The computing cluster is designed to manage compute resources efficiently when many users are competing for the same resources. It also allows for centralisation of the resources, so that individual users do not need a strong machine, which would rarely run at full capacity. Instead of running a job directly/locally, users submit jobs to the scheduler, which then allocate resources based on the request and the availability. The resources are allocated to the jobs individually and exclusively, which minimises the interference between them. If the requested resources are available, the scheduler runs them on a worker node. If not, they are held in a queue.

The scheduler is configured in a way that prevents overloading the worker nodes. For example, a worker node with 32GB memory can run up to 16 jobs that each requires 2GB memory, but only 8 jobs that requires 4GB memory each. For maximum efficiency, it is therefore essential that users have an a priori knowledge about the resources required by the jobs. Requesting excessive amount of resources leads to that fewer jobs will run at the same time, while requesting too little leads to that job may crash or become unresponsive.

You can read more about computing clusters here.

Usefull presentations

Title Presenter Notes Slides
Introduction to RHUL-Psychology computing system Tibor Auer This talk describes of the computing architecture PDF
Using the RHUL-Psychology scheduling system Tibor Auer This talk describes some basic use cases with examples PDF

List of authorized users (for cluster managers only)

Cluster resources

Cluster configuration (e.g. queues)

How to access

Using the cluster