Um estudo sobre o desenvolvimento de aplicações com criação dinâmica de processos em MPI

Abstract

The MPI standard is widely used in the development of parallel programs for distributed memory environments. It is divided into two standards: MPI-1 and MPI-2. The first is characterized by creating processes statically at the start of execution of the application, while in the second the creation of a process takes place dynamically, ie at runtime. However, processes create dynamically generates an extra cost to implement, when compared with static setting. This paper makes a study on the development of applications with dynamic creation of processes, which aims to investigate the impact of this overhead in MPI-2 applications and how it can be bypassed through the use of efficient programming techniques and advanced features provided by the MPI. For this, applications have been implemented with MPI-1 and MPI-2, under two problems (Game of Life and Skyline Matrix Solver) of the suite Cowichan Problems. Our first analysis performed on the Game of Life, it was possible to find the overhead of the dynamic creation of processes in applications with regular workload and whose computation time is similar in both cases. The second analysis performed on the Skyline Matrix Solver possible to analyze the impact of process creation at runtime in applications that have irregular workload and different computation time between processes. Our results show that for applications with regular workload and with little computing time (similar to the Game of Life), at best, the inclusion of dynamic creation of processes impacted by just 0, 2% in the total time of application. For applications with irregular work load considerably useful and time of computation (similar to Skyline Matrix Solver), we show that it is possible to succeed with the creation of dynamic processes, wherein at best, with the implementation of MPI-2 was 6% more efficient than the MPI-1 implementation version.