3343/edge

Extreme-scale Discontinuous Galerkin Environment (EDGE)

edge's Issues

EDGE-V: Usability

As discussed in PR #2:

1. In anticipation of larger meshes, its better to have a typedef for all mesh-specific integral types, such that we can easily replace the ints by longs in future
2. Lets get rid of the using namespace std and using namespace moab.
3. Same goes for preprocessor macros. E.g., std::min() works fine for #define min(X, Y) (((X) < (Y)) ? (X) : (Y)).
4. As far as I know, MOAB's sparse type MB_TAG_SPARSE should be used when only a few entities hold the tag. Since we assign velocities to all elements, dense is probably better.

EDGE-V: Thread Safety

Unfortunately, MOAB is not thread safe:
EDGE-V fails in parallel with MOAB-errors.
Runs fine with OMP_NUM_THREADS=1.

Two options:

1. Thread-private MOAB instances
2. Single threaded access to MOAB

edge/tools/edge_v/src/edge_v.cpp

Line 58 in dd9735e

moab_mesh mMsh;

GOMP: Thread ids and numa-aware inits

EDGE's numa-aware inits through OMP's num_threads:

edge/src/impl/elastic/setup.inc

Line 99 in 466a64d

#pragma omp parallel for num_threads( std::max( edge::parallel::g_nThreads-1, 1 ) )

edge/src/impl/elastic/solvers/common.hpp

Line 355 in 466a64d

#pragma omp parallel for num_threads( std::max( parallel::g_nThreads-1, 1 ) )

edge/src/impl/elastic/solvers/AderDg.hpp

Line 321 in 466a64d

#pragma omp parallel for num_threads( std::max( parallel::g_nThreads-1, 1 ) )

reset the thread-id of the last thread to zero:

edge/src/parallel/global.cpp

Line 31 in 466a64d

int edge::parallel::g_thread = 0;

---

Bug appears in GCC toolchain (4.9.2-10 and 5.3.0).
icpc (18.0.0 20170811) and clang++ (4.0.1) with iomp work fine.

---

Reproducer attached:

alex@deb:~/debug$ g++ -fopenmp global.cpp numt.cpp 
alex@deb:~/debug$ OMP_NUM_THREADS=3 ./a.out running
setting g_thread
printing g_thread
3 0
3 1
3 2
working on some side-effects..............................
printing g_thread
3 1
3 0
3 0

-bash-4.2$ icpc -fopenmp global.cpp numt.cpp 
-bash-4.2$ OMP_NUM_THREADS=3 ./a.out 
running
setting g_thread
printing g_thread
3 2
3 0
3 1
working on some side-effects..............................
printing g_thread
3 1
3 0
3 2

debug.zip

Building with GCC 7 fails with -Werror=implicit-fallthrough

Building from Master with GCC 7 and later fails due to a combination of -Wextra and -Werror. Since GCC 7, -Wextra emits the warning -Wimplicit-fallthrough= when a switch-case block is missing a "break" at the end of each case.

The submodule pugixml seems to have this issue, GCC apparently takes issue with a switch block, sends out the warning, and the terminates due to -Werror.

This can be suppressed by compiling with the flag -Wimplicit-fallthrough=0.