Optimised external computation for the Euro50 MATLAB based integrated model

In previous work we have countered computational demands faced in integrated modelling by developing and using a parallel toolkit for MATLAB. However the use of an increasingly realistic model makes the computational requirements of the model much larger, particularly in wavefront sensing, reaching a point where simulations of several real time seconds were no longer practical taking up to 3 weeks per second. In response to this problem we have developed optimised C code to which MATLAB off loads computation. This code has numerous advantages over native MATLAB computation. It is portable, scaleable using OpenMP directives and can run remotely using Remote Procedure Calls (RPCs). It has opened up the possibility of exploiting high end Itanium and Opteron based shared memory systems, optimised 3rd party libraries and aggressive compiler optimisation. These factors combined with hand-tuning give a performance increase of the order of 100 times. The interface to the rest of the model remains the same so the overall structure is unchanged. In addition we have developed a similar system based on Message Passing Interface version 2, (MPI-2) which allows us to exploited clusters. Here we present an analysis of techniques used and gains obtained along with a brief discussion of future work.