Using DDT with Swift on Cosma

Running on the compute nodes

DDT can submit batch jobs itself but if Cosma is busy this means you go back into the queue if you exit the debugger. An alternative method:

salloc --ntasks=2 --ntasks-per-node=2 -p cosma7 -A dp004 -t 8:00:00

This starts a new shell on the login node. Any MPI programs run in this shell will be started on the allocated compute nodes.

ddt ./swiftsim/build/examples/swift_mpi --verbose=1 --self-gravity --cosmology --velociraptor --pin --threads=14 ./eagle_12.yml

Need to make sure ddt does NOT submit a batch job. The slurm allocation will be cancelled if you exit the shell started by salloc.

DDT memory profiling with a default Swift build using the Intel compiler results in segfaults. It seems to work if you set

CFLAGS="-fno-inline -shared-intel"

when running configure.

Memory profiling Velociraptor is extremely slow unless the option to store stack traces is switched off.