2023/03/09 1WP3 meeting

From cheese2

ChEESE 1st WP3 meeting (March 9th 2023, 14.00h)

  • Agenda: Recall WP3 outlines and define first activities
  • Follow the meeting online:link
    • Access code: rAexyn

Presentation slides

WP3 Meeting presentation folder

Minutes

Attendees

• Filippo Mantovani (BSC) • Giorgio Amati (CIN) • Laetitia Le Pourhiet (SU) • Maria Jose Garcia (CSIC) • Marc Martinez (CSIC) • Mario Wille (TUM) • Marta Garcia (BSC) • Nathanael Schaeffer (CNRS) • Okba Hamitou (BULL) • Piero Lanucara (CIN) • Sebastian Wolf (TUM) • Stephan Jauré (BULL) • Vadim Monteiller (CNRS)

   • Erwan Raffin (BULL)

Applications

  • SeisSol : Sebastian Wolf (TUM)
  • ExaHyPE : Mario Wille (TUM)
  • Specfem3d : Vadim Monteiller (CNRS)
  • Tandem : Alice Gabriel (LMU) - TBC
  • HySEA: Jorge Macias (UMA) - TBC
  • PTatin3d : Leatitia Le Pourhiet (SU)
  • Xshells: Nathanael Schaeffer (CNRS)

Main points (see slides)

  • Work package outlines
  • Short terms activity
  • Link with WP2, especially T2.1 (code audit): as POP not funded yet we should work with the application owner. An initial survey has been set up on application which include mini-apps existence for example. Next common actions to be discussed.
  • Discussion around mini-apps identification and extraction, GPU, HBM and vectorization

Actions

  • ChEESE-1P mini-app (mini-SeisSol, mini-SPECFEM3D (to update), mini-FALL3D): we can start gathering the needs regarding the software stack, including compiler toolchain, libraries, and its impact on the prototypes.
  • ChEESE-2P mini-app (mini-xSHELLS, mini-HySEA, mini-Tandem and mini-pTatin3): Select a version of the code, associated relevant test cases (Small, Medium, Big), results validation

Q&A

  • How to get a Roofline or arithmetic intensity?

→ With a tool like likwid for roofline portable solution and Intel advisor/vtune or by analyzing the algorithm

  • question about mini-apps from Nathanael about CPU/GPU :

→ If GPU code is generic we can work with it, however if it is hardware specific((cuda/hip), better to work with cpu version (we can still play with gpu version)

  • Question about portions of code which are memory bound, portions of code which are cpu bound

→ Analysis should be done on main hotspots hardware choice should consider application as a hole (tradeoffs)

  • question about mini--apps from Laetitia about number of mini--apps

→ Focus on main hotspot, if a code may not have clear hotspots but may be a flat profile, in which case mini--apps could be developed by stripping code from IO and most dependencies