We have developed a compiler, nicknamed "Otter", which translates ordinary MATLAB scripts into C programs targeting parallel computers supporting ScaLAPACK and several other parallel numerical libraries. We outline the phases of the multi-pass compiler and we discuss the design and implementation of the runtime library. We present the performance achieved by ten compiled benchmark scripts on four distinct parallel platforms, relative to the speed of the MATLAB interpreter executing these scripts on a single CPU. Finally, we illustrate how different data distributions and block sizes can affect the performance of a compiled code.