Robin-hpc

From Robin

(Difference between revisions)
Jump to: navigation, search
(SSH from outside UiOs networks)
(Matlab R2019b)
(10 intermediate revisions not shown)
Line 17: Line 17:
| CentOS
| CentOS
|-
|-
-
| worrker node
+
| worker node
| 120 cores/240 vCPU
| 120 cores/240 vCPU
| 460GB
| 460GB
Line 42: Line 42:
=== SSH from outside UiOs networks ===
=== SSH from outside UiOs networks ===
-
To access the machine from outside UiOs networks, see [https://robin.wiki.ifi.uio.no/Deep_Learning_Workstations#Configure_SSH Configure SSH]. To make SSH slightly more pleasant to work with we can create a configuration file at <code>~/.ssh/config</code> which can contain additional information about remote machines. The following is a possible configuration for the Robin-hpc (see [https://en.wikibooks.org/wiki/OpenSSH/Cookbook/Proxies_and_Jump_Hosts#Passing_Through_One_or_More_Gateways_Using_ProxyJump here] for more info). Indiviual key paths can be added with <code>IdentityFile ~/.ssh/jump_key</code>.
+
To access the machine from outside UiOs networks, see [https://robin.wiki.ifi.uio.no/Deep_Learning_Workstations#Configure_SSH Configure SSH]. To make SSH slightly more pleasant to work with we can create a configuration file at <code>~/.ssh/config</code> which can contain additional information about remote machines. The following is a possible configuration for the Robin-hpc (see [https://en.wikibooks.org/wiki/OpenSSH/Cookbook/Proxies_and_Jump_Hosts#Passing_Through_One_or_More_Gateways_Using_ProxyJump here] for more info). Indiviual key paths can be added with <code>IdentityFile ~/.ssh/<key_on_jump_host></code>.
  Host uio-login
  Host uio-login
Line 52: Line 52:
   ProxyJump uio-login
   ProxyJump uio-login
   User <my-username>
   User <my-username>
 +
Now, <code>ssh hpc-robin</code> should directly establish a connection to the server (commands like <code>scp</code> should also work in the same way).
== Access ==
== Access ==
Line 144: Line 145:
== Software ==
== Software ==
-
=== Matlab R2019b ===
+
=== Matlab ===
-
 
+
{{TODO|sebastto}}
-
Todo: sebastto
+
'''Setting up the SLURM job script'''
'''Setting up the SLURM job script'''
Line 196: Line 196:
=== Anaconda ===
=== Anaconda ===
-
See [https://robin.wiki.ifi.uio.no/Deep_Learning_Workstations#Anaconda Deep Learning Workstations] on how to initialize Anaconda.
 
-
Todo: julianpf
+
After initializing Anaconda (see [https://robin.wiki.ifi.uio.no/Deep_Learning_Workstations#Anaconda Deep Learning Workstations]), it is possible to start a job on robin-hpc by using the command [https://robin.wiki.ifi.uio.no/Robin-hpc#SLURM above] and the following template:
 +
 
 +
'''SLURM Anaconda template'''
 +
<pre class="brush: bash">
 +
#!/usr/bin/bash
 +
 
 +
#SBATCH --job-name=<test_name>
 +
#SBATCH --output=log.txt
 +
#SBATCH --ntasks=1
 +
#SBATCH --time=01:00:00
 +
#SBATCH --mem-per-cpu=100
 +
 
 +
# load anaconda and your environment
 +
source ~/.bashrc
 +
conda activate <your_environment>
 +
 
 +
# Run your program with "srun your_command"   
 +
srun python3 <script_name>.py
 +
</pre>
 +
 
 +
The output is written to <code>"log.txt"</code> and can be traced via <code>"tail -f log.txt"</code>.
=== Podman ===
=== Podman ===
  alias docker=podman
  alias docker=podman

Revision as of 14:08, 9 June 2021

Contents

Hardware and network configuration

The robin-hpc is a shared resource for robins researchers and Master students. The strength of the machine is the amount of CPU cores and RAM. Unforunatly, there's no GPU available in this service.

Specs

CPU RAM OS
login node 2 cores/4 vCPU 16GB
CentOS
worker node 120 cores/240 vCPU 460GB CentOS

Storage

The storage on the nodes consists of one 1TB disk where 100GB is reserved for software and 900GB is reserved for the users of the node. However, each user has a soft limit of 20 GB and a hard limit of 100 GB with a grace period of 14 days.

It's important to note that there is no backup of the disk, so do not use the robin-hpc as a cloud storage service. We suggest using rsync/scp to your home area on login.ifi.uio.no.

E.g.

rsync --progress <file>.tar.gz <username>@login.ifi.uio.no:~/<path>/<to>/<wherever>

It is also a possibility to mount your UiO home directory to the robin-hpc.

Network

For security reasons, the robin-hpc is only accessable from ifis-networks. If you use a desktop machine at ifi, you can access the robin-hpc directly. However, from your own laptop you are required to login to ifi's login cluster, namely login.ifi.uio.no, before you can access hpc.robin.uiocloud.no.

URL for the robin-hpc login node is: hpc.robin.uiocloud.no.

SSH from outside UiOs networks

To access the machine from outside UiOs networks, see Configure SSH. To make SSH slightly more pleasant to work with we can create a configuration file at ~/.ssh/config which can contain additional information about remote machines. The following is a possible configuration for the Robin-hpc (see here for more info). Indiviual key paths can be added with IdentityFile ~/.ssh/<key_on_jump_host>.

Host uio-login
 User <my-username>
 Hostname <my-uio-hostname>
Host hpc-robin
 HostName hpc.robin.uiocloud.no
 ProxyJump uio-login
 User <my-username>

Now, ssh hpc-robin should directly establish a connection to the server (commands like scp should also work in the same way).

Access

Apply for access using this link: https://nettskjema.no/a/robin-hpc. More information on how to get access is then given via e-mail.

SLURM

Ropin-hpc uses slurm for job scheduling.

To start a job on robin-hpc you will need to use a job script specifying the resources you want to use. The job is started with the following command:

sbatch nameofthejobscript.sh

(Note: Please do not run your program on the login node, test your implementation before copying the files to robin-hpc. You can easily copy the files needed to run your program to your home area on robin-hpc with scp.)

Below you will find a template for the job script, followed by some convenient commands.

For more information about slurm please visit https://slurm.schedmd.com/quickstart.html.

SLURM job script template

#!/bin/bash                                                                                                                                                                                                 

# This is a template for a slurm job script.
# To start a job on robin-hpc please use the command "sbatch nameofthisscript.sh".                                                                                                                            

# Job name                                                                                                                                                                                                  
#SBATCH --job-name=nameofmyjob                                                                                                                                                                              

# Wall clock time limit (hh:mm:ss). 
# (Note: The program will be killed when the time limit is reached.)                                                                                                         
#SBATCH --time=01:00:00                                                                                                                                                                                     

# Number of tasks to start in parallel from this script.                                                                                                                                                    
# (i.e. myprogram.py below will be started ntasks times)                                                                                                                                                    
#SBATCH --ntasks=1                                                                                                                                                                                          

# CPUs allocated per task                                                                                                                                                                                   
#SBATCH --cpus-per-task=16                                                                                                                                                                                  

# Memory allocated per cpu                                                                                                                                                                                  
#SBATCH --mem-per-cpu=1G                                                                                                                                                                                    

# Set exit on errors                                                                                                                                                                                        
set -o errexit
set -o nounset

# Load your environment                                                                                                                                                                                     
source myenv/bin/activate

# Run your program with "srun yourcommand"                                                                                                                                                                  
# stdout and stderr will be written to a file "slurm-jobid.out".                                                                                                                                            
# (warning: all tasks will write to the same slurm.out file)                                                                                                                                                
srun python3 myprogram.py

Additional options

To run multiple tasks in sequence you can also use the --array option of the sbatch command. You can add the following to the script. You can then get the job array number with SLURM_ARRAY_TASK_ID. The following example would run 900 jobs:

#SBATCH --array=0-899


Convenient commands

View the status of all jobs:

squeue

View the status of your own jobs:

squeue -u yourusername  

Cancel all your jobs:

scancel -u yourusername

Cancel all jobs with name "job_name":

scancel -n job_name

Software

Matlab

TODO - Ventende oppgave for denne siden:
sebastto
Rediger siden


Setting up the SLURM job script

#SBATCH --job-name=matlab_job
#SBATCH --ntasks=1
#SBATCH --cpus-per-task 16

srun matlab -batch "addpath(genpath('/path/to/your/matlab/folder'));run('myScript.m')"

Running Matlab in batch mode is the most safe option for running MATLAB on a HPC. (From Mathworks documentation)[1]:

-batch statement

  • Starts without the desktop

  • Does not display the splash screen

  • Executes statement

  • Disables changes to preferences

  • Disables toolbox caching

  • Logs text to stdout and stderr

  • Does not display modal dialog boxes

  • Exits automatically with exit code 0 if script executes successfully. Otherwise, MATLAB terminates with a non-zero exit code.

The addpath(genpath('/path/to/your/matlab/folder')) part adds all files in the specified directory to the MATLAB search path. Afterwards we run the main script of your program with run('myScript.m').

Utilizing parallel computing in your MATLAB Script

When the SLURM worker node is setting up your job, a number of environment variables is set. We can use the environment variable SLURM_CPUS_ON_NODE to get the number of CPU cores available in our MATLAB script. In fact, we can use that variable to dynamically select the number of workers in the MATLAB parallel pool, so that your script works both on your own computer and on the HPC.

SLURM_CPUS_STR = getenv('SLURM_CPUS_ON_NODE');

% Delete parallel pool from earlier runs
delete(gcp('nocreate'));

if isempty(SLURM_CPUS_STR) 
    % Run on personal computer (with however many cores your CPU has)
    parpool(6);
else
    % Run on SLURM-scheduled HPC
    SLURM_CPUS_NUM = str2num(SLURM_CPUS_STR);
    parpool(SLURM_CPUS_NUM);
end

Anaconda

After initializing Anaconda (see Deep Learning Workstations), it is possible to start a job on robin-hpc by using the command above and the following template:

SLURM Anaconda template

 #!/usr/bin/bash

 #SBATCH --job-name=<test_name>
 #SBATCH --output=log.txt
 #SBATCH --ntasks=1
 #SBATCH --time=01:00:00
 #SBATCH --mem-per-cpu=100

 # load anaconda and your environment
 source ~/.bashrc
 conda activate <your_environment>

 # Run your program with "srun your_command"     
 srun python3 <script_name>.py

The output is written to "log.txt" and can be traced via "tail -f log.txt".

Podman

alias docker=podman
Personal tools
Front page