[Python] Solving sparse linear system on GPU is much slower than CPU

Below is the code which solves a sparse linear system:

import cupyx
import cupyx.scipy.sparse.linalg
import time
import scipy
import scipy.sparse.linalg
import pathlib

file_dir = str(pathlib.Path(__file__).parent.resolve())

# ------------ cpu (scipy) -------------------------------
A_host = scipy.sparse.load_npz(file_dir+'/A.npz')
B_host = scipy.sparse.load_npz(file_dir+'/B.npz')

t00 = time.time()
P_host = scipy.sparse.linalg.spsolve(A_host, B_host)
t01 = time.time()

# ------------ gpu (cupy) -------------------------------
A_device = cupyx.scipy.sparse.csr_matrix(A_host)
B_device = cupyx.scipy.sparse.csr_matrix(B_host)

t10 = time.time()
# 2nd parameter of the below function must be cupy.ndarray
P_seg = cupyx.scipy.sparse.linalg.spsolve(A_device, B_device.toarray())
t11 = time.time()

# ------------ timing results -------------------------------
print('cpu:', t01-t00)
print('gpu:', t11-t10)


The size of matrix A is about 100000x100000, and size of matrix B is about 100000x8. The running results is 0.689s on CPU, 37.774s on GPU.

Why so much slower speed on GPU? A.npz and B.npz can be downloaded here.

Continue reading...