MPI runs
A post-processing script started with mpirun -n 4 python script.py runs every
line on every rank. Without care, four processes would draw the same figure and
write the same file at once. plasma-plots draws and saves on rank 0 only, so
the same script runs in serial and under MPI:
out = Output("sim_1") # on every rankrate = energy.plasma.analysis.growth_rate(window=(5.0, 20.0)) # on every rankout.plot.scalars().save("scalars.png") # drawn and written once, by rank 0Analysis runs on every rank, so every rank has the numbers. On the other ranks
a plot returns a SkippedPlot, whose methods (save, show, to_plotly, …)
do nothing, so chained calls need no checks.
Which rank plots
Section titled “Which rank plots”The rank is read without importing mpi4py, which would initialize MPI in a
serial run: from mpi4py.MPI.COMM_WORLD when the application has already
initialized MPI, otherwise from the variables the MPI launchers export (Open MPI,
MPICH, Intel MPI, Slurm, PMIx, Cray). STRUPHY_MPI=0 turns the detection off,
and every process plots.
import plasma_plots
plasma_plots.mpi_rank() # 0 in a serial runif plasma_plots.is_plotting_rank(): print(f"growth rate {rate.rate:.3e}")Nothing waits for rank 0. A rank that reads a file rank 0 writes must
synchronize first, e.g. with MPI.COMM_WORLD.Barrier().
Plot functions of your own
Section titled “Plot functions of your own”rank_zero makes a function of your own
behave the same way: it runs on rank 0 and returns a SkippedPlot elsewhere.
from plasma_plots.mpi import rank_zero
@rank_zerodef plot_beam(markers): return markers.plasma.plot.phase_space(x="x", v="vx", t=-1)
plot_beam(markers).save("beam.png") # safe on every rankWhere the ranks meet
Section titled “Where the ranks meet”A distributed array is split into blocks, one per rank. Drawing the block boundaries shows where a jump or a noise pattern sits on a rank boundary:
# profiles: dashed lines at explicit values, or from an mpiarray Layoutphi.plasma.plot.lineout(x="x", t=-1, rank_boundaries=layout)# slices: the boundaries of the blocks as grid linesphi.plasma.plot.slice(x="x", y="y", t=-1, overlays={"rank_boundaries": layout})# matrices: where each rank's rows begin, e.g. PETSc's ownership rangesplot_matrix(matrix, rank_boundaries=matrix["ownership_ranges"])A Layout of mpiarray is used duck-typed, through its boundary_values, so
plasma-plots does not import mpiarray.
Profiling
Section titled “Profiling”Timing summaries and profiling plots of scope-profiler output, including one row per rank in the Gantt chart, are in the Profiling guide.