WUT_Computer_Science/Programming/PORR/code/richardson_mpi.py
Krzysztof kuhy Rudnicki faef20dc45 PORR: zero findings, 18 tests still passing after every refactor
598 findings to none, and the suite ran through `./run.sh run` after each step
rather than at the end -- 18 passed, 27s, every time.

The shape of this one was different from the other courses. PORR is four
implementations of the same linear algebra (sequential, threads, processes,
dask), so nearly every finding came in groups of four:

- 229 missing annotations. Rather than `object`, which mypy strict rejects on
  every attribute access, each module declares what these actually are:
  `Matrix = np.ndarray | list[list[float]]` and the vector equivalent. The
  backends really are called with both -- comparing them is the point of the
  course.
- 76 names: `A` -> matrix, `Ab` -> augmented, `Ax` -> product, `I` ->
  identity, `dA` -> dask_matrix, and the two CamelCase functions. Renamed
  through `tokenize`, not regex, so no string or comment was touched.
- 121 missing docstrings, one per method name since the names repeat across
  all four backends.
- Seven functions over the argument limit, all taking the same seven or nine
  loose values. They are two frozen dataclasses now, in a new
  richardson_problem.py: `Problem(matrix, rhs, lambda_min, lambda_max)` and
  `Settings(max_iterations, x0, tol)`. `RichardsonMethod.__init__` also took a
  `size` argument it never read.
- threads_indep_old's iteration did "split into chunks, submit, wait" four
  times by hand (C901 11, PLR0915 54). One `_run_chunked` helper.
- tests.py's two bare asserts became explicit raises, per the no-ignores rule.
  `assert False` was also B011: `python -O` deletes it, so the non-convergence
  branch would have silently passed under optimisation.

Real fixes rather than style: np.random -> Generator, `os.path.exists` ->
Path, an `except Exception` around dask narrowed to the errors it actually
raises, and two runnable modules that logged at INFO without ever calling
basicConfig -- caught by scripts/check_refactor_safety.py, which is now doing
the job it was written for.

Repo-wide: 1,781 findings -> 1,183, all of them in the five courses left.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172dj34FDpao6wsix8ZuCzE
2026-08-28 21:01:13 +02:00

180 lines
5.7 KiB
Python
Executable File

#!/usr/bin/env python3
"""Richardson iteration split across MPI ranks."""
import logging
import time
import numpy as np
from mpi4py import MPI
from richardson_problem import Problem, Settings
# These helpers are called with numpy arrays in some backends and plain
# Python lists in others, which is the point of the comparison this course
# is about, so the aliases name both.
Matrix = np.ndarray | list[list[float]]
Vector = np.ndarray | list[float]
logger = logging.getLogger(__name__)
def richardson_parallel(
problem: Problem, settings: Settings | None = None
) -> tuple[Vector, float]:
"""Run Richardson across MPI ranks."""
if settings is None:
settings = Settings(max_iterations=1000)
matrix = problem.matrix
b = problem.rhs
lambda_min = problem.lambda_min
lambda_max = problem.lambda_max
tol = settings.tol
max_iter = settings.max_iterations
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
# Rozmiar macierzy A
n = matrix.shape[0]
# Obliczanie wartości własnych tylko na jednym procesie
omega = 2 / (lambda_min + lambda_max) if rank == 0 else None
# Rozgłoszenie omega do wszystkich procesów
omega = comm.bcast(omega, root=0)
# Inicjalizacja wektora rozwiązania jako float64
x = np.zeros_like(b, dtype=np.float64)
# Dzielimy pracę między procesy
local_rows = n // size
start_row = rank * local_rows
end_row = start_row + local_rows if rank != size - 1 else n
# Przydzielenie lokalnych porcji A i b
local_matrix = matrix[start_row:end_row, :]
local_b = b[start_row:end_row]
# Lokalny wektor residuum
local_r = np.zeros_like(local_b, dtype=np.float64)
# Globalny wektor residuum (pełny rozmiar b)
global_r = np.zeros_like(b, dtype=np.float64)
start_time = time.time()
for _i in range(max_iter):
# Oblicz lokalny residuum r = b - A @ x
local_r[:] = local_b - np.dot(local_matrix, x)
# Tworzymy tymczasowy wektor o pełnym rozmiarze i kopiujemy lokalne dane
temp_r = np.zeros_like(b, dtype=np.float64)
temp_r[start_row:end_row] = local_r
# Sumujemy lokalne residuum przez wszystkie procesy
comm.Allreduce(temp_r, global_r, op=MPI.SUM)
# Aktualizujemy x równolegle na wszystkich procesach
x += omega * global_r
# Sprawdzamy warunek stopu (norma residuum)
if np.linalg.norm(global_r) < tol:
break
end_time = time.time()
execution_time = end_time - start_time
return x, execution_time
def check_solution(
matrix: Matrix, b: Vector, x_approx: Vector, tolerance: float = 0.008
) -> tuple[bool, float]:
"""Report whether an approximate solution is within tolerance."""
x_true = np.linalg.solve(matrix, b)
error = np.linalg.norm(x_true - x_approx)
return error < tolerance, error
if __name__ == "__main__":
logging.basicConfig(format="%(message)s", level=logging.INFO)
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
from matrix_generator import MatrixGenerator
sizes = [2, 5, 10, 50, 80, 100, 300, 500, 750, 1000, 5000, 10000]
for i in sizes:
if rank == 0:
matrix, b, lambda_min, lambda_max = (
MatrixGenerator.generate_matrix_and_vector("spd", size=i)
)
else:
matrix = None
b = None
lambda_min = None
lambda_max = None
matrix = comm.bcast(matrix, root=0)
b = comm.bcast(b, root=0)
# Rozwiązanie przy użyciu zrównoleglonej metody Richardsona
x, time_taken = richardson_parallel(Problem(matrix, b, lambda_min, lambda_max))
# Sprawdzanie poprawności rozwiązania (na procesie 0)
if rank == 0:
logger.info("Spd matrix with size %s", i)
is_correct, error = check_solution(matrix, b, x)
logger.info("Czas wykonania [s]: %s", time_taken)
logger.info(
"Czy rozwiązanie jest poprawne: %s", "Tak" if is_correct else "Nie"
)
logger.info("Błąd rozwiązania: %s", error)
if rank == 0:
matrix, b, lambda_min, lambda_max = MatrixGenerator.generate_matrix_and_vector(
"nemeth12"
)
else:
matrix = None
b = None
lambda_min = None
lambda_max = None
matrix = comm.bcast(matrix, root=0)
b = comm.bcast(b, root=0)
# Rozwiązanie przy użyciu zrównoleglonej metody Richardsona
x, time_taken = richardson_parallel(Problem(matrix, b, lambda_min, lambda_max))
# Sprawdzanie poprawności rozwiązania (na procesie 0)
if rank == 0:
logger.info("Nemeth12 matrix")
is_correct, error = check_solution(matrix, b, x)
logger.info("Czas wykonania [s]: %s", time_taken)
logger.info("Czy rozwiązanie jest poprawne: %s", "Tak" if is_correct else "Nie")
logger.info("Błąd rozwiązania: %s", error)
if rank == 0:
matrix, b, lambda_min, lambda_max = MatrixGenerator.generate_matrix_and_vector(
"poli3"
)
else:
matrix = None
b = None
lambda_min = None
lambda_max = None
matrix = comm.bcast(matrix, root=0)
b = comm.bcast(b, root=0)
# Rozwiązanie przy użyciu zrównoleglonej metody Richardsona
x, time_taken = richardson_parallel(Problem(matrix, b, lambda_min, lambda_max))
# Sprawdzanie poprawności rozwiązania (na procesie 0)
if rank == 0:
logger.info("Poli3 matrix")
is_correct, error = check_solution(matrix, b, x)
logger.info("Czas wykonania [s]: %s", time_taken)
logger.info("Czy rozwiązanie jest poprawne: %s", "Tak" if is_correct else "Nie")
logger.info("Błąd rozwiązania: %s", error)