Could it be that the system call duplicates everything that is in the virtual memory to start the sister process?

In theory, fork (used to implement system) does exactly that. Modern kernels with virtual memory will set up COW instead of actually copying the entire address space, but this still (usually) requires duplicating page tables, which for 256GiB of data with 4KiB pages themselves fill 512MiB or so. Could you be bumping up against a resource limit? (Look up ulimit for more information.)