Parallel Processing
Written by Ehsan Roohi Golkhatmi   


We investigated the efficiency of a parallel direct simulation Monte Carlo (PDSMC) algorithm in solving the rarefied subsonic flow through a nanochannel. We use MPI library to transfer data between the processors. It is observed that PDSMC solver shows ideal speed up if sufficient workload is provided for each of processors. Additionally, this study shows that the computational time and speed up of the extended PDSMC solver do not depend (or slightly depend) on the number of cells. In contrary, increasing the total number of particles would result in a better efficiency of the PDSMC.
Our test case is the nitrogen flow in a nanochannel having an aspect ratio of 100. The grid consists of 240 by 30 rectangular cells. Around 25 million particles are filled in the domain. Figure 1 shows the speed up of PDSMC solver using up to 80 processors. According to Fig. 1, speed up is almost linear up to 24 processors. However, speed up decreases to 32 for 40 processors and 44 for 80 processors, respectively. High speed up could be achieved if the number of particles, i.e., workloads, is sufficiently large for each processor. However, the speed up and efficiency decrease as soon as the communication overload increases compared with the processors computational workload. It means that PDSMC solver could be used in large scale problems more efficiently.




Fig. 1. Parallel Performance

    ResearchGate  ResearchGate  ResearchGate