FEDGEN HPC Layout

The FEDGEN HPC comprises of five different kinds of nodes: the login node, the compute nodes, the accelerator nodes, the management nodes and the storage nodes. In a typical workflow, users primarily access FEDGEN HPC by connecting to the login node, where they compile and test code before submitting jobs to the SLURM queue which automatically assigns a compute node(s) to the submitted job. This HPC layout is in the diagram below.

Login Node

The Login node is the only node directly accessible over the internet. Users connect to the login from their personal computers/workstation using ssh before accessing other parts of the FEDGEN HPC Cluster. Users initiate their jobs primarily from the login node.

Compute Nodes

Compute nodes perform the actual computations submitted to the cluster. They are often organised into groupings called partitions. A typical compute node has one, two, or four processor sockets on the motherboard to host a Central Processing Unit (CPU). Modern Processors are made of multiple Physical cores, that can be thought of as independent processing units, some of which can possess a feature to further handle two threads (see multi-threading) simultaneously.

Accelerator Nodes

These are special purpose compute nodes equipped with Graphics Processing Units (GPU) for computation intesive tasks. The GPU usually offloads certain computations off the CPU for tasks not ideal for it.

Storage Nodes

Storage nodes are nodes dedicated to cluster data Storage. It is based on shared filesystem that can be accessed on all nodes across the entire cluster

Management Nodes

HPC Clusters generally require support infrastructure or services which are provided by the management nodes. These nodes run management services like User Directory Services, Domain Name Services, monitoring and reporting applications.