Taylor Geospatial Institute Regional AI Learning System (TGI RAILS) User Guide

System status, planned outages, and maintenance information:

Introduction

TGI RAILS is a regional computing system funded by the NSF Campus Cyberinfrastructure program for primary use by the TGI consortium for geospatial related research. RAILS provides a highly capable GPU-focused compute environment for GPU and CPU workloads.

RAILS has three CPU-only nodes and three 8-way NVIDIA H100-based GPU nodes. Every RAILS node has high-performance node-local SSD storage (1.92 TB for CPU nodes, 3.5 TB for GPU nodes), and is connected to the 1 PB VAST filesystem via the high-speed interconnect. The RAILS resource uses the SLURM workload manager for job scheduling.

Contents

• TGI RAILS Status Updates
• System Architecture
  • Compute Nodes
  • Login Nodes
  • Specialized Nodes
  • Network
  • File Systems
• Allocation and Account Administration
  • Requesting an Allocation
  • Managing Users
  • Account and Group
  • Management Tools
  • Configuring Your Account
  • Allocation Policies
• Accessing The System
  • Direct Access login_nodes
  • maintaining persistent sessions: tmux
  • ssh keyboard-interactive
  • Compute Node External Connectivity
• Good Cluster Citizenship
• Data Management
  • File Systems
  • Transferring Data
• Accessing and Transferring Files
  • Transferring a Few Small Files
  • Transferring Many or Large Files With Globus
  • Sharing Files with Collaborators
  • Access Controls
  • Quotas and Policies
• Programming Environment (Building Software)
  • Modules
  • Compiling
  • Serial Programs
  • MPI Programs
  • Python
  • OpenMP Programs
  • Hybrid MPI/OpenMP Programs
  • OpenACC Programs
  • CUDA
• Running Jobs
  • Accounting
  • Performance tools
  • Sample Scripts
  • Interactive Sessions
  • Job Management
  • Job Status
  • Accessing the Compute Nodes
  • Scheduler
  • Partitions (Queues)
  • Job Policies
  • Monitoring a Node During a Job
  • Refunds
• Installed Software
  • modules/lmod
  • Python
  • R
• Visualization
  • Paraview (Client/Server)
  • Visit Client & Server
• Containers
• Services
  • Open OnDemand
  • Globus
• Debugging And Performance Analysis
  • Intel ?? Guide
  • NVIDIA Nsight Systems
• Protected Data
• Acknowledging TGI Rails
• Getting Help