OVERVIEW - Cycle 29 05/25

• Cycle:
  • Betting table: 13.05.2025
  • Review meeting: 03.06.2025 (slides)
• Participants:
  • CSCS: Edoardo, Enrique, Will
  • EXCLAIM: Anurag, Magdalena, Mauro
  • MCH: Christoph

Available people:

• GridTools:
  • Christos: 50% or a bit less (rest on PASC projects -push a bit more at the beginning-)
  • Edoardo: 100%
  • Enrique: 100%
  • Hannes: 0%
  • Philip: 60% (because of LUMI)
  • Rico: 80-90 % (Anticipate sporadic SwissTwins work)
  • Sara (ECMWF): 0% (working on distributed PMAP-GO + PASC poster)
  • Till: 66% (12.05-23.05 only, rest vacation)
  • Ioannis: 50% (because of cornerstone)
• CSCS:
  • Mikael (Simberg): 50%
  • Giacomo: 50% (rest on Flang investigation and Tabascal)
  • Will: 50% (rest: performance of GT4Py muphys, EXCLAIM Symposium)
  • Prashanth: 40% (mainly for I/O bits of ICON4Py)
• EXCLAIM:
  • Chia Rui:
  • Jacopo: 5-10% (helping out with Hannes' open PRs)
  • Magdalena: Full Cycle
  • Nikki: Full Cycle
  • Yilu: almost full cycle (will be away for 4 days)
• MCH:
  • Christoph: 100%
  • Daniel (50%):

Cycle Goal

Meaningful performance baseline for the Dycore granule

Same as Cycle 28, but also improving performance beyond what we see right now. ICON-EXCLAIM running with ICON4Py and GT4Py-DaCe main branches, with minimal Python overhead and meaningful performance benchmarks. This also requires finishing the combined programs and merging all the half-finished features like concat_where, and complete the benchmarking infrastructure to support writing more optimizations in next cycles.

Tasks

Project	Appetite	Developers	Support
[DaCe] Optimization VIII #	full cycle	Philip, Ioannis	Edoardo
[DaCe] Toolchain Runtime Support #	full cycle	Edoardo, Giacomo	Enrique
[GT4Py] Concat where cleanup & merge #	full cycle	Till	Enrique
[ICON4Py] Benchmarking infrastructure #	full cycle	Enrique, Christos, (Magdalena?)	Chia Rui
[ICON4Py] Combined solve non_hydro stencils #	1 week	Christoph (until done)	Nikki
[ICON4Py] Hannes' fixes 1: Static variants of programs #	1 week	Nikki	Edoardo
[ICON4Py] Hannes' fixes 2: Fix skip values #	1 week	Magdalena
[ICON4Py] Hannes' fixes 3: Fix validation errors #	1 week	Jacopo, Daniel	Christoph
[ICON4Py] Small CI improvements #	1 week	Magdalena	Enrique
[ICON4Py] Standalone greenline #	full cycle	Mikael, Yilu	Magdalena
[ICON4Py] Improving performance for CFL condition in dycore #	2 weeks	Christoph, Chia Rui	Magdalena
[Blueline] ICON with Granule deployment #	full cycle	Rico, Will	Christoph
Data Compression Project #	1 week	Nikki, Christos

---

---

Planning Action points

• @egparedes sync with @mluz about the icon4py tasks
  • combined stencils
  • RBF & standalone grid status
  • CI improvements
• @egparedes ask for possible dates for the group brainstorming meeting (if needed) and betting table and send the invites
• @egparedes send betting table invite
• sync and discuss with all people involved on different spack+venv projects (@ricoh @muellch @Will Sawyer, …)

Brainstorming

• benchmarking infrastructure:
  1. StencilTests with larger grids (either standalone or with serialized data)
     • validate with small grid
     • benchmark with large grid
     • accept static args
  2. green-line JW
     • intermediate step: serialize APE with R02B07
       • switch to standalone grid when ready
     • MCH experiment: ICON-CH2-small for validation, Christoph s ICON-CH2 for benchmark
     • add to bencher benchmark diffusion, dycore
  3. blue-line + granule (ape large grid?)
     • add to bencher
• profiling & identifying the low performance in the blueline + granule
• gt4py
  • merge performance counters as soon as they are merged in gt4py
  • finish the concat_where
• dace
  • reduce overhead of the program calls (dace)
  • toolchain instrumentation
  • continue with DaCe optimizations
  • parallel runs with DaCe
• icon4py
  • take over and merge PRs from Hannes handover document
  • finish the combined stencils (one week?)
  • consolidate the serialization points in icon-EXCLAIM (required for the serialization of the larger grids needed for benchmarking) (few days?)
  • standalone: (1 cycle, both tasks could go in parallel )
    • finish RBF validation
    • glue code to start using them and merge and integrate z_ifc branch.
  • Simple CI improvements: integrate nox utils to skip uneeded test execution (few days?)
  • fix CFL reduction in icon4py dycore
  • PASC Poster
• ICON uenvs –> needs larger discussion ??
  • missing experiments:
    • is it possible to run spack-installed uv in the install phase of a dependent spack package in the context of a uenv build? (locally -> uv does not run because spack does not expose libbz2 to it during staging of the dependent package, why?). Possible fix: depend on bzip2 as a build dependency in the icon4py spack package?
      • IF YES -> only needed to add a venv with icon4py for runtime in uenv post-install
      • IF NO -> consider alternatives: pip install instead of uv sync, install icon4py and then ICON both in post-install
  • uenv deployment ?
  • ICON CI (Will, Daniel) ?