METHODS FOR INCREASING SOFTWARE PORTABILITY WITH INCREASED PARALLELIZATION AND PERFORMANCE IN DISTRIBUTED AND MULTICORE SYSTEMS

Abstract

In this thesis, we have solved one of the major practical problems in porting a sequential code on to a multicore platform. Since the exercise of porting to multicore is an expensive and time consuming, it would certainly help to know the speedup one would get. If one has a priori knowledge of possible speedup, one can make an informed decision. The model often used for this is the Amdahl s law based. This method only gives the maximum possible theoretical speedup, and it is often far from what one gets in real life. Thus, the problem is how does one get a better model? In this thesis, we will describe the methodology used for developing a mathematical model, experimentation using benchmark codes and its use for validation of the model, and computation of percentage error in prediction. Our model is called Performance Prediction Model Integrated with Overheads or PPMIO. Before we get into details, let us look at the need for solving the problem. newline