Non-Blocking message passing pada MPI dengan implementasi pada metode preconditioned conjugate gradient


A limitation of many parallel algorithms implemented in a distributed memory architecture is the existence of dependency on non-local data with synchronous communication. Some times the asynchronous communication can be applied to reduce communication cost, i.e. the computation is carried out while the communication is in progress. This study analizes possibility to increase the performance of a parallel system through the implementation of asynchronous communication which is applied to Preconditioned Conjugate Gradient (PCG) method. Basically, PCG method has interdependence among steps so that the synchronous communication is more appropriate. Through reformulation of that method, however, the interdepence among steps can be reduced so that asynchronous communication can be applied. From the experiment, compared to the implementation in synchronous communication. Keywords: asynchroronous communication, parallel efficiency, preconditioned conjugate gradient method. Supeno Djanali dan Imam Machdi; Jurusan Teknik Informatika FTIF, ITS