I was wondering if anyone has successfully compiled and ran ICON 2026.04 on levante so far?
If I use ./config/dkrz/levante.intel --enable-openmp I get errors/warnings in during the model initialization that exactly tell me this before crashing:
146: FINISH PE: 146 mo_nwp_ecrad_init::setup_ecrad: Using Intel compiler with OpenMP for ecRad in ICON has known issues
So I tried using ./config/dkrz/levante.intel without openmp but then later in the initialization it crashes with:
0: Use transient, annually resolved greenhouse gases secenario based on CMIP5
248: [l10030:2067002:0:2067002] Caught signal 11 (Segmentation fault: invalid permissions for mapped object at address 0x7ffff7717bb0)
Maybe the config file has to be updated somehow. But would be nice to know if its a problem in my local set up or with the compiler and the new icon version.
building ICON like u did only provides a binary, that u can use for a couple of simulation setups: atmosphere, ocean, coupled. you can check which components are enabled by default with ./config/dkrz/levante.intel --help
Since you report a runtime error, the building worked for you.
In order to get help for runtime problems it is essential to provide information about the runtime setup:
which test did you run?
which components do you want to use?
did you use a single node? or multiple nodes?
providing path on levante could also help if ppl (like me) have access to the system.
Yes the compiling process works but only without OpenMP. The simulation set up works perfectly fine with icon2025.10 that I build it the same way as explained in the icon getting started website. This is why thought that it could be something with the compiler that makes the issue here. Because I never had issues at that specific moment in runtime.
I use SPICE environment to run a atmosphere only icon-nwp LAM simulation for Januar 2020 and just for reducing pending time I run only on two nodes which also usually always works. As I said this simulation set-up works with the older 2025.10 version.
You find the simulation here:
/work/aa0238/a271113/spice-v2.2/chain/gcm2icon/PanArctic_z0p0003_ctrl_disspheat_snowonice
and
/work/aa0238/a271113/spice-v2.2/chain/work/PanArctic_z0p0003_ctrl_disspheat_snowonice
I just ran the spice test experiment sp001 with this compiled version without openMPI and it works, so I guess I have to check again what the new version does not like about my setup. But it might not be necessarily a compiler issue then. So I would maybe close this issue here unless you have already found something interesting
But thank you so much for your help!
this has zero information. When I read spice I think of Dune or India. Looking into the related scripting is not very helpful, because it seems rather complex (scripting). A link to a documentation about spice could be useful.
Before u mentioned atmosphere only icon-nwp LAM simulation. now u refer to an ICON build without openmpi. Both together hardly work - or did u manage to run a local area setup without MPI at all?
regarding the build of /home/a/a271113/icon_2026.04-clean with openmp: I had zero issues building your icon with /home/a/a271113/icon_2026.04-clean/config/dkrz/levante.intel --enable-openmp && make -j12
So the compiling itself was not a problem but the new icon version will crash when using --enable-openmp . So I thought the Error: “Caught signal 11 (Segmentation fault: invalid permissions for mapped object at address “ would come if you don’t use openmp.
But I changed the values of num_io_procs and nproma and it worked again. So this has to be adapted apparently with the new version.
