Session-UPPMAX: Matlab client on the desktop

Use own computer’s matlab

Would you like to try run batch jobs on the Rackham or Snowy cluster but use the faster graphics that you can achieve on your own computer?
Do you have all your work locally but sometimes need the cluster to do parallel runs?
UPPMAX offers this now.

Warning

This solution is possible only if
- you have an UPPMAX compute project
- a working matlab on your computer with one of the version available on the cluster:
  check with module avail matlab
  
  Examples of the newest ones:
  
  R2020b
  
  R2022a
  
  R2022b
  
  R2023a
  
  R2023b

Want a video?

UPPMAX Matlab client on the desktop (Oct 2024)

Let’s get started together!

Demo

You can try to type-along

The Rackham MATLAB support package can be found at uppsala.Desktop.zip.

Download the ZIP file and start MATLAB.
The ZIP file should be unzipped in the location returned by calling

>> userpath

You can unzip from MATLAB’s Command window.
Configure MATLAB to run parallel jobs on the cluster by calling configCluster. configCluster only needs to be called once per version of MATLAB.

>> configCluster
Username on RACKHAM (e.g. jdoe):

Type your rackham user name.
As a result:

Complete.  Default cluster profile set to "Rackham R2022b".

Note

To submit jobs to the local machine instead of the cluster, run the following:

>> % Get a handle to the local resources
>> c = parcluster('local');

Configuring Slurm details

Prior to submitting the job, various parameters can be assigned, such as queue, e-mail, walltime, etc. The following is a partial list of parameters. See AdditionalProperties for the complete list. Only AccountName, Partition, MemUsage and WallTime.

>> % Get a handle to the cluster
>> c = parcluster;

c =

 Generic Cluster

    Properties:

                   Profile: Rackham R2022b
                  Modified: false
                      Host: UUC-4GM8L33.user.uu.se
                NumWorkers: 100000
                NumThreads: 1

     JobStorageLocation: <path to job outputs locally>
      ClusterMatlabRoot: /sw/apps/matlab/x86_64/R2022b
        OperatingSystem: unix

Set some additional parameters related to Slurm on Rackham

>> % Specify the account
>> c.AdditionalProperties.AccountName = 'uppmax2025-2-272';

>> % Specify the wall time (e.g., 1 day, 5 hours, 30 minutes
>> c.AdditionalProperties.WallTime = '00:30:00';

>> % Specify cores per node
>> c.AdditionalProperties.ProcsPerNode = 20;


[OPTIONAL]

>> % Specify the partition
>> c.AdditionalProperties.Partition = 'devcore';

>> % Specify another cluster: 'snowy'
>> c.AdditionalProperties.ClusterName='snowy'
>> c.AdditionalProperties.ProcsPerNode = 16;

>> % Specify number of GPUs
>> c.AdditionalProperties.GPUsPerNode = 1;
>> c.AdditionalProperties.GPUCard = 'gpu-card';

>> c.saveProfile

To see the values of the current configuration options, display AdditionalProperties.

>> % To view current properties
>> c.AdditionalProperties

Unset a value when no longer needed.

>> % Example Turn off email notifications
>> c.AdditionalProperties.EmailAddress = '';
>> c.saveProfile

Start job

Copy this script and paste in a new file parallel_example_local.m that you save in the working directory where you are (check with pwd in the Matlab Command Window).
- The script is supposed to loop over sleepTime seconds of work nLoopIters times.
- We will define the number of processes in the batch submit line.

function t = parallel_example_local(nLoopIters, sleepTime)
t0 = tic;
parfor idx = 1:nLoopIters
   A(idx) = idx;
   pause(sleepTime);
end
t = toc(t0);

job = c.batch(@parallel_example_local, 1, {16,1}, 'Pool',8,'CurrentFolder','.');

Submission to the cluster requires SSH credentials.
You will be prompted for username and password or identity file (private key).
- It will not ask again until you define a new cluster handle c or in next session.

Jobs will now default to the cluster rather than submit to the local machine.

>> job.State

ans =

    'running'

You can run this several times until it gives

>> job.State

ans =

    'finished'

You can also watch queue

Or on Rackham (it really runs there!):

[bjornc2@rackham2 ~]$ squeue -u bjornc2
          JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
       50827312   devcore MATLAB_R  bjornc2  R       2:20      1 r483

>> job.fetchOutputs{:}

ans =

    2.4853

The script looped over 1 s work 16 times, but with 8 processes.
In an ideal world it would have taken 16 / 8 = 2 s. Now it took 2.5 s with some “overhead”

Run on Snowy

>> c.AdditionalProperties.ClusterName='snowy'
>> c.AdditionalProperties.ProcsPerNode = 16;

Exercises

1. Configure your local Matlab to talk to UPPMAX

Use the instructions above to try to make it work!

2. Run a script on snowy

Try to run a script from the MATLAB GUI and SLURM session
Check in a rackham terminal: squeue -M snowy --me

Keypoints

Steps to configure first time
- download and decompress UPPMAX configure file.
- run configCluster on local MATLAB and set user name
Steps to run
- set parcluster settings, like you do otherwise.
Note: only parcluster will work, not parpool.

Do you get problems?

Send a support ticket to UPPMAX via supr.naiss.se/support

Extra

Helper functions
Function	Description	Applies Only to Desktop
clusterFeatures	List of cluster features/constraints	—
clusterGpuCards	List of cluster GPU cards	—
clusterPartitionNames	List of cluster partition	—
disableArchiving	Modify file archiving to resolve file mirroring issue	true
fixConnection	Reestablish cluster connection (e.g., after reconnection of VPN)	true
willRun	Explain why job is queued	—