parallel make issue in ncdump tests

[edwardhartnett]

As discussed in #2392 there is a parallel make issue (i.e. make -j issue) in the ncdump testing. This causes the following failures:

99% tests passed, 2 tests failed out of 202
Total Test time (real) =  33.53 sec
The following tests FAILED:
	 40 - ncdump_tst_output (Failed)
	 41 - ncdump_tst_nccopy3 (Failed)

The problem is probably test file tst_output_c0.nc and some other files which are created by tst_output.sh and used by tst_nccopy3.sh.

The answer is to make the running of tst_nccopy3.sh dependent on tst_output.sh having completed.

And yet in the Makefile.am I see:

# The tst_nccopy3.sh test uses output from a bunch of other
# tests. This records the dependency so parallel builds work.
tst_nccopy3.log: tst_calendars.log run_utf8_tests.log tst_output.log	\
                 tst_64bit.log run_tests.log tst_lengths.log

So this should be working.

I will take a look at this issue after sorting out some other stuff.

[edwardhartnett]

OK, the problem is in tst_nccopy4.sh, which contains this:

if test -f tst_group_data${ext} ; then ${execdir}/tst_group_data ; fi
if test -f tst_enum_data${ext} ; then ${execdir}/tst_enum_data ; fi
if test -f tst_comp${ext} ; then ${execdir}/tst_comp ; fi
if test -f tst_comp2${ext} ; then ${execdir}/tst_comp2 ; fi

This appears to be an attempt to ensure these tests have been run first. However, this is not the way to achieve this. If this runs at the same time as tst_nccopy3.sh is running, one of these files might be in the process of being (re-)created when it is needed.

Instead, the way to ensure that these files all exist is to add a dependency on tst_nccopy4.log in the Makefile,am (as is already done with nccopy3.sh).

I will submit a PR...

[DennisHeimbigner]

Good. This kind of hidden dependency might occur in other test programs in the same directory.

[edwardhartnett]

I have to work on this PR a bit longer so I have turned it into a draft PR. Once I get it fully sorted out I will mark this PR as ready for review...

[edwardhartnett]

Btw I am also testing with hdf5 1.13.1 and everything seems to be working...

[edwardhartnett]

OK, I'm having this problem again. This time in the CMake build.

[edwardhartnett]

OK, I've got a fix. I will put up a PR.