workaround for BGL nc_test failures

[Robert Latham]

Hi
Not sure how many of you are using pnetcdf on BGL machines.  Those of
you who are have probably read README.bgl where we note that one of
our test cases (nc_test calling the ncmpi_get_vara_*_all and
ncmpi_put_vara_*_all routines) fails with an assertion.  

I found out about a workaround recently, so if any of you have either
run into this problem, or are waiting for this problem to go away,
here's the trick:

By setting the BGLMPIO_TUNEBLOCKING environment variable to 0, the
MPI-IO library will skip the troublesome optimization and fall back to
an unoptimized (but correct) approach.  I would not suggest setting
this environment variable as a matter of course:  for one example, the
FLASH-IO benchmark (which runs correctly w/o the workaround) sees
anywhere from .5% to 20% performance degradation with
BGLMPIO_TUNEBLOCKING set.

In short, try out your pnetcdf apps on BGL.  Only if they die with an
assertion or a seg fault should you then try setting
BGLMPIO_TUNEBLOCKING to 0. 

==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Labs, IL USA                B29D F333 664A 4280 315B