[petsc-users] petsc4py sparse matrix construction time
Cetinbas, Cankur Firat
ccetinbas at anl.gov
Mon Nov 6 10:38:45 CST 2017
It took me a while to respond because I couldn’t get converged results with the iterative solvers (they didn’t give any errors but the results didn’t match with Matlab mldivide and Python spsolve). So we spent some time trying to reconfigure petsc and petsc4py with MUMPS.
I don’t know if locally called mat.createAIJ or setValues come with any overhead but I timed each section of the code locally.
Single core
2 cores
4 cores
8 cores
16 cores
bcast
0.064
0.064
0.071
0.069
local vector init
0.004
0.0028
0.0015
0.00087
scatterv
0.0076
0.0056
0.0079
0.0099
convert to int 32
0.0024
0.0015
0.00094
0.00072
calculate indptr
0.0028
0.0018
0.0011
0.0007
create AIJ
0.044
0.0037
0.022
0.017
0.013
RHS vector
0.0016
0.0017
0.0019
0.0023
solve (MUMPS)
10.94
7.23
5.67
5.78
5.13
total time (s)
10.984
7.3161
5.7694
5.88134
5.22649
MUMPS solver is the most time consuming part so I need to find a way to make iterative solvers converge. The single core MUMPS solver takes almost 9 times longer with respect to mldivide in Matlab. I might be doing something wrong setting up the solver but the results of MUMPS match with Matlab.
ksp = PETSc.KSP()
ksp.create()
ksp.setOperators(pA)
ksp.setType('preonly')
pc = ksp.getPC()
pc.setType('lu')
pc.setFactorSolverPackage('mumps')
ksp.solve(rhv,x)
In addition I don’t understand why broadcast takes so much time w.r.t. scatter. I only pass a small array at the size of number of processes.
Regards,
Firat
From: Matthew Knepley [mailto:knepley at gmail.com]
Sent: Wednesday, November 01, 2017 3:35 AM
To: Cetinbas, Cankur Firat
Cc: Smith, Barry F.; petsc-users at mcs.anl.gov; Ahluwalia, Rajesh K.
Subject: Re: [petsc-users] petsc4py sparse matrix construction time
On Tue, Oct 31, 2017 at 10:00 PM, Cetinbas, Cankur Firat <ccetinbas at anl.gov<mailto:ccetinbas at anl.gov>> wrote:
Hi,
Thanks a lot. Based on both of your suggestions I modified the code using Mat.createAIJ() and csr option. The computation time decreased significantly after using this method. Still if there is a better option please let me know after seeing the modified code below.
At first trial with 1000x1000 matrix with 96019 non-zeros in the matrix, the computation time did not scale with the number of cores : Single core python @ 0.0035s, single core petsc @ 0.0024s, 2 cores petsc @ 0.0036s, 4 cores petsc @ 0.0032, 8 cores petsc @ 0.0030s.
Then I tried with larger matrix 181797x181797 with more non-zeros and I got the following results: Single core python @ 0.021, single core petsc @ 0.031, 2 cores petsc @ 0.024s, 4 cores petsc @ 0.014, 8 cores petsc @ 0.009s, 16 cores petsc @ 0.0087s.
I think the optimum number of nodes is highly dependent on matrix size and the number of non-zeros. In the real code matrix size (and so the number of non-zero elements) will grow at every iteration starting with very small matrices growing to very big ones. Is it possible to set the number process from the code dynamically?
I am not sure how you are interpreting these measurements. Normally, I would say
1) Everything timed below is "parallel overhead". This is not intended to scale with P, instead it will look like a constant, as you observe
2) The time to compute the matrix entires should far outstrip the time below to figure the nonzero structure. Is this true?
3) Solve time is often larger than matrix calculation. Is it?
Thus, we deciding on parallelism, you need to look at the largest costs, and how they scale with P.
Thanks,
Matt
Another question is about the data types; mpi4py only let me transfer float type data, and petsc4py only lets me use int32 type indices. Besides keep converting the data, is there any solution for this?
The modified code for matrix creation part:
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
if rank==0:
row = np.loadtxt('row1000.out').astype(dtype='int32')
col = np.loadtxt('col1000.out').astype(dtype='int32')
val = np.loadtxt('val1000.out').astype(dtype='int32')
m = 1000 # 1000 x 1000 matrix
if size>1:
rbc = np.bincount(row)*1.0
ieq = int(np.floor(m/size))
a = [ieq]*size
ix = int(np.mod(m,size))
if ix>0:
for i in range(0,ix):
a[i]= a[i]+1
a = np.array([0]+a).cumsum()
b = np.zeros(a.shape[0]-1)
for i in range(0,a.shape[0]-1):
b[i]=rbc[a[i]:a[i+1]].sum() # b is the send counts for Scatterv
row = row.astype(dtype=float)
col = col.astype(dtype=float)
val = val.astype(dtype=float)
else:
row=None
col=None
val=None
indpt=None
b=None
m=None
if size>1:
ml = comm.bcast(m,root=0)
bl = comm.bcast(b,root=0)
row_lcl = np.zeros(bl[rank])
col_lcl = row_lcl.copy()
val_lcl = row_lcl.copy()
comm.Scatterv([row,b],row_lcl)
comm.Scatterv([col,b],col_lcl)
comm.Scatterv([val,b],val_lcl)
comm.Barrier()
row_lcl = row_lcl.astype(dtype='int32')
col_lcl = col_lcl.astype(dtype='int32')
val_lcl = val_lcl.astype(dtype='int32')
indptr = np.bincount(row_lcl)
indptr = indptr[indptr>0]
indptr = np.insert(indptr,0,0).cumsum()
indptr = indptr.astype(dtype='int32')
comm.Barrier()
pA = PETSc.Mat().createAIJ(size=(ml,ml),csr=(indptr, col_lcl, val_lcl)) # Matrix generation
else:
indptr = np.bincount(row)
indptr = np.insert(indptr,0,0).cumsum()
indptr = indptr.astype(dtype='int32')
st=time.time()
pA = PETSc.Mat().createAIJ(size=(m,m),csr=(indptr, col, val))
print('dt:',time.time()-st)
Regards,
Firat
-----Original Message-----
From: Smith, Barry F.
Sent: Tuesday, October 31, 2017 10:18 AM
To: Matthew Knepley
Cc: Cetinbas, Cankur Firat; petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>; Ahluwalia, Rajesh K.
Subject: Re: [petsc-users] petsc4py sparse matrix construction time
You also need to make sure that most matrix entries are generated on the process that they will belong on.
Barry
> On Oct 30, 2017, at 8:01 PM, Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>> wrote:
>
> On Mon, Oct 30, 2017 at 8:06 PM, Cetinbas, Cankur Firat <ccetinbas at anl.gov<mailto:ccetinbas at anl.gov>> wrote:
> Hello,
>
>
>
> I am a beginner both in PETSc and mpi4py. I have been working on parallelizing our water transport code (where we solve linear system of equations) and I started with the toy code below.
>
>
>
> The toy code reads right hand size (rh), row, column, value vectors to construct sparse coefficient matrix and then scatters them to construct parallel PETSc coefficient matrix and right hand side vector.
>
>
>
> The sparse matrix generation time is extremely high in comparison to sps.csr_matrix((val, (row, col)), shape=(n,n)) in python. For instance python generates 181197x181197 sparse matrix in 0.06 seconds and this code with 32 cores:1.19s, 16 cores:6.98s and 8 cores:29.5 s. I was wondering if I am making a mistake in generating sparse matrix? Is there a more efficient way?
>
>
> It looks like you do not preallocate the matrix. There is a chapter on this in the manual.
>
> Matt
>
> Thanks for your help in advance.
>
>
>
> Regards,
>
>
>
> Firat
>
>
>
> from petsc4py import PETSc
>
> from mpi4py import MPI
>
> import numpy as np
>
> import time
>
>
>
> comm = MPI.COMM_WORLD
>
> rank = comm.Get_rank()
>
> size = comm.Get_size()
>
>
>
> if rank==0:
>
> # proc 0 loads tomo image and does fast calculations to append row, col, val, rh lists
>
> # in the real code this vectors will be available on proc 0 no txt files are read
>
> row = np.loadtxt('row.out') # indices of non-zero rows
>
> col = np.loadtxt('col.out') # indices of non-zero columns
>
> val = np.loadtxt('vs.out') # values in the sparse matrix
>
> rh = np.loadtxt('RHS.out') # right hand side vector
>
> n = row.shape[0] #1045699
>
> m = rh.shape[0] #181197 square sparse matrix size
>
> else:
>
> n = None
>
> m = None
>
> row = None
>
> col = None
>
> val = None
>
> rh = None
>
> rh_ind = None
>
>
>
> m_lcl = comm.bcast(m,root=0)
>
> n_lcl = comm.bcast(n,root=0)
>
> neq = n_lcl//size
>
> meq = m_lcl//size
>
> nx = np.mod(n_lcl,size)
>
> mx = np.mod(m_lcl,size)
>
> row_lcl = np.zeros(neq)
>
> col_lcl = np.zeros(neq)
>
> val_lcl = np.zeros(neq)
>
> rh_lcl = np.zeros(meq)
>
> a = [neq]*size #send counts for Scatterv
>
> am = [meq]*size #send counts for Scatterv
>
>
>
> if nx>0:
>
> for i in range(0,nx):
>
> if rank==i:
>
> row_lcl = np.zeros(neq+1)
>
> col_lcl = np.zeros(neq+1)
>
> val_lcl = np.zeros(neq+1)
>
> a[i] = a[i]+1
>
> if mx>0:
>
> for ii in range(0,mx):
>
> if rank==ii:
>
> rh_lcl = np.zeros(meq+1)
>
> am[ii] = am[ii]+1
>
>
>
> comm.Scatterv([row,a],row_lcl)
>
> comm.Scatterv([col,a],col_lcl)
>
> comm.Scatterv([val,a],val_lcl)
>
> comm.Scatterv([rh,am],rh_lcl)
>
> comm.Barrier()
>
>
>
> A = PETSc.Mat()
>
> A.create()
>
> A.setSizes([m_lcl,m_lcl])
>
> A.setType('aij')
>
> A.setUp()
>
> lr = row_lcl.shape[0]
>
> for i in range(0,lr):
>
> A[row_lcl[i],col_lcl[i]] = val_lcl[i]
>
> A.assemblyBegin()
>
> A.assemblyEnd()
>
>
>
> if size>1: # to get the range for scattered vectors
>
> ami = [0]
>
> ami = np.array([0]+am).cumsum()
>
> for kk in range(0,size):
>
> if rank==kk:
>
> Is = ami[kk]
>
> Ie = ami[kk+1]
>
> else:
>
> Is=0; Ie=m_lcl
>
>
>
> b= PETSc.Vec()
>
> b.create()
>
> b.setSizes(m_lcl)
>
> b.setFromOptions()
>
> b.setUp()
>
> b.setValues(list(range(Is,Ie)),rh_lcl)
>
> b.assemblyBegin()
>
> b.assemblyEnd()
>
>
>
> # solution vector
>
> x = b.duplicate()
>
> x.assemblyBegin()
>
> x.assemblyEnd()
>
>
>
> # create linear solver
>
> ksp = PETSc.KSP()
>
> ksp.create()
>
> ksp.setOperators(A)
>
> ksp.setType('cg')
>
> #ksp.getPC().setType('icc') # only sequential
>
> ksp.getPC().setType('jacobi')
>
> print('solving with:', ksp.getType())
>
>
>
> #solve
>
> st=time.time()
>
> ksp.solve(b,x)
>
> et=time.time()
>
> print(et-st)
>
>
>
> if size>1:
>
> #gather
>
> if rank==0:
>
> xGthr = np.zeros(m)
>
> else:
>
> xGthr = None
>
> comm.Gatherv(x,[xGthr,am])
>
>
>
>
> --
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/
--
What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/<http://www.caam.rice.edu/~mk51/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20171106/6bdd5336/attachment-0001.html>
More information about the petsc-users
mailing list