TimeSeries readable only on same number of procs

[David]

TimeSeries can only be read if the script is run on the same number of processors as were used for writing the TimeSeries.

Example: Run the MWE on 1 process and then on 2. The serial TimeSeries is not read correctly on 2 processes. If the number of processors is equal, it works.

from dolfin import *
import numpy as np

mesh = UnitSquareMesh(64, 64)
V = FunctionSpace(mesh, 'P', 1)
u = interpolate(Expression('sin(20*x[0])*sin(20*x[1])', degree=1), V)

ts = TimeSeries('ts_parallel')

rank = MPI.rank(mesh.mpi_comm())
size = MPI.size(mesh.mpi_comm())


# store if running in serial
if size == 1:
    ts.store(u.vector(), 0.)


# read TimeSeries
u2 = Function(V)
ts.retrieve(u2.vector(), 0., False)
print('{}: |u-u2| = {}'.format(rank, np.linalg.norm(u.vector().get_local() -
                                                    u2.vector().get_local())))

I don't know if this is the intended behavior, at least I did not find any documentation about it.

Best wishes David

[David]

Ok, I just noticed this comment in HDF5File.cpp, in the write() function for dolfin functions. So one has to store functions with HDF5File to be able to read/write data on different numbers of procs.

  // FIXME: This routine is long and involves a lot of MPI, but it
  // should work for the general case of reading a function that was
  // written from a different number of processes.  Memory efficiency
  // could be improved by limiting the scope of some of the temporary
  // variables

It would be useful if this was mentioned somewhere in the docs. For instance, there are no docstrings for HDF5File.write and TimeSeries.write in python.