Hello Daniel,
I think the answer to your question comes down to a combination of workflow, convenience, and licensing. In many circumstances, it's perfectly valid and even recommended to use a client MATLAB session on a compute node. As an example, let's say that you have access to a university HPC cluster that has compute nodes with 128 cores each and access to a Campus-Wide License. You may even be able to access the GUI MATLAB desktop through something like Open OnDemand. If your parallel processing computation fits onto that single node, using the local Process pool would be ideal for your needs.
One benefit of adding Parallel Server to the mix is remote job submission. With this setup, you can have MATLAB installed on your local machine. As long as the proper configuration is in place, you can submit remote batch jobs from your machine to the cluster without needing to be heavily familiar with SSH, Linux, and HPC schedulers. You can submit several batch jobs to the HPC queue, close your local MATLAB session, then return at a later time to fetch your results. Even if these are all single node jobs, it gives you some additional flexibility in how you use the cluster.
Licensing is another consideration that may make single-node jobs better with Parallel Server. Without going too far into the weeds, client MATLAB sessions pull a license seat for MATLAB and every Toolbox used in your code, while instead MATLAB Parallel Server pulls a license seat for every worker used in your job. Often it makes more sense to use dedicated Parallel Server license seats rather than client MATLAB seats, though this does depend on your particular environment.
Overall, both are valid ways of running MATLAB code on an HPC cluster and my exact recommendation on which to use would depend on your particular circumstances.
Let me know if that helps and if there's anything else I can elaborate on!
-Damian
