weighted_dec0_telescope_probe Namespace Reference

Functions
	sphere_sdf (o, sz, n, per_axis=2, rfrac=0.35)
	bed_weights (n, bed, lam, tilt=0.0, seed=7)
	blocks (comm, s)
	describe (tag, comm, blks, w, n)
	run_steps (comm, s, k)
	main ()

Detailed Description

SCALING_ISSUES #2 TRAP probe — does a WEIGHTED level-0 decomposition collapse the pressure
multigrid's telescope to one rank?

The claim (read from `src/mac_cutcell_mg.hpp`, not measured when it was written): after
`diagnostics.rebalance_by_weights(w)` the level-0 blocks are no longer even on the coarsenable
axes, the telescope fires at L = 0, and the agglomeration-depth search falls through to d = 0 —
the whole level-0 residual gathered onto ONE rank every V-cycle.

This measures it, flow-only (the coupling's `CfdDem.rebalance()` reaches the same call):

  1. build a distributed Solver on the default equal-weight aligned ORB, run a few steps;
  2. record the per-rank blocks (origin, size, parity) and the pressure iterations / step time;
  3. `rebalance_by_weights(w)` with a particle-bed-like weight (1 + Poisson counts in the bottom
     `--bed` fraction of z), record the same again.

The MG ladder (levels, ranks per level, which transitions TELESCOPE) is printed by flow itself on
rank 0 at every hierarchy build when PECLET_FLOW_MG_DEBUG=1 (a trace-only env var; it changes no
result). The ranks column of the level after a telescope point is the size of the roots'
sub-communicator, i.e. 2^d for the depth d the search selected (np a power of two).

    PECLET_FLOW_MG_DEBUG=1 OMP_NUM_THREADS=2 OMP_PROC_BIND=false \
      /usr/bin/mpirun -np 8 --map-by slot:PE=2 --bind-to core python tests/study/weighted_dec0_telescope_probe.py --n 96 --levels 4

Function Documentation

sphere_sdf()

weighted_dec0_telescope_probe.sphere_sdf	(		o,
			sz,
			n,
			per_axis = 2,
			rfrac = 0.35
	)

Periodic simple-cubic array of spheres (per_axis^3 of them), sampled at cell centres of this
rank's block (cell units, origin 0). Negative inside the solid.

Definition at line 35 of file weighted_dec0_telescope_probe.py.

Referenced by main().

bed_weights()

weighted_dec0_telescope_probe.bed_weights	(		n,
			bed,
			lam,
			tilt = 0.0,
			seed = 7
	)

1 + Poisson(lam) particle counts below a bed surface, 1 above — a settled bed, as
`CfdDem.rebalance(gamma=1)` would build it (1 + gamma * count). tilt = 0 is a FLAT bed at
z = bed*n (uniform in x and y, so the x/y ORB splits stay at the equal-cell positions); tilt > 0
is a HEAP whose surface falls linearly in x and y, z_s = bed*n*(1 + tilt*(1/2 - x/n) +
tilt*(1/2 - y/n)), which moves every split. Global, x-fastest (Fortran ravel), identical on
every rank (same seed).

Definition at line 50 of file weighted_dec0_telescope_probe.py.

Referenced by main().

blocks()

weighted_dec0_telescope_probe.blocks	(		comm,
			s
	)

Definition at line 65 of file weighted_dec0_telescope_probe.py.

Referenced by main().

describe()

weighted_dec0_telescope_probe.describe	(		tag,
			comm,
			blks,
			w,
			n
	)

Definition at line 71 of file weighted_dec0_telescope_probe.py.

Referenced by main().

run_steps()

weighted_dec0_telescope_probe.run_steps	(		comm,
			s,
			k
	)

Per step: pressure iterations, wall time (barrier to barrier), and the max over ranks of the
solver's own projection and momentum timers (`last_step_timers`).

Definition at line 92 of file weighted_dec0_telescope_probe.py.

Referenced by main().

main()

weighted_dec0_telescope_probe.main

(

)

Definition at line 109 of file weighted_dec0_telescope_probe.py.

References bed_weights(), blocks(), describe(), main(), run_steps(), and sphere_sdf().

Referenced by main().