Log In | Users | Register
COSMO Web
Edit | Attach | New | Raw | Diff | Print | | Tools
You are here: ICON » ICONmeet20190710

ICON meeting March 28, 2019

Participants:

Zane Dedekind ZD
Remo Dietlicher RD
Saskia Drossaart van Dusseldorp SD
Bernhard Enz BE
Sylvaine Ferrachat - excused, but reported by Colombe SF
Philippe Marti PM
Katie Osterried KO
Colombe Siegenthaler CS
Colin Tully CT

Reports:

* Saskia Drossaart van Dusseldorp (SD):
  • SD is waiting for Sylvaine.
  • SD will work on hail prediction in ICON (NWP physics) over Switzerland.
    • RD indicates that a set-up for ICON similar to COSMO-1 (i.e. 1km resolution) running on Daint exists at Meteoswiss.

* Colin Tully (CT):

  • CT is waiting for Sylvaine for running ICON. CT will use ICON-HAM.

* Colombe Siegenthaler (CS):

  • CS finished writing some python scripts which clone the official ICON organization from MPI repository and push ONLY the tags to the C2SM Github organization. CS is testing the script to be sure it does what she wants. As soon as it is finished, CS will push ICON-2.5.0 in the C2SM-ICON organization.
  • CS was at the ICON-HAM Hackathon in Oxford in the 1st week of July 2019. The two goals of the Hackathon were
    • to couple HAM to the "Sapphire package" (higher resolution package used in MPI).
    • to begin the development of HAM-Lite (simplified version of HAM, using fewer aerosol modes and simplified aerosol physics).
  • Here is a brief summary of the outcome of the Hackathon:
    • For the coupling of HAM to higher resolution package used in MPI, it has been decided to couple HAM in the icon-aes master branch. However, Monika Esch advised to wait until the merge of all the work done at GPU Hackathon in Zurich. This big merge should happen next week.
    • The development of HAM-Lite was indeed initiated.
    • CS worked on preparing the HAMMOZ-specific input files used by ECHAM-HAM for ICON-HAMMOZ.

* Sylvaine Ferrachat (SF):

  • Concerning her work with ICON, SF has been working on two projects in parallel:
    • ICON-HAM-GPU (eniac project) : re-basing HAM on the icon-eniac branch. The code is technically coupled, but not fully functional code (i.e., functional on CPUs).
    • ICON-HAM in Sapphire package : HAM was coupled to ICON (echam physics) by Marc Salzmann some years ago. However this existing ICON-HAM is based on an old tag (1.3.0) of icon-aes. SF is re-basing HAM onto the most recent icon-aes master. So far the code compiles but still crashes at initialization phase.
    • Remark from SF: my above ICON-HAM task in some way contributes to my preliminary ICON-HAM-GPU task, because the icon-aes version I worked on shares some common features with icon-eniac, meaning that the bugs I solved to have HAM running on top of icon-aes-master are also relevant for getting HAM to run on top of icon-eniac.

* Zane Dedekind (ZD):

  • ZD will work with ICON-LAM over Switzerland in around 6 months.

* Remo Dietlicher (RD):

  • RD is porting ICON NWP on GPU in the project IMPACT.
  • RD defined his working set-up : 80'000 grid-points (~ 10 km) over Switzerland.
  • RD will put his set-up on the ICON wiki (https://wiki.c2sm.ethz.ch/ICON/WebHome).

* Katie Osterried (KO):

  • KO was at the GPU Hackathon in Zürich. KO and RD adapted the COSMO and ICON testsuites for ICON, allowing for example to test for a threshold-tolerance (instead of bit identity).
  • The latter testsuite is working with Jenkins. KO wants to incorporate the testsuite in Buildbot, but for the moment Buildbot is broken.
    • PM added : the stability tests in Buildbot were disabled recently to make the tests successful.
  • Concerning Extpar : The Extpar technical testsuite used for development now includes the DWD ICON operational version as well as a standard ICON setup from MPI. There were a lot of code developments in the last few months because of the high resolution external parameters files needed for the QUBICC experiment. One outstanding problem with Extpar is that the resolution of the source files used by Extpar is not anymore higher that the output resolution needed, so some of the interpolation algorithms used are no longer sufficient.

* Philippe Marti (PM):

  • PM is attending the C2SM ICON meeting for the last time since PM will leave at the end of August 2019.
  • All the climate physics are ported on GPU, except for the radiation scheme (which will be the RRTMGP scheme) and JSBACH. PSRad is abandoned as it cannot be ported to GPU.
  • Next week all the developments done at GPU Hackathon (and afterwards) with GPU will be merged together. This does not include radiation and JSBACH. We should have at the end of next week (so on 20th of July) a ICON version running on GPU.
  • A GPU version of the radiation scheme RRTMGP already exists, but does not give the right results. It is very probably an interface problem.
  • Valentin is nearly done with JSBACH. After Valentin leaves, the maintenance of CLAW will be done by Xavier. CLAW is needed by COSMO for operational model runs.
  • The performance of GPU vs CPU :
    • radiation is not ported (cf above)
    • dynamical core, graupel scheme and vdiff (turbulence) : 3 times faster.
    • other parts (cloud, convection, aerosols, etc) are slower or as fast on GPU as on CPU (but not faster). Some of the slower parts are coming from a problem with atomic operations. Once PM has resolved them and other obvious problems, no part should not be slower on GPU anymore, but might not be faster. PM does not know if it matters yet, because he measured the performance piecewise, and does not know if the time spent in the slow pieces is proportionally important compared to the total computational time.
  • There will be a new build system for ICON in the coming release or the next one. It will work with Spack. To include JSBACH on GPU, the build system needs to include CLAW-specific pre-processing (but CLAW will have to be installed on the machine separately).
  • Compilers for the GPU version:
    • Cray was abandoned (Cray is not supporting OpenAcc anymore)
    • gcc : probably version 10 would work. The OpenAcc standard 2.6 need to be supported.
    • PGI : icon-eniac needs version 18.10 as minimum. On Daint, the default pgi 18.5.0 which will not work. On Daint, the cray environment for pgi is not right. PM has an separate hand-written module file (not part of the icon-eniac distribution) which unloads and loads the right modules (cray does it wrong). You are welcome to open tickets at cscs requesting a correct environment for pgi.

* Bernhard Enz (BE):

  • BE has a running version of ICON-LAM on Daint.
  • BE encountered a strange bug indicating that the restart might be broken: You have to stop the experiment at the same time as the restart files are written. If you do not, you get empty output files.
    • advice by PM: try a newer version (-> as soon as CS updates the C2SM-ICON organization with ICON-2.5, BE will give a try). If the problem persists, a bug report has to be opened.
    • advise II : try longer runs.

Next meeting:

October-November 2019, before the next ICON developer/infrastructure meeting (12.-14. November 2019 at Evangelische Tagungsstätte, Hofgeismar (DE)) 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!