For offload, thread stack size (default 2MB) is set by MIC_STACKSIZE:
https://software.intel.com/sites/products/documentation/doclib/iss/2013/compiler/cpp-lin/GUID-17EE519D-7D60-4294-B87F-87E5F51EFF2C.htm
For processes running in a shell on the MIC (e.g. MPI), you must set primary stacksize in that shell e.g.
ulimit -s unlimited
as well as possibly adjusting thread stack size (default 4MB) by OMP_STACKSIZE or KMP_STACKSIZE (which over-rides OMP_STACKSIZE). Raising thread stack size is almost certain to require unlimited primary stack (and may still limit the number of threads which can be used effectively). For example, I used 30 threads in each of 6 MPI ranks with OMP_STACKSIZE=9m.
Clearly, you must not set MIC_STACKSIZE or OMP_STACKSIZE to a ridiculous value. Nothing stops you from tying the coprocessor in knots by attempting a total of 200GB stack.
When I last heard (well over a year ago), there was no expert consensus on changing anything in the scheme, although there were opinions that something should be done (such as making ulimit -s unlimited the default). Note that the newest information I could find by browser search is that old.
