage nodes that collectively offer approximately 2.94 PB of multi-tiered storage. These systems utilize both high-speed NVMe drives and large-capacity HDDs to accommodate various storage requirements.
Storage Node Summary¶
- Processors: Intel Xeon Gold (varied models)
- Cores per Node: 8 to 32
- Memory per Node: 512 GB to 1 TB
- Storage per Node: 25.6 TB to 583.68 TB
Storage Tiers¶
- Home: Persistent personal storage for user scripts and configuration files.
- Scratch: High-performance temporary storage space subject to periodic purging.
- Work: Long-term storage for research outputs and work purposes.
Network Architecture¶
The network backbone of REPACSS supports high-throughput and low-latency communication among compute and storage nodes. * Switching Infrastructure: Dell PowerSwitch S5248-ON and S5232-ON * Storage Interconnect: InfiniBand network * Features: * Reliable and fast data movement * Efficient support for multi-node parallel processing (MPI) * Secure remote access capabilities
System Specifications Summary¶
| Node Type | Total Nodes | CPU Model | CPUs/Node | Cores/Node | Memory/Node | Storage/Node | GPUs/Node | GPU Model |
|---|---|---|---|---|---|---|---|---|
| CPU Nodes | 110 | AMD EPYC 9754 | 2 | 256 | 1.5 TB DDR5 | 1.92 TB NVMe | - | - |
| GPU Nodes | 8 | Intel Xeon Gold 6448Y | 2 | 64 | 512 GB | 1.92 TB SSD | 4 | NVIDIA H100 NVL (94 GB HBM) |
| Login Nodes | 3 | AMD EPYC 9254 | 2 | 48 | 256 GB | 1.92 TB NVMe | - | - |
| Storage Nodes | 9 | Intel Xeon Gold (varied) | 2 | 8–32 | 512 GB–1 TB | 25.6–583.68 TB | - | - |