!> Please note that we are effectively treating the design and implementation of SDF as a green field deployment. As such certain technologies and tools may not be available. Please contact us if you wish for anything to be added or removed.
It is not recommended to store your conda environments in your $HOME due to 1) quota limits, and 2) an inability to share conda environments across groups. We generally recommend that you install software into your $GROUP space (eg /sdf/group/<group_name>/sw - please see $GROUP storage).
Download the latest version of Miniconda from the conda website and follow the Instructions. Change the installion prefix to point to an appropriate $GROUP directory:
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O /tmp/Miniconda3-latest-Linux-x86_64.sh
bash /tmp/Miniconda3-latest-Linux-x86_64.sh -p /sdf/group/<group_name>/sw/conda/Replacing <group_name> appropriately.
We can also modify our ~/.condarc file as follows;
channels:
- defaults
- anaconda
- conda-forge
- pytorch
envs_dirs:
- /sdf/group/<group_name/sw/conda/envs
pkgs_dirs:
- /sdf/group/<group_name/sw/conda/pkgs
auto_activate_base: falseConda environments are a nice way of switching between different software versions/packages without multiple conda installs.
There is no unique way to create a conda environment, we illustrate here how to do so from a .yaml file (see the conda documentation for more details).
In order to create an environment called mytest with python=3.6 and numpy and pandas in it, create mytest-environment.yaml:
name: mytest
dependencies:
- python=3.6
- numpy
- pandasThen run the following command: conda env create -f mytest-environment.yaml.
If successful, you should see mytest when listing your environments: conda env list.
You can now activate your environment and use it: conda activate test. To double-check that you have the right packages, you can type conda list once in the environment and check that you see numpy and pandas.
For a detailed explanation of how to use Singularity on SDF, you can go through Yee-Ting Li's presentation here where you can find both the slides and the zoom recording.
As pulling and building a new image can use quite a lot of disk space, we recommend that you set the appropriate cache paths for singularity to not use your $HOME directory. Therefore, before pulling or building an image, define the following environment variables:
export DIR=/scratch/${USER}/.singularity
mkdir $DIR -p
export SINGULARITY_LOCALCACHEDIR=$DIR
export SINGULARITY_CACHEDIR=$DIR
export SINGULARITY_TMPDIR=$DIRTo pull an image from DockerHub, do:
singularity pull docker://<user or organization>/<repository>:<tag>We offer various of GCC through (Redhat Software Collections](https://developers.redhat.com/products/softwarecollections/overview). These are wrapped up in Environment Modules and are available with a module load devtoolset/<version>.
We also provide symlinks to the appropriate GCC version to the above devtoolset packages with module load gcc/<version>.
We provide tunnel-less access to jupyter instances running on SDF via an easy to use web portal where you can 'bring-your-own' jupyter environments. See Interactive Jupyter for more information.
The first step is to set up matlab for your account. These steps will create a .matlab directory under your account with our site's configuration. You only need to do this once per version of MATLAB.
ssh sdf-login
module load matlab
matlab
configCluster
exit
Slurm batch submissions of Matlab actually start up 2 jobs. The initial job is just to bring up matlab which, once successful, submits a second job which runs your matlab code.
If you do not have access to a priority partition or you don't specify a partition, your jobs will run in shared. If you do have access and want to run your jobs in a priority partition then you need to specify the partition in both the sbatch script which launches matlab and as a matlab setting in the matlab job. You can set the partition for the matlab job either in the matlab job itself which you will then need to do everytime you submit. For example, if you have access to the spearx partition, your matlab code would begin:
c=parcluster;
c.AdditionalProperties.QueueName='spearx';
p=c.parpool(200);
or you can save the partition to your matlab profile like so:
module load matlab
matlab
c=parcluster
c.AdditionalProperties.QueueName='spearx'
c.saveProfile
and then you won't need to specify it in any future matlab scripts.
You may need to increase the ulimit for processes in your .bashrc depending on the size of the parpool you need. The default is 4096 and too small for say a parpool of 100. Include something like
ulimit -u 19200
Following are an example sbatch and a test.m matlab file which can be used to test after you have followed the steps to set up your environment.
Create the matlab job test.m:
c=parcluster;
p=c.parpool(100);
tic
n = 200;
A = 500;
a = zeros(1,n);
parfor i = 1:n
a(i) = max(abs(eig(rand(A))));
end
rtime = toc;
fileID = fopen('readtime_out.txt','w');
fprintf(fileID,'%f',rtime);
fclose(fileID);
p.delete
Create the file testrun.sbatch and modify the "matlab -batch" line to add the path to the directory containing your test.m file. It is invoked as just "test" without the ".m".
#!/bin/sh
#SBATCH --partition=spearx
#SBATCH --job-name=test
#SBATCH --output=output-%j.txt
#SBATCH --error=error-%j.txt
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=2
#SBATCH --mem-per-cpu=4g
#SBATCH --time=5:00
source /usr/share/Modules/init/sh
module load matlab
matlab -batch "addpath ('/sdf/home/v/vanilla/matlab');test"
Note that this submission script only needs to request enough resources to launch matlab. Matlab will figure out how many cores it needs from the parpool request, or you can give matlab more specifics about nodes and processes as described in the next section.
Run the job:
[vanilla@sdf-login02]~% sbatch testrun.sbatch
Submitted batch job 722162
[vanilla@sdf-login02]~%
[vanilla@sdf-login02]~%
[vanilla@sdf-login02]~%
[vanilla@sdf-login02]~%
[vanilla@sdf-login02]~% squeue | grep vanilla
722180 spearx Job1 vanilla R 0:14 2 rome[0244,0251]
722179 spearx test vanilla R 0:42 1 rome0133
In this example, jobid 722179 is the ntasks=1 job that starts up matlab and 722180 is the job that actually runs test.m. You should end up with two files in the location pointed to by addpath, the slurm output-jobid.txt file with the details about the matlab submission, and the output file specified in the matlab job that has the results of your computation:
-rw-r--r-- 1 vanilla sf 591 Jun 29 11:27 output-722179.txt
-rw-r--r-- 1 vanilla sf 8 Jun 29 11:27 readtime_out.txt
[vanilla@sdf-login02]~% cat output-722179.txt
Non-Degree Granting Education License -- for use at non-degree granting, nonprofit,
educational organizations only. Not for government, commercial, or other organizational use.
Starting parallel pool (parpool) using the 'sdf R2020b' profile ...
additionalSubmitArgs =
'--ntasks=100 --cpus-per-task=1 --ntasks-per-core=1 -p spearx'
Connected to the parallel pool (number of workers: 100).
Parallel pool using the 'sdf R2020b' profile is shutting down.
[root@sdf-login01 ~vanilla]# cat readtime_out.txt
4.291791
You can manage the number of nodes when running your parallel server jobs. Our rome nodes have 128 cores. If you don't tell matlab to use some certain number of nodes it will spread the job out and use whatever cores it finds available. It may be preferrable to restrict matlab jobs to fewer or even exclusive nodes and avoid any contention that may come from other jobs.
To restrict your matlab job which is requesting a parpool of 230 to 2 nodes, you would add this to your matlab job:
c.AdditionalProperties.NumNodes=2;
c.AdditionalProperties.ProcsPerNode = 116;
This tells matlab to run on 2 nodes, and the procspernode should be the amount of the parpool, in this case 230, divided by the number of nodes, and then add one, which in this case would be 230/2 and add one 115+1=116.
If you are running in a priority partition, and you want your jobs to run exclusvely even if your parpool doesn't require all the cores, you can add:
c.AdditionalProperties.AdditionalSubmitArgs = '--exclusive';
to your matlab job, but still keep the NumNodes and ProcsPerNode lines.
Parallel computing documentation: https://www.mathworks.com/help/parallel-computing/index.html
Parallel computing coding examples: https://www.mathworks.com/products/parallel-computing.html
Parallel computing tutorials: https://www.mathworks.com/videos/series/parallel-and-gpu-computing-tutorials-97719.html
Parallel computing videos: https://www.mathworks.com/videos/search.html?q=&fq[]=product:DM&page=1
Parallel computing webinars: https://www.mathworks.com/videos/search.html?q=&fq[]=product:DM&fq[]=video-external-category:recwebinar&page=1