-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Present: Ian, Eloisa, Seth, Josh, Yosef, Erik, Frank, Peter, Roland, Steve, Matt Kinsey
ET release: * most tests pass on all machines, exceptions are GRHydro's weno and slow_sector tests. Roland will look into them (since he put them into the repo). Suspicions are that they were generated using gcc but are now tested using intel (thing --fast-math) * Peter reports some test failures with new intel compilers as well (on his laptop)
* currently have 34 open tickets flagged for the release. Please have a look. 6 of them are in review, some are even reviewed ok.
* Ian now automatically builds documentation for ET, will not use it on ET page yet since it is build from trunk and the page should be the release code. Ian will also build the stable documentation and we can link to that one.
* Frank to ask Bruno to give the documentation a quick check
Yours, Roland
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Hello all,
ET release: * most tests pass on all machines, exceptions are GRHydro's weno and slow_sector tests. Roland will look into them (since he put them into the repo). Suspicions are that they were generated using gcc but are now tested using intel (thing --fast-math)
I looked into this and so far it is not clear to me where the differences come from. The points that differ are in the outer boundary in the "---" corner of the domain. (in 2D) They form a square of size of the ghost zones. This is idependent of using -O2 or -O0. Also compiling the WENO routines with gfortran (and everything else with ifort) makes no difference.
I also looked into the issue that many tests fail on bethe when oversubscribing the machine using 2 MPI processes and 8 threads each (there are only 8 physical cores). This turns out to be related to vectorization/OpenMP in the Kranc generated thorns, in particular McLachlan. The errors go away if I either run without openmp, or fewer openmp threads, eg. 1x16 fails but 1x8 is fine. They also go away if I disable vectorization. Changing various Vectors options for ALIGNED storage and read/write seems to make no difference.
This actually seems serious if really there is a bug in the interaction between vectorization and OpenMP (note that since the error goes away and comes back if I disable OpenMP only in eg. the ML_BSSN_enforce.cc file it does not seem to be related to reductions being not bit identical when using OpenMP.).
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://keys.gnupg.net.
users@lists.einsteintoolkit.org
-
Roland Haas -
Roland Haas