I want to measure the efficiency of the paralleling for different mesh sizes. My problem is quite simple: a flow in a tube divided in 6 domains by METIS (the 6 processor are on the same machine). It is quite surprising for me to see that when one augments the refinement, the efficiency decreases after an optimum. I was thinking that the efficiency increases in such a case.
The efficiency is here measured as: TCPU(1) / ( 6 * TCPU(6) ), where TCPU(N) is the CPU time for parallel computations on N processors.
I thought first it was linked to the fieldTransfer=1 (fieldTransfer are inverted on the attached graph), so I change it. The results are a little closer to what expected, but I still observe the general behavior depicted above. Finally, maybe are my results normal, but I have no explanation for them. What do you think about this ?
|All times are GMT -4. The time now is 15:49.|