Dear developers,
I have been running into problems with CCSDT(Q) calculations
on closed-shell, neutral molecules, using the stand-alone MRCC
code (V. 2014-07-10; compile log appended; for comments on
the 2015-02-04 version see below) when requesting large amounts
of memory.
If, on one particular machine with 256 GB memory and 24 cores, I
run the sample input of acetylene (below) requesting 60000 MB of
memory, it just runs fine. If I increase the request up to 123000 MB,
it runs fine as well. However, if I go to 124000 MB or beyond, then
the calculation runs into a problem at the very beginning (when
generating an initial SCF guess). Below I include the relevant section
of the output (Sample output 1).
Note that the last 5 lines (2 empty, 3 with messages "Fatal error in...",
"echo ...", "Program will stop.") are repeated over and over again,
at a fast pace. The output file easily grows to hundreds of MB
within minutes!
Obviously this much memory is not needed for the small example demonstrated
here, but the very same problem occurs for all larger cases for which I need
a lot of memory. The problem always appears to occur at the same cutoff on
that particular cluster (up to ca 123 GB fine, problems beyond). I experience
the same problems on other machines, but at different cutoffs (usually smaller,
like 40 GB). Note that I always request less memory in the input than I request
in the batch queuing script and than the machine actually has. Obviously this
problem prevents me from doing certain large calculations as the memory requirements
scale (almost) linearly with the number of OMP-cores used and I wish to use
16-24 cores to keep the total wall-clock time within queuing system limits.
I have tested the same example with the new version of MRCC (2015-02-04, compiled
the same way on the same machine), and I experience similar problems, actually
a little worse:
If I request 65000 MB or less the job just runs fine, if I request 124000 MB, it
shows very similar behavior as described above but the output is a little
different (Sample output 2).
Apparently the error appears even earlier now and the message
"Fatal error in pwd > junkscript."
is replaced by
"Fatal error in which dmrcc > mrccjunk1".
If I request 85000 MB, 100000 MB, or 120000 MB, the program proceeds until
"Generating initial guess for the SCF calculation..." but produces no further
output. I have waited for at least 10 minutes before I killed the job.
If I request 123000 MB I get the final message:
Fatal error in cd mrccjunk/; ./junkscript > /dev/null.
Program will stop.
but no endless repetition of the 5 lines mentioned above.
Do you have any idea what the problem might be and how it may be
fixed?
Thank you,
best regards,
Dirk Bakowies
