It is possible to compile an application on MATLAB and then submit that application to run on the cluster. The documentation is here and there is an example of the workflow here.
Note also that there are other ways to tackle the problem, too. For example, if you have a common filesystem that the Hadoop cluster nodes have access to, you could place a MATLAB installation on the common filesystem and set the ClusterMatlabRoot property to use that, no local installation required -- but this also makes a lot of assuptions about your computing environment -- that you have a common filesystem, that you have the network capacity to run MATLAB over the network at your scale, etc.
Another approach is that you can install only the toolboxes you require. MDCS is recommended to be installed with all toolboxes so that any arbitrary command from any arbitrary client will have access to the same functions on the cluster and avoid "function not found" errors. If you know that you only use functions from a few toolboxes, you can reduce the installation size by using that smaller installation, but this takes more tending and upkeep than just installing everything.