5.1. Understanding Storage Options#
There are several storage options available on the Kempner Institute HPC cluster. Each storage option has its own unique features and is designed to meet different requirements. This document provides an overview of the storage options available on the cluster and their key concepts.
5.1.1. Default persistent home directory#
Each user has 100GB of persistent storage that is their home directory. This is for long-term storage of files, checkpoints, datasets, etc. Your home directory is located at:
/n/home<number>/<your user name>
There are 15 different home numbers, and your home directory will be allocated to one of them. For example, Jonathan Frankle’s home directory is at:
/n/home09/jfrankle
This storage is only accessible only to you. There is no cost associated with this storage.
Tip
Check your home directory usage while on login node by running the following command:
df -h ~/
Read more about how to manage a full home directory situation here.
5.1.2. Default persistent lab directory#
Each lab has 4 TB of persistent storage. This is for long-term storage of files, checkpoints, datasets, etc. Your lab directory is located at:
/n/holylabs/LABS/<your lab name>
For example, Jonathan Frankle’s lab directory is at:
/n/holylabs/LABS/jfrankle_lab
This storage is only accessible to members of the lab. There is no cost associated with this storage.
5.1.3. Temporary scratch storage (Lustre)#
Each lab has 50 TB of scratch
storage space. This storage is high-performance, and it is intended to be where you keep data you are actively using for a job (e.g., datasets you’re actively using for a job, checkpoints you’re storing from a job, etc.). That data should be copied from the persistent directories above. Data in scratch folders will be deleted after 90 days, and you should treat it as if it could be deleted at any time.
Warning
Please be aware that employing any methods to alter data in the scratch directory to circumvent the 90-day deletion policy is strictly forbidden and will lead to administrative action by the RC team. For further information, please consult the following resource: RC Scratch Directory Policy.
Your scratch directory is located at:
/n/holyscratch01/<your lab name>
For example, Jonathan Frankle’s lab directory is at:
/n/holyscratch01/jfrankle_lab
This storage is only accessible to members of the lab. The prefix of that path may change in the future, so you can use the $SCRATCH environment variable to refer to the prefix of the path:
cd $SCRATCH/jfrankle_lab
There is no charge for scratch storage.
Note
In the scratch
storage space under the Users
directory, you can have private directory with your username. This will not exist by default. You need to file a ticket with FASRC on the portal to as for a private directory in the scratch
storage space.
5.1.4. Temporary scratch storage (VAST)#
The VAST storage offers improved performance for AI workflows, especially those requiring high I/O in computer vision ML workflows. In addition to the default holyscratch01
scratch space, which is a Lustre-type filesystem, Kempner Cluster users also have access to VAST scratch space, mounted at the following path on the cluster.
/n/vast-scratch/<your lab name>
For example, Jonathan Frankle’s lab directory on VAST scratch storage is at:
/n/vast-scratch/kempner_jfrankle_lab
This VAST scratch storage is only accessible to members of the lab. Each lab has 25 TB of VAST scratch storage space.
Note
The VAST scratch storage mounted on /n/vast-scratch
is accessible only from Kempner cluster compute nodes (holygpu8a17*
for H100 GPU nodes and holygpu8a19*
for A100 GPU nodes), and FASRC is working to mount it on login nodes as well. You have access to this storage during an interactive session on any Kempner partition, or your batch job can access this storage once submitted to a Kempner partition.
5.1.5. Scratch Storage Summary Table#
The following table summarizes the details of the default scratch storage (holyscratch01
) and VAST scratch storage (VAST
) offerings:
Feature |
Default Scratch Storage |
VAST Scratch Storage |
---|---|---|
Name |
holyscratch01 |
VAST |
Filesystem Type |
Lustre |
VAST |
Each Lab Quota |
50 TB |
25 TB |
File Retention |
90 days |
90 days |
Lab Scratch Path |
|
|
Use Cases |
Optimized for Varieties of Workflows |
Optimized for High I/O AI Workflows |
Note
The current VAST scratch storage, mounted on /n/vast-scratch
, serves as a temporary solution. Users may need to migrate to a more permanent VAST storage later. This page will be updated with the most recent information, and users will be notified about any changes to the scratch storage in advance.
The following table summarizes the storage options available on the cluster, visit data storage for more information.