Overview¶
Here you find a description of available LUMI storage systems, what kind of systems they are and on what hardware partitions they are located at.
Important information regarding data lifetime
The data lifetime on LUMI is closely related to the project lifetime.
When a project ends, it is deactivated for compute, meaning you will no longer be able to submit jobs. However, you will still have access to LUMI and the data stored on it for 90 days. You are responsible of the backup any data you may still need within this 90-day period. After the 90 days, the data will be deleted.
If you start a new project within 90 days of your previous project's end and wish to reuse the data, you are responsible for transferring the data from one project to the other.
On LUMI you have access to network file system storage based on the LUMI-P and LUMI-F hardware partitions, as well as to the object storage on LUMI-O. The use of storage is billed according to the billing policy.
Where to store data?¶
Each user has a home directory ($HOME) that can contain up to
20 GB of data. It is intended to store user configuration files and personal
data. The user home directory is purged once the user account expires. However, LUMI, and therefore the user home directory, is not accessible if the user is not a member of an active project, even though the user has still a valid account on LUMI.
Persistent storage intended to share data amongst the members of a project.
You can see this disk area as the project home directory. Typically, this
space is used to share applications and libraries compiled for the project.
The project persistent storage is located at
/project/project_<project-number>. The project persistent directory is
purged once the project expires.
Temporary storage for input, output, or checkpoint data of your application. When running jobs on LUMI, this is the main storage you should use for your disk I/O needs.
You are not supposed to use the scratch space as long-term storage.
A high performance variant of project scratch. Use this when running jobs on LUMI that need very fast disk I/O operations. Note the difference in billing of the project scratch and project flash as detailed on the billing page.
The project flash space is only meant for very short term file storage.
LUMI network file system disk storage areas¶
On LUMI, there are several network-based disk storage areas. An overview is provided in the tables below. Please familiarize yourself with the characteristics of the hardware partitions before using the different storage areas.
| Path | Intended use | Hardware partition used | |
|---|---|---|---|
| User home |
/users/<username> |
User home directory for personal and configuration files |
LUMI-P |
| Project persistent |
/project/<project> |
Project home directory for shared project files |
LUMI-P |
| Project scratch |
/scratch/<project> |
Temporary storage for input, output or checkpoint data |
LUMI-P |
| Project flash |
/flash/<project> |
High performance temporary storage for input and output data |
LUMI-F |
| Quota | Max files | Expandable | Retention | Billing rate |
|
|---|---|---|---|---|---|
| User home |
20 GB | 100k | No | User lifetime | NA |
| Project persistent |
50 GB | 100k | Yes, up to 500GB |
Project lifetime | 1x |
| Project scratch |
50 TB | 2000k | Yes, up to 500TB |
Project lifetime* | 1x |
| Project flash |
2 TB | 1000k | Yes, up to 100TB |
Project lifetime* | 10x |
| Quota | Max buckets |
Max objects per bucket |
Retention | Billing rate |
|
|---|---|---|---|---|---|
| Object storage | 150 TB | 1000 | 500 000 | project lifetime | 0.5x |
Note that, except for the user home directory, data storage is allocated per project. When a storage space is marked as expandable, it means that you can request more space if needed. Please contact the User Support Team to request more storage space.
*) Data retention policies on LUMI
Please remove the files that are no longer needed by your project on a regular basis. If the storage space on LUMI gets too full at some point, automatic cleaning of project scratch and flash might be enabled. In this case information would be sent to LUMI users at least three months in advance.
About the number-of-files quota¶
For reasons related to performance, we are particularly attentive to the number of files present on the parallel file system. A lot of small files negatively impact all users by stressing the file system metadata servers. Therefore, any requests to increase the number-of-files quota will be evaluated carefully by the User Support Team and must be fully justified to be granted.
Examples of requests that will be rejected include:
- You are installing a lot of small files, e.g. using Conda. Please see the Python packages installation guide for an overview of better alternatives.
- The compilation of your application is generating too many files for your home or project directory: you should compile your application from scratch and then install it in your home or project directory. Exceptions can be made if you are developing an application on LUMI and you want to keep the source and object files in the long term.
In general, applications that generate a lot of small files per process are not well suited for LUMI. If you are the developer of such application, you should consider tools like HDF5 or SIONlib.
You can check the memory and file usage quotas of your projects with the following command:
Temporary storage on compute nodes¶
The /tmp directory on the compute nodes resides in memory. The memory used
for /tmp is included in the job memory allocation. If you use /tmp, you
must allocate memory for it to avoid running out of memory.
Data storage on LUMI is provided by the LUMI-P parallel file system hardware partition, the LUMI-F flash based parallel file system hardware partition, and the LUMI-O object storage hardware partition for a total of 117 PB of storage space.