▾ Graeme Stewart - HSF
▾ Points from the talk
• HSF created to face the computing & software challenges for HL-LHC in light of needs ramping up faster than technology
• Idea: bottom up effort, do-ocracy
▾ Foundation (no profit/funding only from host universities in terms of members time)
• Does not have funding / commercialize software, but helps members apply for funding (e.g. IRIS-HEP)
• Has grown from roadmap whitepaper to a more organized working group structure
▾ Points from Q&A
• Computing is also an interest of the HSF, we have joint yearly meetings with the WLCG and are in touch regularly
• Good synergies with nuclear physics (ALICE, FAIR), disussions on data management (DOMA) in WLCG
• All software from collaborations is open source, licensing is also important
▾ Encouraging a modular approach to software, so it is not all ready at once
• Projects that follow that idea and are rather mature: Rucio, DD4hep, ACTS [add links]
• Projects that could be useful: EM shower generation
▾ HSF does not mean to write code for the experiments, but aims to be a great help
• e.g. by making sure experiments can find and take off-the-shelf pieces that are useful for their software
• Work on approximate statistical methods in event generation may be happening in some groups, not aware of official efforts yet though
• Role of ML: significant contributions in analysis (including object identification) - simulation and reconstruction are harder
▾ Carlos Munoz Camacho - AGATA and EIC
▾ Points from the talk
• Experiments covered: AGATA, EIC
• Some unique challenges, in terms of reconstruction and physics space
▾ (Actual) real-time analysis, using streaming readout
• Difference wrt LHC real-time analysis: timescales, LHC is a few hours and the data is stored in a buffer for as long as it’s necessary for the calibration, here one wants to avoid writing raw data altogether
• A key challenge of real time analysis is fast calibration / self-calibration
▾ Rates: like LHCb, so not immense, and can think outside the box
• Exascale computing brings accelerators and that's a real challenge for software writers
• There is an EIC software group, HSF is in contact with them
▾ Points from Q&A
• Do you share code with lattice QCD calculations (main consumers of HPC)
▾ Giovanni Lamanna - ESCAPE
▾ Points from discussion
• Data lake being co-developed with HEP
• Idea of a virtual research environment, with connections to HSF
• Economy of scale: choose to use existing building blocks, rather than rederive everything
▾ Since there will be many more communities, propose a modular workflow (with containers) that different collaborations can adapt to
• Build those example workflows around science cases
• Researchers become software writers, not only data users
• Teams are more diverse, including computing scientists, physicists and data scientists (more professionals needed)
• European Open Science Cloud infrastructure can provide funding, including for people
▾ Investment in training is very important as students have diverse backgrounds (computer science, physics, data science)
• HSF could contribute to the ASTERICS school in LAPP by extending to HEP as well
• This would be a way to start collaborating straight away, HSF will email responsible from ESCAPE
▾ Chris Tunnell - Experience from Direct Detection community
▾ Points from discussion
• DD is a large but more heterogeneous community wrt HEP
▾ Nevertheless, bottom-up effort has started, community-building stage
• First identify the needs, then the solutions
• Idea of short-term, limited-scope inter-collaboration efforts - could extend beyond DD?
• DANCE workshop @ RICE: https://dance.rice.edu
▾ Paschal Coyle - KM3NeT/ANTARES
▾ Points from the talk
Data is quite managable
Simulation is very expensive, especially with large detectors (ML, GPU interesting?)
Accelerated photon transport - JUNO experiment simulations
▾ Event-like data, like HEP
• Machine learning is having a big impact
▾ Data still manageable, total of 995 TB * 3 building blocks
• Possibly the time to start thinking about archiving and opening data
• Starting to try DIRAC for MC generation / data analysis, thinking about using Grid
▾ Points from Q&A
▾ Challenge: large detector —> big simulation overhead
• Same as CTA, only a few days of data taking
• This is an issue for machine learning, as training data is limited
• Could look into accelerated photon transport from JUNO (plenary at CHEP)
▾ Plugged into a real-time alert system sensitive to supernovas
• Even in that case, event rate does not go up significantly
▾ Final discussion and outcomes
▾ Strengthen links between HSF and ESCAPE
• Opportunity for further funding through this cluster, also for recruitment
• Could raining (some urgency on this - school is in June)
▾ Software catalogs
• Many communities would benefit from a classification of useful/supported/documented software e.g. on peak finders, filtering, compression
▾ Trying a physics case: dark matter?
• The know-how exists (and it is actively building up in direct detection), we could use dark matter searches as a prototype
▾ What next
• Advertise workshops interesting
• Possibly put in an Expression Of Interest to APPEC-NuPECC-ECFA for support towards continuing this discussion