use netcdf for reading Amber netcdf but netCDF4 for writing

- netcdf reads MUCH faster than netCDF4 but writes MUCH, MUCH
  slower than netCDF4: always read with netcdf but write with
  netCDF4 if available, otherwise use slow netcdf and warn
- implements @swails 's solution from ParmEd/ParmEd#722 -- thank you!!
- minimal testing: write the same trajectory with netcdf and with
  netCDF4
- NOTE: netCDF4 is not installed by default, use the re-introduced
  [AMBER] install target

     pip install MDAnalysis[AMBER]

  to request its installation but Amber users should also be aware
  of potential issues with the bundled netcdf library of Amber; see
  #506 (comment)
  for details
- closes #506

package/CHANGELOG

@@ -24,19 +24,19 @@ Deprecations
 Fixes
   * correctly read little-endian TRZ files on big-endian architectures (issue
     #1424)
-    
+
 Changes
   * remove deprecated TimeSeriesCollection
   * remove deprecated analysis.align.rms_fit_trj
   * remove deprecated analysis.contacts.ContactAnalysis
   * remove deprecated analysis.contacts.ContactAnalysis1
   * remove deprecated analysis.hbonds.hbond_analysis 1-indexing
   * remove deprecated analysis.rms `target` keyword from functions
-  * use scipy.io.netcdf pure python implementation for reading of Amber
-    netcdf3 trajctories instead of netCDF4 + netcdf lib (see also
-    Issue #506)
+  * use fast scipy.io.netcdf pure python implementation for reading of Amber
+    netcdf3 trajctories instead of netCDF4 but use netCDF4 for fast
+    writing (if available) or fall back to netcdf (see also Issue #506)
   * bundled scipy.io.netcdf (v0.19.1) as lib.netcdf as a fallback for
-    scipy.io.netcdf so that Amber NC trajectories can always be read
+    scipy.io.netcdf so that Amber NC trajectories can always be processed
 
 
 mm/dd/17 richardjgowers, rathann, jbarnoud, orbeckst, utkbansal

package/MDAnalysis/coordinates/TRJ.py

@@ -152,6 +152,11 @@
     logger.debug("Using the bundled lib.netcdf (from scipy 0.16.1) instead.")
     from ..lib import netcdf
 
+try:
+    import netCDF4
+except ImportError:
+    netCDF4 = None
+    logger.warn("netCDF4 is not available. Writing AMBER ncdf files will be slow")
 
 class Timestep(base.Timestep):
     """AMBER trajectory Timestep.
@@ -629,6 +634,9 @@ class NCDFWriter(base.WriterBase):
        Added ability to write velocities and forces
     .. versionchanged:: 0.11.0
        kwarg `delta` renamed to `dt`, for uniformity with other Readers
+    .. versionchanged:: 0.17.0
+       Use fast netCDF4 for writing but fall back to slow scipy.io.netcdf
+       if netCDF4 is not available.
     """
 
     format = 'NCDF'
@@ -722,7 +730,15 @@ def _init_netcdf(self, periodic=True):
                 errno.EIO,
                 "Attempt to write to closed file {0}".format(self.filename))
 
-        ncfile = netcdf.netcdf_file(self.filename, mode='w', version=2)
+        if netCDF4:
+            ncfile = netCDF4.Dataset(self.filename, 'w', format='NETCDF3_64BIT')
+        else:
+            ncfile = netcdf.netcdf_file(self.filename, mode='w', version=2,
+                                        mmap=False)
+            warn_msg = "Could not find netCDF4 module. Falling back to MUCH slower "\
+                       "scipy.io.netcdf implementation for writing."
+            logger.warn(warn_msg)
+            warnings.warn(warn_msg)
 
         # Set global attributes.
         setattr(ncfile, 'program', 'MDAnalysis.coordinates.TRJ.NCDFWriter')

package/setup.py

@@ -512,6 +512,7 @@ def dynamic_author_list():
           # you might prefer to use the version available through your
           # packaging system
           extras_require={
+              'AMBER': 'netCDF4',
               'analysis': [
                   'seaborn',  # for annotated heat map and nearest neighbor
                               # plotting in PSA

testsuite/MDAnalysisTests/coordinates/test_netcdf.py

@@ -32,6 +32,7 @@
     # fall back (should ALWAYS work)
     from MDAnalysis.lib import netcdf
 
+import pytest
 from nose.plugins.attrib import attr
 from numpy.testing import (assert_, assert_equal, assert_array_almost_equal,
                            assert_array_equal,
@@ -185,7 +186,7 @@ def tearDown(self):
         del self.Writer
         del self.tmpdir
 
-    def test_write_trajectory(self):
+    def _test_write_trajectory(self):
         t = self.universe.trajectory
         with self.Writer(self.outfile, t.n_atoms, dt=t.dt) as W:
             self._copy_traj(W)
@@ -208,6 +209,14 @@ def test_write_trajectory(self):
                 err_msg='ncdf time output not float32 '
                         'but {}'.format(time[:].dtype))
 
+    def test_write_trajectory_netCDF4(self):
+        pytest.importorskip("netCDF4")
+        return self._test_write_trajectory()
+
+    @block_import('netCDF4')
+    def test_write_trajectory_netcdf(self):
+        return self._test_write_trajectory()
+
     def test_OtherWriter(self):
         t = self.universe.trajectory
         with t.OtherWriter(self.outfile) as W: