User Tools

Site Tools


hpc:storage_on_hpc

Storage

There are different types of storage on the clusters. This is important to understand where to store which type of data.

Important: please note that Boabab and Yggdrasil have both the same architecture and logic regarding storage, however the data is not shared between the clusters.

Cluster storage

We use a distributed parallel filesystem on our clusters called BeeGFS. Any data you put in $HOME or in $HOME/scratch folders are thereby accessible on the login node and on each compute node of the cluster. This means you don't need to waste time or resources to copy data on the nodes.

This is the storage space we offer on our clusters

Cluster Path Total storage size Nb of servers Nb of targets per servers Backup Quota size Quota number files
Baobab /home/ 138 TB 4 1 meta, 2 storage Yes (tape) 1 TB -
/srv/beegfs/scratch/ 1.0 PB 2 1 meta, 6 storage No - 10 M
/srv/fast 5 TB 1 1 No 500G/User 1T/Group -
Yggdrasil /home/ 495 TB 2 1 meta, 2 storage Yes (tape) 1 TB -
/srv/beegfs/scratch/ 1.2 PB 2 1 meta, 6 storage No - 10 M

We realize you all have different needs in terms of storage. To guarantee storage space for all users, we have set a quota on home and scratch directory, see table above for details. Beyond this limit, you will not able to write to this filesystem. We count on all of you to only store research data on the clusters. We also count on your to periodically delete old or unneeded files and to clean up everything when you will leave UNIGE. Please keep on reading to understand when you should use each type of storage.

Home directory

Your home folder is located here: $HOME.
It is available on the login node and on each compute nodes of the cluster.

In your home directory you can store any file needed for running your jobs : data, code, software, etc.

However, this is not a personal storage for data such as emails, private pictures, etc.

Please always use the the $HOME variable in your script instead of using the full path. This is a good practice and will save you a lot of trouble.

A backup of your $HOME folder is done daily.

For temporary storage, please read about the “Scratch directory” below.

Your $HOME is only accessible by you (permission 0700) ; you are not allowed to change the permissions and if you do, they are automatically reset every day. If you need to share files, please check Sharing files with other users.

Quota

As the storage is shared by everyone, this ensure a fair home usage and prevent users from filling it. It is important to lower the disk usage on home as this location is backuped and this takes a lot of time, especially if you have a lot of files.

What does it mean for you: if your home directory usage is higher than 1TB, you won’t be able to write to it anymore.

Error message:

Disk quota exceeded

To resume the situation, you should clean up some data in your home directory and or migrate some data to your scratch directory.

Scratch directory

Your scratch folder is located here: $HOME/scratch.

It is available on the login node and on each compute nodes of the cluster. It has more space than $HOME.

N.B. : $HOME/scratch is a symbolic link to your user's folder in /srv/beegfs/scratch/.

The scratch directory allows you to store any data that is not unique or that can be regenerated. Please use this to store any file that doesn't need a backup. You will typically use it as a storage when your application writes temporary data to disk. We thank you for your cooperation.

It is also acceptable to store for instance a large dataset that you use as an input data during your computation.

The content of this folder is persistent and accessible from any node, but there is no backup.

Your $HOME/scratch is only accessible by you (permission 0700) ; you are not allowed to change the permissions and if you do, they are automatically reset every day. If you need to share files, please check Sharing files with other users.

Also, the scratch directory is not a permanent storage solution, we strongly advise you to move/clean useless/unused data after your project.

Quota

As the storage is shared by everyone, this ensure a fair scratch usage and prevent users from filling it. We setup a quota based on the number of files you own, not the file size.

The maximum file count is currently set to 10M.

What does it mean for you: if the number of files in your scratch space is higher than 10M, you won’t be able to write to it anymore.

Error message:

Disk quota exceeded

To resume the situation, you should clean up some data in your scratch directory.

Fast directory

A new fast storage is available dedicated for jobs using multiples nodes and scratchlocal need to be shared between nodes.

Cluster path Total storage size Nb of servers Backup Quota size Quota number files
Baobab /srv/fast 5 TB 1 No 500G by user & 1TB by group -
This storage will be erased on each maintenances.

Quota

As the storage is shared by everyone, this ensure a fair scratch usage and prevent users from filling it. We setup a quota based on the total size.

You should clean up some data in your fast directory as soon as your jobs are finished.

Local storage

Scratch directory (local on each node)

On each compute node, a local scratch folder is available and is located here : /scratch.
It is available on all compute node of the cluster. Unlike $HOME/scratch, the local scratch can only be used from the node itself and while your job is running.

It is usually fast and there is no overhead for using the network. It is also more efficient at dealing with a large amount of small files.

You usually use it for temporary files generated while your job is running and that do not need to be accessible at the end of the job.

Important : at the end of the computation, all files in that directory are automatically deleted. If you need to keep some files, you can add a command to your Slurm sbatch script to move them in the $HOME or $HOME/scratch directory at the end of the job.

While all the Yggdrasil nodes have local SSDs, on Baobab we still have HDDs (i.e. mechanical disks) on the following nodes:
node[001-005,007-021,023-056,058-059,061-066,093-101,103-111,140-156,180,186,203].

Temporary private space

On each compute node, you can use the following private ephemeral spaces:

  • /dev/shm: fastest storage, on RAM, be careful to request enough RAM)
  • /var/tmp: on local disk
  • /tmp: on local disk

Those places are private and only accessible by your job.

Temporary shared space

If you need to access the data from more than one node, you can use a space reachable from all your jobs running on the same compute node. When you have no more jobs running on the node, the content of the storage is erased.

The path is the following: /share/users/${SLURM_JOB_USER:0:1}/${SLURM_JOB_USER}

See here for a usage example: https://hpc-community.unige.ch/t/local-share-directory-beetween-jobs-on-compute/2893

Sharing files with other users

Sometimes you need to share files with some colleagues or your research group.

We provide two types of shared folders :

  • in “home” (/home/share/) - to share scripts and shared libraries for a common project
  • in “scratch” (/srv/beegfs/scratch/shares/) - to share datasets for instance

If you need one, please contact us by email :

  • send the email to hpc@unige.ch with your PI in cc
  • provide the following :
    • do you need a “home” or “scratch” shared folder (or both) ?
    • the list of people (email and username) who needs to access the shared folder
    • the desired name of the shared folder
You are not allowed to change the permission of your $HOME/$SCRATCH folder on the clusters. Even if you did, our automation scripts will break what you did.
For easy sharing you need to set umask 0002 (thus new files and directories will be created with 660 and 770 permissions, respectively), otherwise you will be asked for confirmation every time you want to modify a file or, even worse, you will not be able to create new files/folders.

This is a side-effect of the default permissions on Red Hat-based systems without User Private Groups (i.e. when the UID/GID differs, as it is the case on Baobab/Yggdrasil 1) ).

Since we use ACL to set the user right, you can't rely on sticky bit to force the new files to belong to a group which is not your primary group. You have the following options:
  • You can request to change your primary group: every file that you create on the cluster will belong to this group
  • You can set your umask to 0002 as explained previously
  • You can launch on a regular basis as script that “fix” the group. Example: find . -type f -exec chown :share_xxx {} \;

Best practices

I/O performance

Source : 2)

All the BeeGFS storage are shared. Although it is a pretty robust solution, the performance depends of the workload of all the users and can be impacted negatively. This is especially true on a cluster, where a single job may perform I/O from many different threads at the same time, even on multiple nodes.

BeeGFS uses two kind of storage servers.

  • metadata servers: they are involved when you type ls, find, etc. and you don’t read the content. The metadata is stored on fast NVMe disks. See here for details about the infrastructure.
  • storage servers: they are involved when you read or write a file. The data are stored on regular hard drives disks. See here for details about the infrastructure.

The bottleneck is usually the metadata servers and you notice it while trying to browse files.

A performance killer is for example to read/write a lot of very small files or to traverse a directory with thousands of files. This is usually something the user himself has a control on and can avoid, for example by limiting the number of files per directory.

Also, think twice before launching those tools on a shared storage:

  • updatedb
  • find
  • ncdu
  • du
  • or anything that takes a long time to scan all files on the storage

Best practices :

  • limit the number of files per directory (for example, 1000)
  • do not test the performance of the storage by launching a “benchmark”, as you’ll waste precious resources
  • work with bigger files (1MB instead of 1KB for example). In this case, size matters and the bigger, the better.
  • avoid reading many times the same file, try to cache it instead.

Check disk usage on the clusters

Check disk usage on home and scratch

Since /home and /srv/beegfs/scratch/ have quota enabled and enforced, we can quickly check the disk usage by fetching the quota information.

The script beegfs-get-quota-home-scratch.sh gives you a quick summary :

(baobab)-[sagon@login2 ~]$ beegfs-get-quota-home-scratch.sh
home dir: /home/sagon
scratch dir: /srv/beegfs/scratch/users/s/sagon
 
        user/group                 ||           size          ||    chunk files
storage     |   name        |  id  ||    used    |    hard    ||  used   |  hard
----------------------------|------||------------|------------||---------|---------
home        |          sagon|240477||  530.46 GiB| 1024.00 GiB||  1158225|unlimited
scratch     |          sagon|240477||    2.74 TiB|   unlimited||   436030| 10000000

This includes all your data in $HOME, $HOME/scratch, but also any data in /home/share and /srv/beegfs/scratch/shares that belongs to you (if you are using a shared directory).

The column “chunk files” doesn't correspond exactly to the number of files you own. It corresponds to the number of chunks you own. Each file has at least one chunk. The current chunk size is 512kB. If a file is bigger, it will be split.

Check disk usage on NASAC

If you have space as well in /acanas (The NASAC) you can check your quota and usage like this:

-bash-4.2$ quota --hide-device  -s -f /acanas
Disk quotas for user XXX (uid NNN):
     Filesystem   space   quota   limit   grace   files   quota   limit   grace
   1975G      0K   3072G          15725k       0       0

File transfer (between nodes)

There is no need to transfer files between compute nodes and the login as they share the same storage space ($HOME and $HOME/scratch).

However, if for some reason you need specifically to transfer something from the login node to the local storage of a compute node, you can use sbcast (but you probably shouldn't).

Backup

Your home ($HOME) directory is backed up every day.

To avoid backing up temporary data, we ask you to be fair and store them in the $HOME/scratch/ directory.

There is no backup of the scratch folder.

If you deleted something from the $HOME directory, please contact us by email :

  • send the email to hpc@unige.ch
  • provide the following :
    • the full path of the folder or file you deleted
    • the date and time (as precise as possible) when you deleted it
    • the date and time (as precise as possible) of the last time you used the file
    • we will try to retrieve the file(s) from the most recent backup available before deletion.

Please be aware that the backup retention is limited. The backup ruls are :

  • For “active” files, we keep 10 versions or the ones modified in the last 30 days.
  • For “inactive” files (file that are not present on disk, either because they were deleted or moved), we keep the last 2 versions for 90 days. After 90 days, those copies are deleted for good.

Thus we cannot guarantee to recover a version of a file more recent than 24h, or on the contrary a file you deleted a long time ago. Don't bargain based on the above mentioned rules, just contact us as soon as possible and we will see what is doable.

Archive

If you need to archive you results, you can use the following services:

  • NAS academic, to store archives on a CIFS or NFS share at the price of 75.-CHF / TB / year.
  • Yareta for long-term preservation.

Baobab isn't a long term archive service!

Access external storage

If you need to transfer files from “outside” the cluster, please refer to this documentation

NASAC

GIFS is not working due to a dummy patch integrated by the fast network supplier (Infiniband by Mellanox/Nvidia).

2024-12-05 Update: Since Rocky9 was deployed on Bamboo during the last maintenance, the module now seems to be available again. You should be able to mount your NASAC share on both the login nodes and compute nodes.

Rocky9 will also be deployed on Baobab and Yggdrasil during the next maintenances.

For more information: https://hpc-community.unige.ch/t/2024-current-issues-on-hpc-cluster/3245/15

If you need to mount an external share (NAS for example) on Baobab from command line, you can proceed as follow in your terminal.

Launch dbus:

[sagon@login2 ~] $ dbus-launch bash

mount the share, smb in this example:

[sagon@login2 ~] $ gio mount smb://server_name/share_name

This will prompt for the username, password and domain. If you are mounting an UNIGE network share such as the NASAC, it’s your ISIS credentials and the domain is ISIS.

The share will be mounted on /run/user/your_uid/gvfs/

You can access the files using standard POSIX tools such as cp ls etc but keep in mind that those tools aren't meant to be used on non reliable media such as network share. If you face an error when accessing a specific file, you can use gio copy or rsync as cp replacement which works better and handle network errors. The same for ls etc.

When you don’t need to access the data anymore, you may unmount the share:

[sagon@login2 ~] $ gio mount -u smb://server_name/share_name
The data are only available on the login2 node. If you need to access the data on the nodes, you need to mount them there as well in your sbatch script.

If you need to script this, you can put your credentials in a file in your home directory.

Content example with the credendtials stored in the file .credentials.

username
domain
password

Mount example using credentials in a script:

[sagon@login2 ~] $ gio mount smb://server_name/share_name < .credentials

Do not store the credentials in the file .netrc as the format conflict with other tools using it such as wget.

Troubleshooting

Where does gio mounts my data?

The process that lets you access GVfs/Gio shares via CLI is called gvfsd-fuse and its first argument is the folder where GVfs/Gio shares are exposed to.

You can find such folder with the following command:

[sagon@login2 ~] $ ps ux | grep -e '[g]vfsd-fuse'
sagon    196919  0.0  0.0 387104  3376 ?        Sl   08:49   0:00 /usr/libexec/gvfsd-fuse /home/sagon/.gvfs -f -o big_writes

In this case, it means the mount will be done in /home/sagon/.gvfs

List the user DBUS process

[sagon@login2 ~] $ pgrep -a -U $(id -u) dbus
196761 /usr/bin/dbus-daemon --fork --print-pid 4 --print-address 6 --session
224317 /usr/bin/dbus-daemon --fork --print-pid 4 --print-address 6 --session

reference: 3)

CVMFS

All the compute nodes of our clusters have CernVM-FS client installed. CernVM-FS, the CernVM File System (also known as CVMFS), is a file distribution service that is particularly well suited to distribute software installations across a large number of systems world-wide in an efficient way.

A couple of repository are mounted on the compute and login node such as:

  • atlas.cern.ch
  • atlas-condb.cern.ch
  • atlas-nightlies.cern.ch
  • sft.cern.ch
  • grid.cern.ch

The content is mounted using autofs under the path /cvmfs. It means that the root directory /cvmfs may appears empty as long as you didn't access explicitly one of the child directory. Doing so will mount the repository for a couple of minutes and unmount it automatically.

Other flaghship repository available without further configuration:

  • unpacked.cern.ch
  • singularity.opensciencegrid.org (container registry)
  • software.eessi.io (
[root@node002 ~]# ls /cvmfs
[root@node002 ~]# ls -dl /cvmfs/grid.cern.ch
drwxrwxr-x 56 cvmfs cvmfs 2 Feb  1  2010 /cvmfs/grid.cern.ch
[root@node002 ~]# ls /cvmfs/
cvmfs-config.cern.ch  grid.cern.ch

The EESSI did a nice tutorial about CVMFS readable on multixscale git repo.

Robinhood

Robinhood Policy Engine is a versatile tool to manage contents of large file systems. It daily scans the scratch beegfs filesystems. It makes it possible to schedule mass action on filesystem entries by defining attribute-based policies.

Policies

Each file matchinga policy will be recorded in a robinhood report and accessible by the owner.

LargeDir: Triggered if the directory contains more than 500 files

The more files a directory contains, the more impact it has on the performance of the filesystem and therefore on other users. So we stronghly encourage you to avoid The LargeDir .

OldFiles: Triggered if a file has not been accessed for more than 60 days

The colder the data, the less likely it is to be reused. This rule is widely verified on most storage systems. As scratch is not a permanent storage, cold data tends to take up space unnecessarily. Some files have not been opened/modified since 2013. What is the probability that it will be used again soon. (close to 0%)

Eventually, we may apply certain actions to policies, such as deleting old and unused files, to free up space and allow users who need it to continue their projects. So stay tuned to the HPC team's communication, to the best practices described in our documentation and to the evolution of robinhood.

Report

If some files have been triggered by robinhood, a report will be displayed on the connection node:

(baobab)-[terminator@login2~]$ 

###############################################################################
                                Robinhood Report
###############################################################################

Cluster storage usage:

       USER >     /home    | /srv/beegfs/scratch
       root >  107.48 GiB  |    1.79 TiB 


Policies report:
    * LargeDir: 173
      Triggered if the directory contains more than 500 files

    * OldFiles: 12307
      Triggered if a file has not been accessed for more than 60 days



Please check /srv/beegfs/scratch/log/robinhood/report/t/terminator.log 
for more details and clean them up if  possible.

###############################################################################

This is not a decorative item and should be taken seriously, this report is intended to inform you of files that do not comply with the storage policy. This list of files should get your attention and prompt you to sort through your files, delete or migrate your data to permanent storage, or reduce the number of files in a directory to avoid a drop in performance.

hpc/storage_on_hpc.txt · Last modified: 2024/12/09 11:10 by Adrien Albert