Difference between revisions of "Usage of a HPC cluster"

From Wiki Max
Jump to navigation Jump to search
Line 30: Line 30:
  
 
   <user>:  Unix username on the cluster login node
 
   <user>:  Unix username on the cluster login node
 +
  <machine_host_name>:  hostname (or IP) of the target HPC machine
 +
 +
In order to access you need to have a working Unix username and passwd on the target machine.
 +
 +
==Scheduler==
 +
 +
Multiple schedulers are available. Examples include PBS Torque, IBM LoadLeveler and SLURM.
 +
Here we just cover some examples, providing the main commands to be used within the shell.
 +
 +
===PBS Torque===
 +
* submitting a job
 +
  qsub  file.sh
 +
 +
*  monitoring a job
 +
  qstat              # general query, all jobs are shown
 +
  qstat -u $USER    # only my jobs
 +
  qstat JID
 +
 +
* checking queue properties
 +
  qstat -q
 +
 +
* deleting a job
 +
  qdel JID
 +
 +
  JID (Job ID) is given by qstat
 +
 +
* Interactive use on compute nodes
 +
  qsub -I -q <queue_name>
 +
 +
==Environment==
 +
 +
Use the <code>module</code> command to manage the unix environment
 +
 +
Make sure you load a proper parallel environment (incl MPI-aware compilers)
 +
 +
Some examples:
 +
 +
  module purge
 +
  module load psxe_2015
 +
 +
These commands can be also place in the .bashrc file and most of the time need to be present
 +
in submission scripts
 +
 +
In order to enquire the module system:
 +
  module list      # lists loaded modules
 +
  module avail    # lists available modules

Revision as of 09:29, 30 April 2021


Here we collect some general (and by no means complete) information about usage and policies of a HPC cluster.

Structure of a HPC cluster

Logical structure HPC cluster.png

The structure of a HPC system is sketched in the picture above. These are the main logical building blocks:

  • a login node is exposed to users for access (typically via ssh),
  • a dedicated scheduler (the queuing system) dispatches computational jobs to the compute nodes
  • computation happens therefore asynchronously (in batch mode), and not on the login node.
  • a specific software environment is provided on the login node and on the compute node to run parallel jobs
  • the above cluster structure is used from commodity local machines to large scale HPC architectures (as shown below for the Summit machine at Oak-Ridge Natl Lab (ORNL), TN, USA)

Summit-ONL.jpg

Connecting

Unless other means are provided, you typically connect using the ssh protocol.

From a shell terminal or a suitable app:

 ssh -Y <user>@<machine_host_name>  or
 ssh -Y -l <user> <machine_host_name>
 <user>:  Unix username on the cluster login node
 <machine_host_name>:  hostname (or IP) of the target HPC machine 

In order to access you need to have a working Unix username and passwd on the target machine.

Scheduler

Multiple schedulers are available. Examples include PBS Torque, IBM LoadLeveler and SLURM. Here we just cover some examples, providing the main commands to be used within the shell.

PBS Torque

  • submitting a job
 qsub  file.sh
  • monitoring a job
 qstat              # general query, all jobs are shown
 qstat -u $USER     # only my jobs
 qstat JID
  • checking queue properties
 qstat -q
  • deleting a job
 qdel JID
 JID (Job ID) is given by qstat
  • Interactive use on compute nodes
 qsub -I -q <queue_name>

Environment

Use the module command to manage the unix environment

Make sure you load a proper parallel environment (incl MPI-aware compilers)

Some examples:

 module purge
 module load psxe_2015

These commands can be also place in the .bashrc file and most of the time need to be present in submission scripts

In order to enquire the module system:

 module list      # lists loaded modules
 module avail     # lists available modules