Our clusters are accessed through a login node[^1], that login node can by used low cpu work like software development (testing, compiling porting, small test), job management, data transfer...
However all heavy duty computing must go through the SLURM manager, and non complying processes can be killed without prior notice.
Usage examples are provided in the [cluster specific](Home) pages of this wiki, but all facilities share a common architecture.
## Useful links
*[The Slurm cheat sheet](https://slurm.schedmd.com/pdfs/summary.pdf): the most usual commands and options.
*[The Rosetta page](https://slurm.schedmd.com/rosetta.pdf): if you're used to a popular job scheduler, you will find translations of the most common commands.
# Submitting different types of jobs
All the examples are [provided in this repository](/home#examples).
How to connect to a node for an interactive session.
## Partition usage
There are different partitions adapted to different kind of jobs. The time limit for each partition can be obtained with the command [sinfo](https://slurm.schedmd.com/sinfo.html).
### Licallo
**seq:* for sequential/multi-thread jobs.
**short-seq:* for short sequential/multi-thread jobs.
**x40:* for parallel jobs, each node has 40 core.
**short:* for short parallel jobs
**fdr:* should be avoided.
**besteffort:* for idempotent jobs. This partition has few constraints, but jobs of this partition will be interrupted whenever theirs resources are needed (advanced).
**bash:* for interactive jobs.
# Controlling/querying jobs
*[When is my job going to start](Running-the-jobs/When is my job going to start)
# Tips and tricks
*[Pack arrays of sequential jobs](Running-the-jobs/Pack-arrays-of-sequential-jobs) to efficiently handle job array
*[Connect on a running job's node](Running-the-jobs/Connect-on-a-running-job's-node) for debugging purpose
*[Use local copies for IO intensive work](Running-the-jobs/Use-local-copies-for-IO-intensive-work)
How to use local file access when the share disk space cannot handle the load.
*[Job dependencies](Running-the-jobs/Job-Dependencies) to handle restarts and multi stage jobs
*[Moving data around](Running-the-jobs/Moving-Data-Around)
*[Using a specific account](Running-the-jobs/Using-a-specific-account)
*[Special tools](Running-the-jobs/Special-Tools)
Running Mathematica etc..
[^1]:there might be more than one actual machine to ensure load balancing and fault tolerance. But you only need to remember the logical name. Like *licallo.oca.eu* for the HPC cluster.
Size job max:
"En général, la façon la plus rapide (mais par essai erreur) de voir ce genre de choses, c'est de soumettre le job puis de faire un squeue -u alainm.
If the job is stuck (statut PD) by a limit, this will indicate which.
Otherwise, the max number of cores is generally given by the MaxTRES property the QOS (Quality Of Service) qosoca-par, at the moment it is 600:
15:48:58 [alainm@pollux rel]# sacctmgr show QOS qosoca-par format=name,MaxTRES
Name MaxTRES
---------- -------------
qosoca-par cpu=600
15:49:16 [alainm@pollux rel]#
Mais cela peu bouge en fonction de la charge par exemple.
TRES signifie Trackable RESources, qui chez nous sont les cœurs (parce que c'est important de rester intuitif).
[[_TOC_]]
# Running the job though Slurm
Our clusters are accessed through a login node[^1], that login node can by used low cpu work like software development (testing, compiling porting, small test), job management, data transfer...
However all heavy duty computing must go through the SLURM manager, and non complying processes can be killed without prior notice.
Usage examples are provided in the [cluster specific](Home) pages of this wiki, but all facilities share a common architecture.
## Useful links
*[The Slurm cheat sheet](https://slurm.schedmd.com/pdfs/summary.pdf): the most usual commands and options.
*[The Rosetta page](https://slurm.schedmd.com/rosetta.pdf): if you're used to a popular job scheduler, you will find translations of the most common commands.
# Submitting different types of jobs
All the examples are [provided in this repository](/home#examples).
How to connect to a node for an interactive session.
## Partition usage
There are different partitions adapted to different kind of jobs. The time limit for each partition can be obtained with the command [sinfo](https://slurm.schedmd.com/sinfo.html).
### Licallo
**seq:* for sequential/multi-thread jobs.
**x40:* for parallel jobs, each node has 40 core.
**short:* for short parallel jobs
**fdr:* should be avoided.
**besteffort:* for idempotent jobs. This partition has few constraints, but jobs of this partition will be interrupted whenever theirs resources are needed (advanced).
**bash:* for interactive jobs.
# Controlling/querying jobs
*[When is my job going to start](Running-the-jobs/When is my job going to start)
# Tips and tricks
*[Pack arrays of sequential jobs](Running-the-jobs/Pack-arrays-of-sequential-jobs) to efficiently handle job array
*[Connect on a running job's node](Running-the-jobs/Connect-on-a-running-job's-node) for debugging purpose
*[Use local copies for IO intensive work](Running-the-jobs/Use-local-copies-for-IO-intensive-work)
How to use local file access when the share disk space cannot handle the load.
*[Job dependencies](Running-the-jobs/Job-Dependencies) to handle restarts and multi stage jobs
*[Moving data around](Running-the-jobs/Moving-Data-Around)
*[Using a specific account](Running-the-jobs/Using-a-specific-account)
*[Special tools](Running-the-jobs/Special-Tools)
Running Mathematica etc..
[^1]:there might be more than one actual machine to ensure load balancing and fault tolerance. But you only need to remember the logical name. Like *licallo.oca.eu* for the HPC cluster.
Size job max:
"En général, la façon la plus rapide (mais par essai erreur) de voir ce genre de choses, c'est de soumettre le job puis de faire un squeue -u alainm.
If the job is stuck (statut PD) by a limit, this will indicate which.
Otherwise, the max number of cores is generally given by the MaxTRES property the QOS (Quality Of Service) qosoca-par, at the moment it is 600:
15:48:58 [alainm@pollux rel]# sacctmgr show QOS qosoca-par format=name,MaxTRES
Name MaxTRES
---------- -------------
qosoca-par cpu=600
15:49:16 [alainm@pollux rel]#
Mais cela peu bouge en fonction de la charge par exemple.
TRES signifie Trackable RESources, qui chez nous sont les cœurs (parce que c'est important de rester intuitif).
