HPC cluster

Kindly assist me with the following problems: 1) How does Matlab partition large matrices and what does the algorithm look like? 2) I intend to use configuration manager to setup job submission for a HPC cluster (with a head node (multicore), 5 compute nodes (multicore) and some client nodes). Please on which of the nodes do I need to install Matlab (with parallel toolbox)? Do I need to install parallel Matlab on all the nodes? On which of the node (matlab) do I setup the configuration manager? Answer to these questions will help me to know the number of licenses to buy.

Thank you.

Answer 1

[Edric Ellis]

The partitioning of distributed arrays is documented here, although you can choose the partition differently if you wish. (There's not often a good reason to do this).

You need to install MATLAB + Parallel Computing Toolbox on the nodes from which you wish to submit jobs. You also launch the configuration manager on that node.

You need to install MATLAB Distributed Computing Server on the worker nodes. If you're going to use the MathWorks jobmanager on the head-node, you need to install MDCS there too - but the jobmanager process does not consume any licences.

[Margaret]

As one cannot connect to the outside network from most compute nodes, how do you set up the license file?

[Thomas]

If I remember right THE DCS handles the license manager and the compute nodes do not have to worry about the license .. Only your head node (on which the DCS is running) connects to the FLexLM license manager and checks out a DCS license for how many ever nodes of DCS license you have..