Hello,
This is not very surprising, as 15000 elements is a rather small mesh. However our tests with many more processors on very big meshes show that the speed up can be still interesting up to a few hundreds of points per sub-domain. You can try the Malpasset test case in 2D (26000 elements), which can run in about 4 s on a 12-core machine (about 10 s here at EDF on a 8-core Linux). It could be that clusters are not so optimised.
With best regards,
Jean-Michel Hervouet