Cluster (HPC Cluster)

A cluster is a set of computers linked together so that their resources can be pooled and applied collectively to computational jobs, making many machines behave like one larger system. The term is not exclusive to HPC, a Kubernetes cluster, for example, is a group of computers configured to run Kubernetes together, but in high performance computing it usually refers to a Beowulf-style cluster built to run computational jobs. In everyday use, clusters of any kind are simply called a "cluster" without a modifier.

What an HPC cluster is

A modern HPC cluster is a Beowulf-style cluster designed specifically to run computational jobs. It can be built as a general-purpose system supporting a wide variety of workloads, or configured with more specialized components to target particular workloads such as AI training, molecular dynamics, or computational fluid dynamics. Either way, the defining idea is the same: commodity servers joined by a fast network and coordinated by software so they can tackle problems no single node could handle alone. Terms like "HPC cluster" and "supercomputer cluster" refer to the same concept.

How a cluster is put together

A cluster is composed of several coordinated parts:

• Compute nodes, The individual servers that do the actual work; see compute nodes.
• Interconnect, A high-speed, low-latency network that lets nodes exchange data; see HPC interconnect.
• Shared storage, Often a parallel file system that all nodes can read from and write to.
• Head and management nodes, Systems that handle login, scheduling, and provisioning of the compute nodes.
• Scheduler, Software that queues jobs and assigns them to available nodes.

These pieces map directly onto the HPC architecture that governs how a cluster is designed, and their balance is captured by the HPC triangle of compute, storage, and networking.

Why clusters matter

Clusters make HPC economically practical. Rather than buying a single monolithic machine, organizations combine off-the-shelf hardware into a system that scales by adding nodes, an approach known as scaling out. This uses parallel computing to divide a workload across many processors that run simultaneously, delivering throughput far beyond a single server while keeping cost and flexibility manageable.

Cluster versus a single large server

	Cluster	Single large server
Growth model	Scale out by adding nodes	Scale up one machine
Cost	Commodity hardware	Premium components
Fault tolerance	Loss of a node is survivable	Single point of failure
Best for	Parallel HPC and AI workloads	Workloads that fit one box

A cluster trades the simplicity of one machine for scalability and resilience, which is why it is the dominant architecture for HPC.