Go to the Indico Home Page Oldest event Older event Up to category Newer event Newest event Printable version
  • Compact style
  • Indico style
  • Indico style - inline minutes
  • Indico style - numbered
  • Indico style - numbered + minutes
  • Indico Weeks View
    • Switch to the management area of this event
    Login

    EP R&D Software Working Group Meeting

    → Europe/Zurich
    32/1-A24 (CERN)

    32/1-A24

    CERN

    40
    Show room on map
    Live Notes
    EP R&D Software Working Group Meeting
    Zoom Meeting ID
    62893115044
    Host
    Graeme A Stewart
    Alternative host
    Andre Sailer
    Useful links
    Join via phone
    Zoom URL
    Hide

    Software R&D Working Meeting Minutes

    2022-12-14

    Apologies: Dalila

    Room: Graeme, Andre, Erica, Paul, Javier, Vincenzo, Pere, Piyush, Jonas, Valentin

    Remote: Jakob, Juraj, Witek, Marco, Alberto, Anna, Giovanna, Vincenzo Innocente, Andi, Pascal

    Introduction (Graeme)

    Noted that for hardware, one big advantage is that we can cover more phase space (different I/O systems and accelerators). We also share access between tasks. However, we do need a global optimisation look at the hardware request - at the moment it's rather generic.

    Proposed timetable for the second round of Phase II proposal agreed.

    Efficient Analysis (Javier)

    How much does RNTuple know about the underlying DAOS? How much does it need to know? Could be sensitive to site decisions.
    - Current results had no specific tuning, in fact. For any resonable storage setup we should get good results.

    Writing is faster, because the system needs to accept the data, then it can buffer the write. For reads the data needs to be read off disk, then delivered.
    - Can data be cached in the OS? Yes. But there is a cache management issue that needs to be understood.

    Other storage system features (slide 6). Are the other storage systems working on similar features?
    - Proposal on the HDF5 community to make HDF5 multi-threaded, so there is R&D there. Not sure about the others.

    Analysis throughput between Parquet and RNTuple - is this due to smaller reads (higher compression) or more efficient decompression or transient unpacking/repacking costs?
    - It's to do with metadata processing. Our metadata is tiny, could be a Parquet issue? We did see previously an issue in Parquet, which is now fixed. We optimise to a tighter use case.

    AOB

    There is a Compute Accelerator Forum meeting on running tracking on GPUs with ACTS this afternoon: https://indico.cern.ch/e/1160623

    There are minutes attached to this event. Show them.