Contents


Introduction

Welcome to the lustre user guide. This guide details the steps required to access the lustre high speed storage at eRSA.

What is lustre?

The lustre file system is a high speed, parallel file system, specifically designed for high throughput systems like HPC. Data written to the lustre filesystem is striped (or split up) across multiple servers which increases read and write performance. The eRSA lustre storage system has 100TB of available space.

How much space can I use?

A general rule of thumb is 1-2 TB per eRSA account. If you require more, please contact the service desk.

Is the lustre filesystem backed up?

Currently the lustre filesystem is not backed up. Please copy your files to your home or allocation directory if you require backups.

How long can I leave files on the lustre filesystem?

Currently, we are implementing a policy of 120 days from the last access time of a file (atime). However, we will also take into account the amount of free space on the filesystem.

Use cases for lustre

Lustre is excellent at handling large sequential reads and writes. Files above 2 Megabytes in size will benefit greatly from using the lustre filesystem over the existing NFS server. Lustre is not good at reading and writing many small files (below 4 Kilobytes in size). This type of workload should be performed on the local SATA disk of a HPC compute node.

Please note: HPC Jobs that hinder the performance of the lustre filesystem will be suspended from the queue.


Transferring files to and from lustre
Copying files into and out of the lustre filesystem needs to happen from inside your job script. See the Scratch on Tango guide

Lustre performance

eRSA account

Lustre really shines with workloads that generate parallel read or write I/O from multiple processes. 12 processes appears to be the sweet spot.

 

Single Process (2 MB blocks):
Read: 249MB/sec
Write: 296MB/sec

  

12 Processes (2 MB blocks):
Read: 1210MB/sec
Write: 1166MB/sec

 



Did you find this information useful? Share your feedback here.