Step 4: Checking Parallel Performance

  Generally, you will use the parallel solver for large, computationally intensive problems, and you will want to check the parallel performance to determine if any optimization is required. Although the example in this tutorial is a simple 3D case, you will check the parallel performance as an exercise.

For details, see this chapter in the separate User's Guide.

Parallel $\rightarrow$ Timer $\rightarrow$ Usage

Performance Timer for 176 iterations on 2 compute nodes
  Average wall-clock time per iteration:              0.141 sec
  Global reductions per iteration:                      147 ops
  Global reductions time per iteration:               0.000 sec (0.0)
  Message count per iteration:                          383 messages
  Data transfer per iteration:                        0.217 MB
  LE solves per iteration:                                7 solves
  LE wall-clock time per iteration:                   0.030 sec (21.2)
  LE global solves per iteration:                         2 solves
  LE global wall-clock time per iteration:            0.000 sec (0.0)
  LE global matrix maximum size:                        11
  AMG cycles per iteration:                          12.506 cycles
  Relaxation sweeps per iteration:                      314 sweeps
  Relaxation exchanges per iteration:                   146 exchanges

  Total wall-clock time:                             24.866 sec
  Total CPU time:                                    49.813 sec

  The most accurate way to evaluate parallel performance is by running the same parallel problem on $1$ CPU and on $n$ CPUs, and comparing the Total wall-clock time (elapsed time for the iterations) in both cases. Ideally you would want to have the Total wall-clock time with $n$ CPUs be $1/n$ times the Total wall-clock time with $1$ CPU. In practice, this improvement will be reduced by the performance of the communication subsystem of your hardware, and the overhead of the parallel process itself. As a rough estimate of parallel performance, you can compare the Total wall-clock time with the Total CPU time. In this case, the CPU time was approximately twice the Total wall-clock time. For a parallel process run on two compute nodes, this reveals very good parallel performance, even though the advantage over a serial calculation is small, as expected for this simple 3D problem.

Note:    The wall clock time, the CPU time, and the ratio of iterations to convergence time may differ depending on the type of computer you are running (e.g., Windows32, Linux 64, etc.).