« Previous - Version 11/21 (diff) - Next » - Current version
Onkar Sahni, 07/25/2012 06:40 pm

Meeting notes

July 25, 2012 Meeting and Webinar

Missing website content

Note: send all missing content requests to

  • Overview text for whole site
  • Institutional profiles
    • Columbia university (Mark Adams)
      • POCs and Description of activites
    • UCB (Jim Demmel)
      • POCs and Description of activites
  • People content missing
    • photos, one sentence description and contact info (email, phone, addr and URL)
    • all LBNL folks
    • Mark Adams (have picture but nothing else)
    • Karen email
  • Research descriptions
    • Most folks sent images embedded in a word document. If you did that, please send images separately in an email at highest resolution possible.
    • If you used latex, please send the latex file in addition to the PDF
  • Software section
    • Short description, URL and download
    • Still missing...
      • boxlib, chombo, moab, phasta, cgm, ML, sundials, mesh-adapt, gmi
      • summary for ugrid technologies
      • complex geometry section? separately or put in with ugrid?
        • Tim: role into ugrid stuff; Phil: can go either way; Mark: putting them together makes more sense.
      • High order discretizations?
        • Phil: part of structured vs. unstructured
      • Particle methods?
        • Phil: who is responsible rep. for particles?
        • Tim: Interested because it fits in MOAB data model but not from applications perspective yet.
        • Mark S: C.S. Chen's fusion project involves particles and unstructured meshes
        • Phil: take first cut for structured grid perspective.
        • Lori: Ok to have two sections there; one for structured and unstructured
        • XXXX: Particles should stand separate from grids
      • Solvers is complete
      • Interoperability of FASTMath technology. Still missing...
        • Mesh to mesh coupling
        • Common Interface development
        • Particle to mesh coupling
      • Leadership class facility use
        • Won't worry about this for time being

SIAM CS&E meeting

  • Would like FASTMath mini-symposiums on several topics. What do folks think?
  • Deadline for submitting is mid-august
  • How many and who will actually be able to go?
  • If each area is one mini, need at least 4 speakers
  • Is it ok to get a speaker outside of FASTMath but a related topic.
    • Ok but not preferred
  • Another option is to have one mini on FASTMath with 4 topics covered
    • Could have 8 speakers during this mini
    • Solicit interest to see who would like to give a talk
    • Have experience doing same with ITAPS but turns out to be a project review session. Its better to have topic-oriented sessions and wind up getting interest from folks on those topics.
  • What about user's of FASTMath software/applications?
    • FASTMath in action

SciDAC PI Meeting

  • Attendance is significantly constrained
  • FASTMath has 20 slots
  • In some cases we have multiple choice for people to fill some slots
  • Won't be able to have full team at the PI meeting
  • Meeting is Sep. 10-12 in WASH DC at Rockville Hilton
  • Meeting is NOT approved yet so don't plan travel until get formal word of approval
  • Expect 7 posters

Mesh adaptive loops (Onkar)

  • Need to frequently adapt to maintain efficiencies
    • May do mesh adaption every ten solution cycles
  • Interested in doing the integration of two or more simulation components via memory (e.g. not via files)
    • Flow solver and adaptive meshing library for example. Could be a separate error estimator and correction component too.
  • Each component may pre-exist and have thousands of man-years of development preceding our work here
  • Slide 6 discussion...
    • One executable, the yellow box, and two libraries being driven by it.
    • Implies need to have the simulation components adjusted to act more like a library than a standalone application
      • Non-trivial cost but much lower than actual effort to develop the simulation
    • Once you do this kind of work for one solver, its easier for the next
  • Relationship to coupled simulation in general?
    • Generic transfer of data part is a piece of the coupling problem; transfering the mesh is an issue
    • Really thinking of adaptive loops only not multi-physics type of coupling
    • The mesh part of things is a bit trickier. You wind up having two copies of mesh. Adaptive meshing contains a rich mesh database while the solver can get by with a reduced sized variant.
      • Don't you really have like 4 meshes; two copies in each component? How do you handle the issue of relating degrees of freedom on either side of the adaptation.
        • Ans: Don't have 4 copies. We can free up the last version of the mesh before the adaptation
  • Inter-language interoperability issues
    • Returning allocated data from Fortran to C
    • Much success with Fortran with iso_c_binding (Fortran 2003 standard - almost all compilers support this)
      • Tim T: Better way to go than the way ITAPS originally went with Fortran interoperability
  • Do not necessarily get bit-for-bit comparable results between file based and memory based approaches. This is due to order of operations variations between the two approaches. But do get numerically comparable results which differ only in insignificant decimal positions.
  • Mark M suggested to collect peak memory usage from two integration approaches (file-based and in-memory) - that will be good to compare

FASTMath_CSE_Agenda-Actions.pdf - Lori's notes from CSE Meeting (88.2 KB) Mark Miller, 04/03/2013 05:55 pm