Trace: • iscinventory

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revisionPrevious revision
Next revision		Previous revision
infra:howto:runjob [2025/02/19 08:55] – [Environment] remiinfra:howto:runjob [2026/03/11 10:25] (current) – [Quotas] marc
Line 10: Line 10:
  
  
 +===== Connection =====
 +Currently, you have to SSH directly on one of the compute node, either server [[infra:chacha|Chacha]] or [[infra:disco|Disco]] :
 +
 +  $ ssh firstname.lastname@disco.hevs.ch
 +  $ ssh firstname.lastname@chacha.hevs.ch
 +  
 +** NOTE : Please note you have to connect from the school network or the HEVS VPN to be able to connect on those servers **
 +
 +TODO : Change doc when the jump host will be ready
  
 ===== Environment ===== ===== Environment =====
  
-On either server [[infra:chacha|Chacha]] or [[infra:disco|Disco]], you have a symlink **datasets** in your home directory that is linked to the local storage of the server : its purpose is to give you a proper space to put all the data you will be working on.+ 1. On either server [[infra:chacha|Chacha]] or [[infra:disco|Disco]], you have a symlink **datasets** in your home directory that is linked to the local storage of the server : its purpose is to give you a proper space to put all the data you will be working on.
  
-You have also another symlink **shared_dataset** for jobs that needs to be run on several nodes : this filesystem is shared between nodes.+ 2. You have also another symlink **shared_dataset** for jobs that needs to be run on several nodes : this filesystem is shared between nodes.
  
 + 3. Your .bashrc / .zshrc contains by default the variable **APPTAINER_TMPDIR** set to **/home/user.name/.apptainer/** : this allows you to build containers without using the system /tmp that is restricted with a low quota, and use your larger dataset quota instead.
  
-NOTE : By default, you are the only one seeing your data : If you are working as a team on these data, **please ask for a group creation** so we can add members in it and apply suitable permissions.+ 4. By default, you are the only one seeing your data : If you are working as a team on these data, **please ask for a group creation** so we can add members in it and apply suitable permissions.
  
  
Line 27: Line 37:
 To avoid having everyone installing their libraries installed on the system or on their user directly on the physical servers, we need you to keep them cleanly packed in a container : That way you can both install what you want inside this container, and you can do it without needing any root priviledge on the server you are sharing with other researchers. To avoid having everyone installing their libraries installed on the system or on their user directly on the physical servers, we need you to keep them cleanly packed in a container : That way you can both install what you want inside this container, and you can do it without needing any root priviledge on the server you are sharing with other researchers.
  
-See [[infra:howto:apptainer_sample|How to create a simple apptainer container]]+**For examples, see : [[infra:howto:apptainer_sample|How to create a simple apptainer container]]**
  
  
 ===== Run your application via SLURM  ===== ===== Run your application via SLURM  =====
  
-Now that you have your container ready in the Apptainer file format : for example application.sif+To be able to run a job on the ISC Compute Center, you **MUST** run it under [[https://slurm.schedmd.com/overview.html|SLURM]]. Ressource usage is managed by Slurm on this cluster.
  
-To run your job via SLURM you need to create a sbatch script : +**For examples, see : [[infra:howto:slurm_sample|How-to create a simple SLURM job]]**
  
-(Note: for now we don't have Modules or LMod installed, we might add it later : hence the module commands commented) 
  
-==== Example for a serial job ==== 
  
-<code> +More information on our **[[infra:compute:slurmconfig|SLURM cluster]]** 
-#!/bin/bash +===== Storage considerations =====
-#SBATCH --job-name=application   # create a short name for your job +
-#SBATCH --nodes=1                # node count +
-#SBATCH --ntasks=1               # total number of tasks across all nodes +
-#SBATCH --cpus-per-task=1        # cpu-cores per task (>1 if multi-threaded tasks) +
-#SBATCH --mem=4G                 # total memory per node (4 GB per cpu-core is default) +
-#SBATCH --time=00:05:00          # total run time limit (HH:MM:SS) +
-#SBATCH --mail-type=begin        # send email when job begins +
-#SBATCH --mail-type=end          # send email when job ends +
-#SBATCH --mail-user=email@domain.org +
-#module purge +
-apptainer run application.sif <arg-1> <arg-2> ... <arg-N> +
-</code>+
  
 +==== Filesystems structure ====
  
-==== Example for a parallel MPI code ==== +On each server, your user is created with the following home directory : 
-<code> +  ls -l /home/user.name 
-#!/bin/bash +  lrwxrwxrwx  1 root              root                51 Feb  4 10:21 .apptainer -> /data/disk01/apptainer/user.name/.apptainer/ 
-#SBATCH --job-name=solar         # create a short name for your job +  lrwxrwxrwx  1 root              root                38 Feb  4 10:21 datasets -> /data/space/datasets/user.name/ 
-#SBATCH --nodes=1                # node count +  lrwxrwxrwx  1 root              root                31 Feb  4 10:21 results -> /data/shared/user.name/results/ 
-#SBATCH --ntasks=4               # total number of tasks across all nodes +  lrwxrwxrwx  1 root              root                31 Feb  4 10:22 shared_datasets -> /data/shared/user.name/shared_datasets/ 
-#SBATCH --cpus-per-task=1        # cpu-cores per task (>1 if multi-threaded tasks) +  
-#SBATCH --mem-per-cpu=4G         # memory per cpu-core (4G per cpu-core is default) +
-#SBATCH --time=00:05:00          # total run time limit (HH:MM:SS) +
-#SBATCH --mail-type=begin        # send email when job begins +
-#SBATCH --mail-type=end          # send email when job ends +
-#SBATCH --mail-user=email@domain.org +
-#module purge +
-#module load openmpi/gcc/4.1.2 +
-srun apptainer exec solar.sif /opt/ray-kit/bin/solar inputs.dat +
-</code>+
  
 +  * **.apptainer** is your cache directory for your containers, use ''apptainer cache clean'' every week to clean unused containers.
 +  * **datasets** is your main directory for all your work material, do NOT use directly /home/user.name/ to put data or software on it : it would uselessly fill the root partition.
 +  * **results** is located on a shared filesystem between all Slurm nodes to allow you to run jobs anywhere : as long as you write the output of your job in results, you will be able to access it whatever node the job had run on.
 +  * **shared_datasets** is located on a shared filesystem between all Slurm nodes to allow you to run jobs anywhere : when your job needs to get the same data from several nodes, use this directory. **NOTE :** you will have less performances for fast reads/writes since it is a network shared filesystem : if this is a concern, please consider using the faster local datasets directory above and specify your work node in your sbatch using --nodelist=xxx.
  
-==== Example using GPUs ==== 
  
-<code> +==== Data retention ====
-#!/bin/bash +
-#SBATCH --job-name=tensorflow    # create a short name for your job +
-#SBATCH --nodes=1                # node count +
-#SBATCH --ntasks=1               # total number of tasks across all nodes +
-#SBATCH --cpus-per-task=4        # cpu-cores per task (>1 if multi-threaded tasks) +
-#SBATCH --mem-per-cpu=4G         # memory per cpu-core (4G per cpu-core is default) +
-#SBATCH --time=00:05:00          # total run time limit (HH:MM:SS) +
-#SBATCH --gres=gpu:1             # number of gpus per node +
-#SBATCH --mail-type=begin        # send email when job begins +
-#SBATCH --mail-type=end          # send email when job ends +
-#SBATCH --mail-user=email@domain.org +
-#module purge +
-apptainer exec --nv ./tensorflow.sif python3 tensor.py +
-</code>+
  
-Note the **--nv** flag that allows you to use the GPU from your container without being root.+On the ISC Compute Center, you need to be aware of how much time data is kept on which location.
  
-==== Example using GPU sharding ====+  - **Local datasets :** Data life is short here, for now cleaning is not automated but will probably be in the future. Useful for a large volume of data to be processed on one server, and with the fastest I/O available. 
 +  - **Shared datasets and results :** Data life is short here, for now cleaning is not automated but will probably be in the future. Useful for a large volume of data to be processed on several servers, and with reasonable I/O. (with some network latency due to NFS) 
 +  - **Filer mount :** Data life is long, and backuped : this is where you can store data, results and other files for a longer time. By default you don't have a filer mount, it is the [[https://hessoit.sharepoint.com/sites/VS-Intranet-SInf|Sinf]] who can create you a shared directory on their main filer (filer01.hevs.ch) on their infrastructure, which is backuped and also isolated with limited access. [[https://servicedesk.hevs.ch/hesso_portal?id=sc_category&sys_id=944044f987aca110a741cd0a0cbb35ae|Create this filer mount with their "Demande d'obtention de droits d'accès réseau" request ]]  
 +**NOTE :** Please note the filer storage is slow, it is not advised to run a job directly using those data.
  
-Sharding is a generic way of SLURM to use fragments of a GPU for a job, leaving room for other researchers, or running several jobs needing each one a fragment of GPU.+**TODO :** set auto-cleaning for old data on each filesystem
  
-<code> 
-#!/bin/bash 
-#SBATCH --partition=Dance        # Partition to run the job on 
-#SBATCH --job-name=tensorflow    # create a short name for your job 
-#SBATCH --nodes=1                # node count 
-#SBATCH --ntasks=1               # total number of tasks across all nodes 
-#SBATCH --cpus-per-task=4        # cpu-cores per task (>1 if multi-threaded tasks) 
-#SBATCH --mem-per-cpu=4G         # memory per cpu-core (4G per cpu-core is default) 
-#SBATCH --time=02:00:00          # total run time limit (HH:MM:SS) 
-#SBATCH --gres=shard:24          # number of gpu shards to use 
-#SBATCH --mail-type=begin        # send email when job begins 
-#SBATCH --mail-type=end          # send email when job ends 
-#SBATCH --mail-user=email@domain.org 
-#module purge 
-apptainer exec --nv ./tensorflow.sif python3 tensor.py 
-</code> 
  
-Note the **--nv** flag that allows you to use the GPU from your container without being root. Here, the partition has been requested to be specifically run on the Disco node, which has max 80 shards. On Chacha the limit is 96 shards. 
  
 +==== Quotas ====
 +  - The root filesystem (/, including /home) for every researchers is 20GB for convenience. For students, this quota is lowered to 10GB to encourage proper infrastructure usage / coding with a lower threshold.
 +  - On local and shared datasets, quota will be set on a case by case basis
  
-==== Example using an interactive shell ====+==== Cleaning ==== 
 +  - Users must regularly clean the Apptainer cache using ''apptainer cache clean''. Every week is a good starting point, we might automate cleaning if this is not respected : please don't think the container is a persistent storage for your data, it is not. 
 +  - Don't keep data directly on your /home/user.name/ directory : you are limited to 20GB on it just for convenience : no large file should be there. 
 +  - When you create new apptainer containers, the .sif files can be quite big : don't store them in /home/user.name/, prefer using either datasets/ or if needed, archive them in your filer mount.
  
-**NOTE : Debugging your application directly on the ISC compute is to be avoided ** 
- 
- 
-To debug your application on a test Slurm + Apptainer infrastructure, you can use srun with the --pty argument to run your container : 
- 
-<code> 
-# For a simple compute container : 
-srun --cpus-per-task=12 --time=3:00:00 --mem=24G --pty apptainer shell /home/user.name/example_apptainer.sif 
- 
-# Or a container using GPUs : 
-srun -G 1 --cpus-per-task=12 --time=3:00:00 --mem=24G --pty apptainer shell --nv /home/user.name/example_apptainer.sif 
-</code> 
-==== Execute your batch file ==== 
- 
-Then you can run your sbatch script : 
- 
-  sbatch ./application_sbatch.sh 
  
  
 ===== Good practice ===== ===== Good practice =====
  
-==== Storage considerations ====+==== Coding considerations ====
   - **Don't use VScode SSH plugin** to avoid constant file scanning on our servers :    - **Don't use VScode SSH plugin** to avoid constant file scanning on our servers : 
    * [[https://earlruby.org/2021/06/fixing-vscode-when-it-keeps-dropping-ssh-connections/]]     * [[https://earlruby.org/2021/06/fixing-vscode-when-it-keeps-dropping-ssh-connections/]] 
Line 147: Line 99:
    * [[https://github.com/microsoft/vscode-cpptools/issues/5362]]    * [[https://github.com/microsoft/vscode-cpptools/issues/5362]]
    * [[https://learn.microsoft.com/en-us/answers/questions/1221136/visual-studio-2022-clear-local-caches]]    * [[https://learn.microsoft.com/en-us/answers/questions/1221136/visual-studio-2022-clear-local-caches]]
-  - Currently your vscode caches will be automatically removed from our servers. +  - Currently your vscode caches are automatically removed from our servers.
-  - Users must regularly clean the Apptainer cache using ''apptainer cache clean''. Every week is a good starting point, we might automate cleaning if this is not respected : don't think the container is a persistent storage for your data, it is not. +
- +
-==== Filesystems structure ==== +
- +
-On each server, your user is created with the following home directory : +
-  ls -l /home/user.name +
-  lrwxrwxrwx  1 root              root                51 Feb  4 10:21 .apptainer -> /data/disk01/apptainer/user.name/.apptainer/ +
-  drwx------  2 user.name         user.name         4096 Dec 19  2023 .cache/ +
-  lrwxrwxrwx  1 root              root                38 Feb  4 10:21 datasets -> /data/space/datasets/user.name/ +
-  lrwxrwxrwx  1 root              root                31 Feb  4 10:21 results -> /data/shared/user.name/results/ +
-  lrwxrwxrwx  1 root              root                31 Feb  4 10:22 shared_datasets -> /data/shared/user.name/shared_datasets/ +
-   +
- +
-  * **.apptainer** is your cache directory for your containers, use ''apptainer cache clean'' every week to clean unused containers. +
-  * **datasets** is your main directory for all your work material, do NOT use directly /home/user.name/ to put data or software on it : it would uselessly fill the root partition. +
-  * **results** is located on a shared filesystem between all Slurm nodes to allow you to run jobs anywhere : as long as you write the output of your job in results, you will be able to access it whatever node the job had run on. +
-  * **shared_datasets** is located on a shared filesystem between all Slurm nodes to allow you to run jobs anywhere : when your job needs to get the same data from several nodes, use this directory. **NOTE :** you will have less performances for fast reads/writes since it is a network shared filesystem : if this is a concern, please consider using the faster local datasets directory above and specify your work node in your sbatch using --nodelist=xxx.+
  
 +==== Automation ====
  
 +If you need to use Cron to schedule something, you need to ask for your user to be added to the /etc/cron.allow whitelist.
Back to top