SergeAS

Computational Core of my oldest 2D NS-solver now open source (LGPLv3),
meet OpenHyperFLOW2D
(include serial and parallel MPI,OpenMP and CUDA versions)

Dependence of efficiency on the number of cores is a very important characteristic of scalability solver based on which one can predict the scaling limit. For simplicity we consider only one option for the solver, using the asynchronous version of the MPI exchanges Isend() + Irecv(). It is seen that the dependence of the efficiency of parallelization of the solver is linear or nearly linear, with a gradual drop in effectivity with increasing number of cores used
...

...use a blocking call Sendrecv() instead of a pair of Send() + Recv() slightly improves the performance of the code but does not result to improved scalability
...

Let us consider in more detail how works domain decomposition in the simplest 1D case.
The computational domain is divided into several subdomains along one (usually the longest) dimension. The number of such subdomains usually corresponds to the number cores involved. The main data exchanges take place between neighboring subdomains. The size of transmitted data between adjacent subdomains in the general case depends on many factors (number of equations in the system, the size of the computational...

Looking the latest HPCSource I came across very well-known picture, illustrating the Amdahl's law:



which shows the dependence of the speedup parallel code on number of processor cores for different fraction of time () spent in the part that was not parallelized. This diagram is interested in me for the...