[Users] Meeting minutes for 2021-11-18

[Roland Haas]

Hello Wolfgang,

> Roland: Finished the review. Code is in Excellent state! RePriMand has
> been added to the thornlist.
Concerning your question about running arbitrary executables other than
the main Cactus code for the test. It occurred to me that we have the
capability to postprocess test data:

https://www.einsteintoolkit.org/usersguide/UsersGuide.html#verbatim-92

and just above using the:

POSTPROC tool.exe regex-of-files

notation.

Nothing really stops you from having a "dummy" Cactus test that
produces an output file and abusing the postproc option to run the
actual test executable (and completely ignoring the output file
content).

Horrible, but maybe a first try.

Note that there are issues with this. Namely the postproc utility will
run on on whichever node ends up running the testsuite driver Perl
script. This could be your laptop, the login node of a cluster or a
compute node in a cluster (or a launch node in something like a Cray XE
system).

There is no guarantee that a login node will be able to run code
compiled for the compute nodes (which your tool would be since Cactus
*only* knows how to compile for compute nodes, not for the login
nodes). This is an issue when cross-compiling. Note that compiling for
KNL compute nodes using say a Broadwell login node (or even SkyLake) is
cross compiling since the KNL have instructions that the login node
cannot execute (and the Intel compiler inserts startup code to
main() that checks for them and aborts if they are not found).

Similarly some clusters may just not let you run any compiled code
without the equivalent of mpirun (again Crays are good examples for
this). Since the testsuite code does not prefix the tools by mpirun,
they would fail. Similarly some clusters do not let you execute mpirun
on the login nodes.

There are also (to no ones surprise I a guess) issues with executing a
secondary tool from within cactus_sim. Namely many MPI implementations
do not let you. See eg:

https://www.open-mpi.org/faq/?category=tuning#fork-warning

and note that cactus_sim is an MPI process (it is started via mpirun,
it calls MPI_Init() in its main()). Other MPI implementations have
similar restrictions.

Yours,
Roland

-- 
My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from http://pgp.mit.edu .
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
Url : http://lists.einsteintoolkit.org/pipermail/users/attachments/20211118/6c1d59fe/attachment.bin