Log In | Users | Register
COSMO Web
Edit | Attach | New | Raw | Diff | Print | | Tools
You are here: ICON » Meetings » ICONmeet20200211

ICON meeting February 11, 2020

Participants:

Colombe Siegenthaler (CS); Doris Folini (DF); Sylvaine Ferrachat (SF); Sven Kotlarski (SK); Colin Tully (CT); Katie Osterried (KO); Michael Steiner (MS); Bernhard Enz (BE); Remo Dietlicher (RD); Will Sawyer (WS) .

Reports:

* Colombe Siegenthaler (CS):
  • CS has been supporting CT with running ICON-HAMMOZ.
  • CS ran some benchmarks with ICON on Piz Daint for a CSCS proposal:
    • Overall conclusions from the benchmarking: Intel is the fastest compiler by far, gcc is very slow.
    • WS suggests a few things that can be done to optimize performance of ICON on Piz Daint:
      • OpenMP bug is fixed in latest release of ICON, and using OpenMP can give a good speedup
      • Using hyperthreading also gives a performance boost
    • The benchmarks also showed some unexpected behavior- the run with 6 hourly output ran faster than the run with 1 monthly output. WS suggests the use of IO processors to optimize the IO operations of ICON.
    • Full details of the benchmark can be found on the wiki on this page.

* Sylvaine Ferrachat (SF):

  • SF is planning on porting the HAMMOZ portion of ICON-HAMMOZ to gpus, potentially using the CLAW compiler, but work has not yet begun in earnest.

* Sven Kotlarski (SK):

  • SK is leading a team at MeteoSwiss working on climate change scenarios. They are interested in the possibilities of using ICON in limited area mode for climate simulations, but have not yet begun working with the model.

* Colin Tully (CT):

  • CT is running ICON-HAMMOZ on Piz Daint.
  • CT would like to use restarts, but they are not working correctly in the (older) version of ICON that he is using. CS reports that restarts seem to work correctly in later versions of the model. However, CT might continue to use the older version of the model in order to compare results with a paper published by MPIM.
  • CT is working on scientific validation of ICON-HAMMOZ, which is not yet giving completely satisfactory results.

* Katie Osterried (KO):

  • KO is the source code administrator of the Extpar code, which is used to prepare external parameter fields for the ICON model. Extpar is currently in a state of very active development due to the presence of an intern working solely on it. Also, there are plans for ICON developers Luis Kornblueh from MPIM and Jürgen Helmert from DWD to come to ETHZ in April to spend a week working solely on Extpar development.
  • KO will work with RD this year to port some of the COSMO physical parameterizations onto gpu in the ICON model using OpenACC. The parameterizations themselves are already ported, but the interfaces of the code need to be ported and adapted. Work will probably begin with the TERRA soil model.

* Michael Steiner (MS):

  • MS is doing an internship with Dominik Brunner at EMPA and is interested in using ICON-ART.
  • MS has worked through the exercises from the ICON tutorial and, after some difficulty with incorrect grid files, was able to get ICON to run the exercises on Piz Daint. However, the ICON-ART code did not work in full chemistry mode.
  • MS will attend the ICON training course at the end of March in Germany.
  • The research group at EMPA is interested in using ICON-ART on gpus. This is not yet possible as the ART portion of the code as not yet been ported. WS reports that there have been preliminary discussions with people at KIT about how to go about porting ART to gpu, but there is no firm plan in place at the moment.

* Bernhard Enz (BE):

  • BE is running simulations with the ICON model in limited area mode, and is able to generate hurricanes in the model domain of the expected intensity.
  • The simulations are run with ICON version 2.5 on Piz Daint with a 13km grid resolution.

* Remo Dietlicher (RD):

  • RD works on the IMPACT project at MeteoSwiss, which has the goal of porting the COSMO (NWP) physics into the ICON model on gpu.
  • RD began by incorporating a tolerance checker into the testsuite that is running on Buildbot. This allows for correctness testing of the cpu version against the gpu version during the gpu porting.
  • RD has also incorporated a test into Buildbot that tests the NWP physics in limited area mode.
  • RD has finished porting the turbulence and microphysics paramaterizations to gpu in ICON, and this has been merged into the icon-nwp repository.
  • The goal is to have a working version of the NWP physics in ICON on gpu by the end of 2020. The biggest obstacle to this goal is the radiation parameterization, because the COSMO radiation scheme has been discarded and the ICON NWP radiation scheme is new and therefore will have to be ported completely from scratch.
  • RD will attend the ICON infrastructure meeting at ICCARUS along with Carlos Osuna.

* Will Sawyer (WS):

  • WS reports that the ICON model in climate mode (ECHAM physics) is now mostly ported to gpu. The radiation scheme has been ported, but is still undergoing scientific validation and is not yet merged into a released version of ICON.
  • The latest release of ICON (2.6) contains all the other gpu code as well as a new build system.
  • The current gpu porting of ICON was done mostly with OpenACC directives, and there are a few CUDA kernels used for data compression.
  • WS shows some benchmarks on Piz Daint using the gpu version of ICON: on a node to node comparison using the P100 gpu gives a 4.9x speedup over using one Haswell cpu.
  • WS explains that there are two kinds of communication used: the traditional MPI communication through the cpus, and one that is able to bypass some of the cpu route to allow the gpus to communicate more directly. This "gpudirect" communication method is shown to be 7% or so faster according to the benchmarks on Piz Daint.
  • The full presentation from WS can be found here.

Discussion topics:

* Nesting in ICON:
  • WS reports that the nesting feature of ICON has been mostly ported to gpu, and in theory should work on gpus. However, it still needs to be debugged and tested, and it is not clear who will do this work and when it will happen.

* C2SM releases of ICON

  • CS asks if it is possible to be able to squash merge the commits from the official ICON git repository on Gitlab into the C2SM-ICON repository on Github. Currently she has to overwrite the files and make a new commit for each release, which is a cumbersome way of using a git repository, but necessary due to the terms of the ICON license with ETH. WS will provide CS with an invitation to the ICON infrastructure meeting at ICCARUS in March, so that she may pose this question there to the ICON gatekeepers.

Next meeting:

Early July 2020 spacer
This site is managed by the Center for Climate Systems Modeling (C2SM).
This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors. Ideas, requests, problems? Send feedback!