매니코어용 마이크로 커널 상의 MPI 응용프로그램 성능 평가

Abstract

The multi-core architecture term appears due to the limitation of the single-core architecture which the improvement of clock frequency is difficult to be gained. A multi-core processor is typically made by two or four independent processor cores in a single chip which are connected through an on-chip bus. Because of the number of cores are increasing from 4 to 8, 16, and so on, a many-core term is appeared. Therefore a system which uses many-core architecture is supposed to be highly parallel. A parallel program makes a task into threads or processes which it should be spread across multiple cores. In microkernel-based systems, the communication of each process is assisted by Inter Process Communication (IPC). The message-passing allows the processes to communicate with each other using IPC. The parallel programming that uses message passing standard is called Message Passing Interface (MPI). In this thesis, MPI library is ported to the L4Linux on top of L4/Fiasco microkernel over many-core architecture. The MPICH is chosen to be used as the MPI library routines. The performance comparisons from three different scenarios have been done to check the feasibility of MPI application on microkernel OS. From the benchmarking application, we observed how MPI library takes influence in the communication between processes which occurs in the many-core system. The performance improvement is shown by the matrix multiplication test program that proves the feasibility of MPI on microkernel OS over many-core system.