Falcon HPC Cluster

Description

The Falcon cluster at the Computer Science Department is an HPC cluster which supports both CPU and GPU jobs.
The hardware specifications are listed as follows:

  • 4 x AMD EPYC 74F3 CPUs (192 cores)
  • 2Tb RAM
  • 8 Nvidia A100 80Gb GPUs (640Gb total GPU memory)
System overview

The Falcon Cluster can only be reached from within the CSU network or via CSU’s VPN from off-campus. The login nodes serve as the user’s entry point and are where jobs are prepared and submitted. SLURM is used to schedule and manage computational tasks on the compute nodes. The CSU Computer Science network file system is accessible to all compute nodes and login nodes.

image

Cluster Nodes

Here is a brief description of the types of nodes:

Login nodes

Users work on these nodes directly. This is where users log in, access files and submit jobs to the scheduler to be run on the compute nodes. Furthermore, these are the only nodes that can be accessed via SSH/SCP from outside of the cluster.

Compute nodes

These nodes are where the actual computing gets done. Users do not log into these nodes directly. Instead jobs are created and submitted from the login nodes and a scheduler decides which compute resources are available and when to run the job.

Management node

There is one management node, which is reserved for administration of the cluster. It is not available to users.

Queueing System

On Falcon, we use the open-source scheduler Slurm, to manage all compute jobs. The process is described in detail in the Job handling section.

Storage

The following storage locations are available on the Falcon cluster.

CS NFS Space

The cluster has read/write access to the same NFS shares as any other CS Linux system

Scratch Space

All compute nodes have a local /scratch directory on a high speed SSD, and it is generally about a tera-byte in size. This is used by the scheduler to store temporary job files. Users have access to this folder only for the duration of their job. After jobs are completed, this space is purged automatically. This space is shared between all users, but your data is accessible only to you.