Preliminary results from a parallel MATLAB compiler

We are developing a compiler that translates ordinary MATLAB scripts into code suitable for compilation and execution on parallel computers supporting C and the MPI message-passing library. We report the speedup achieved for several MATLAB scripts on three diverse parallel architectures: a distributed-memory multicomputer (Meiko CS-2); a symmetric multiprocessor (Sun Enterprise Server 4000); and a cluster of symmetric multiprocessors (Sun SPARCserver 20s). By generating code suitable for execution on parallel computers, our system multiplies the gains achievable by compiling, rather than interpreting, MATLAB scripts. Generating parallel code has an additional advantage: the amount of primary memory available on most parallel computers makes it possible to solve problems too large to solve on a single workstation.