[petsc-users] Python PETSc performance vs scipy ZVODE

Zhang, Hong hongzhang at anl.gov
Mon Aug 14 17:37:36 CDT 2023


PETSs is not necessarily faster than scipy for your problem when executed in serial. But you get benefits when running in parallel. The PETSc code you wrote uses float64 while your scipy code uses complex128, so the comparison may not be fair.

In addition, using the RHS Jacobian does not necessarily make your PETSc code slower. In your case, the bottleneck is the matrix operations. For best performance, you should avoid adding two sparse matrices (especially with different sparsity patterns) which is very costly. So one MatMult + one MultAdd is the best option. MatAXPY with the same nonzero pattern would be a bit slower but still faster than MatAXPY with subset nonzero pattern, which you used in the Jacobian function.

I echo Barry’s suggestion that debugging should be turned off before you do any performance study.

Hong (Mr.)

On Aug 10, 2023, at 4:40 AM, Niclas Götting <ngoetting at itp.uni-bremen.de> wrote:


Thank you both for the very quick answer!

So far, I compiled PETSc with debugging turned on, but I think it should still be faster than standard scipy in both cases. Actually, Stefano's answer has got me very far already; now I only define the RHS of the ODE and no Jacobian (I wonder, why the documentation suggests otherwise, though). I had the following four tries at implementing the RHS:

  1.  def rhsfunc1(ts, t, u, F):
    scale = 0.5 * (5 < t < 10)
    (l + scale * pump).mult(u, F)
  2.  def rhsfunc2(ts, t, u, F):
    l.mult(u, F)
    scale = 0.5 * (5 < t < 10)
    (scale * pump).multAdd(u, F, F)
  3.  def rhsfunc3(ts, t, u, F):
    l.mult(u, F)
    scale = 0.5 * (5 < t < 10)
    if scale != 0:
        pump.scale(scale)
        pump.multAdd(u, F, F)
        pump.scale(1/scale)
  4.  def rhsfunc4(ts, t, u, F):
    tmp_pump.zeroEntries() # tmp_pump is pump.duplicate()
    l.mult(u, F)
    scale = 0.5 * (5 < t < 10)
    tmp_pump.axpy(scale, pump, structure=PETSc.Mat.Structure.SAME_NONZERO_PATTERN)
    tmp_pump.multAdd(u, F, F)

They all yield the same results, but with 50it/s, 800it/, 2300it/s and 1900it/s, respectively, which is a huge performance boost (almost 7 times as fast as scipy, with PETSc debugging still turned on). As the scale function will most likely be a gaussian in the future, I think that option 3 will be become numerically unstable and I'll have to go with option 4, which is already faster than I expected. If you think it is possible to speed up the RHS calculation even more, I'd be happy to hear your suggestions; the -log_view is attached to this message.

One last point: If I didn't misunderstand the documentation at https://petsc.org/release/manual/ts/#special-cases, should this maybe be changed?

Best regards
Niclas

On 09.08.23 17:51, Stefano Zampini wrote:
TSRK is an explicit solver. Unless you are changing the ts type from command line,  the explicit  jacobian should not be needed. On top of Barry's suggestion, I would suggest you to write the explicit RHS instead of assembly a throw away matrix every time that function needs to be sampled.

On Wed, Aug 9, 2023, 17:09 Niclas Götting <ngoetting at itp.uni-bremen.de<mailto:ngoetting at itp.uni-bremen.de>> wrote:
Hi all,

I'm currently trying to convert a quantum simulation from scipy to
PETSc. The problem itself is extremely simple and of the form \dot{u}(t)
= (A_const + f(t)*B_const)*u(t), where f(t) in this simple test case is
a square function. The matrices A_const and B_const are extremely sparse
and therefore I thought, the problem will be well suited for PETSc.
Currently, I solve the ODE with the following procedure in scipy (I can
provide the necessary data files, if needed, but they are just some
trace-preserving, very sparse matrices):

import numpy as np
import scipy.sparse
import scipy.integrate

from tqdm import tqdm


l = np.load("../liouvillian.npy")
pump = np.load("../pump_operator.npy")
state = np.load("../initial_state.npy")

l = scipy.sparse.csr_array(l)
pump = scipy.sparse.csr_array(pump)

def f(t, y, *args):
     return (l + 0.5 * (5 < t < 10) * pump) @ y
     #return l @ y # Uncomment for f(t) = 0

dt = 0.1
NUM_STEPS = 200
res = np.empty((NUM_STEPS, 4096), dtype=np.complex128)
solver =
scipy.integrate.ode(f).set_integrator("zvode").set_initial_value(state)
times = []
for i in tqdm(range(NUM_STEPS)):
     res[i, :] = solver.integrate(solver.t + dt)
     times.append(solver.t)

Here, A_const = l, B_const = pump and f(t) = 5 < t < 10. tqdm reports
about 330it/s on my machine. When converting the code to PETSc, I came
to the following result (according to the chapter
https://petsc.org/main/manual/ts/#special-cases)

import sys
import petsc4py
petsc4py.init(args=sys.argv)
import numpy as np
import scipy.sparse

from tqdm import tqdm
from petsc4py import PETSc

comm = PETSc.COMM_WORLD


def mat_to_real(arr):
     return np.block([[arr.real, -arr.imag], [arr.imag,
arr.real]]).astype(np.float64)

def mat_to_petsc_aij(arr):
     arr_sc_sp = scipy.sparse.csr_array(arr)
     mat = PETSc.Mat().createAIJ(arr.shape[0], comm=comm)
     rstart, rend = mat.getOwnershipRange()
     print(rstart, rend)
     print(arr.shape[0])
     print(mat.sizes)
     I = arr_sc_sp.indptr[rstart : rend + 1] - arr_sc_sp.indptr[rstart]
     J = arr_sc_sp.indices[arr_sc_sp.indptr[rstart] :
arr_sc_sp.indptr[rend]]
     V = arr_sc_sp.data[arr_sc_sp.indptr[rstart] : arr_sc_sp.indptr[rend]]

     print(I.shape, J.shape, V.shape)
     mat.setValuesCSR(I, J, V)
     mat.assemble()
     return mat


l = np.load("../liouvillian.npy")
l = mat_to_real(l)
pump = np.load("../pump_operator.npy")
pump = mat_to_real(pump)
state = np.load("../initial_state.npy")
state = np.hstack([state.real, state.imag]).astype(np.float64)

l = mat_to_petsc_aij(l)
pump = mat_to_petsc_aij(pump)


jac = l.duplicate()
for i in range(8192):
     jac.setValue(i, i, 0)
jac.assemble()
jac += l

vec = l.createVecRight()
vec.setValues(np.arange(state.shape[0], dtype=np.int32), state)
vec.assemble()


dt = 0.1

ts = PETSc.TS().create(comm=comm)
ts.setFromOptions()
ts.setProblemType(ts.ProblemType.LINEAR)
ts.setEquationType(ts.EquationType.ODE_EXPLICIT)
ts.setType(ts.Type.RK)
ts.setRKType(ts.RKType.RK3BS)
ts.setTime(0)
print("KSP:", ts.getKSP().getType())
print("KSP PC:",ts.getKSP().getPC().getType())
print("SNES :", ts.getSNES().getType())

def jacobian(ts, t, u, Amat, Pmat):
     Amat.zeroEntries()
     Amat.aypx(1, l, structure=PETSc.Mat.Structure.SUBSET_NONZERO_PATTERN)
     Amat.axpy(0.5 * (5 < t < 10), pump,
structure=PETSc.Mat.Structure.SUBSET_NONZERO_PATTERN)

ts.setRHSFunction(PETSc.TS.computeRHSFunctionLinear)
#ts.setRHSJacobian(PETSc.TS.computeRHSJacobianConstant, l, l) #
Uncomment for f(t) = 0
ts.setRHSJacobian(jacobian, jac)

NUM_STEPS = 200
res = np.empty((NUM_STEPS, 8192), dtype=np.float64)
times = []
rstart, rend = vec.getOwnershipRange()
for i in tqdm(range(NUM_STEPS)):
     time = ts.getTime()
     ts.setMaxTime(time + dt)
     ts.solve(vec)
     res[i, rstart:rend] = vec.getArray()[:]
     times.append(time)

I decomposed the complex ODE into a larger real ODE, so that I can
easily switch maybe to GPU computation later on. Now, the solutions of
both scripts are very much identical, but PETSc runs about 3 times
slower at 120it/s on my machine. I don't use MPI for PETSc yet.

I strongly suppose that the problem lies within the jacobian definition,
as PETSc is about 3 times *faster* than scipy with f(t) = 0 and
therefore a constant jacobian.

Thank you in advance.

All the best,
Niclas


<log.log>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20230814/44f6afa2/attachment-0001.html>


More information about the petsc-users mailing list