HSF Generator Workshop follow-up meeting, 11 December 2018

Post-Workshop

• Short presentation done (Graeme) at the WLCG Overview Board
• Presentation in last week’s ATLAS Weekly meeting (Simone A) and talk from Josh in this week’s ATLAS Software and Computing Workshop
• Common themes…
  • Following up on GPU versions of the code
  • Support for generator codes from experiments / software engineers
• Tangential: Frameworks paper being published to arXiv.
  • Would there be interest in writing the CWP generators paper as an outcome?
• Write up proceedings of workshop (technical)?
  • Sounds like we could weave in original generator CWP paper, but we are further on.
  • Sets a good baseline.
  • Peer review from workshop participants, to put on arXiv.
  • Physics motivation is a key driver, but hard to be answer in any definitive way. Then describe benchmarking.
  • Target outline for before the end of the year. Finish proceedings end of January.

Take Home Messages / Actions / Planning

• Generator performance comparisons work
  • Broaden this activity
  • Understand differences between ATLAS and CMS - e.g. two things to understand
    • Comparison of sherpa and MG for the same process
    • Are the two experiments producing very different types of samples or configurations? Is ATLAS producing more or more expensive events?
    • How much is “user driven”? Learn from databases.
      • Samples requested
      • Generators used
        • With parameters…?
        • Benchmarks
        • Is it possible to actually get apples-apples comparisons? (also consider phase-space slicing)
    • Document these setups in github so anyone can re-run them
• LHCb is different from ATLAS/CMS as generation focuses on decays
• Regression tests (for physics results AND timing performance) and best code practice
  • The HSF can help with this (a lot of technical experts)
  • Can the experiments help with their infrastructure for timing regression?
• Projection of CPU needs for Run4
  • The fractional budget of event generation will increase as experiments move to fast simulation and faster reconstruction
  • Model for use of LO, NLO and NNLO?
  • Tails will be rather important.
• Best ways to support technical work in generator field
  • Experiments
    • Joint PhD students (CDTs in UK)
    • Lab positions
  • Engineers (how to fund this - grant applications?)
  • Prospects for ‘permanent’ jobs (i.e. not just two year grants)
    • Talking to senior management about this
    • Real career prospects for people doing this
  • Iterate again between ATLAS/CMS and Generator Groups
  • BEST4HEP ITN somewhat related
• Generators critical part of what experiments do. Aware that “reward system” has limitations, need to improve on this.
  • Physics improvements are what gets noticed in the theory community and therefore gets funding and jobs. Example speed/memory improvement in Herwig (matrix elements and NLO integration) would be “many years of work for little recognition”. (see Peter’s presentation)
  • Referencing can be improved.
    • Citations to general papers rather than specific ones where large improvements were made.
    • Important to cite papers with few authors where individual contributions can be easily recognized (rather than annual papers with all updates that came during one year). Many physics processes generated by the experiments rely on specific improvements and massages on the generators for those specific channels and precisions.
    • Simon: I think the important point is a fair citation policy, which in my opinion is hard to achieve by a rule of thumb. It is really up to the event generator’s authors to convey to the experiments/users what should be cited, including all relevant publications – this is where we need to improve. I think the minutes should reflect this.
• GPU ports / code portability? Are experiments interested in using/validating, particularly Madgraph. ATLAS yes. We should try this (get MG team to minimally validate).
  • Supercomputers/HPCs are going more and more towards GPUs.
  • At the moment not so incentivised for the generator teams.
• A dedicated performance workshop in the future (Spring?)
  • Say 5 days of real hands-on work
  • Spring too early, but still useful - benchmarklng is a preparation.
• Negative weights are a problem (but only in NLO, there are none in LO). These come from cancellations of real and virtual contributions and are unavoidable. “Unweighted” events with negative weights means that events have weight +1 or -1 essentially.
  • Need better metrics to understand the impact of negative weights. Example: scalar metrics integrating the knowledge of the weight distribution and also the relative cost (in CPU budget) of detector simulation vs generation.
  • Mainly need this for V+jets and ttbar
• Can a common hook be added to generators so that we can set generator level cuts in a standard way? (Ask Zach?)
  • Cuts “easy”, but how to make that efficient? Bias to the events that will be accepted.
• Experiments sharing inputs
  • Start with integration grids? These are smaller and more flexible.
  • LHEFs slightly more complicated.
  • One additional selling point and possible solution to the statistical correlation issue would be for experiments can make each others systematic uncertainty samples (if such samples are anyway used on each side).
• HSF has Working Groups in important areas of work for HEP. This area could benefit from the same.
  • Some convenors willing to carry this forward and organise?
    • Andrea, Steve, Josh, Taylor, Stefan
  • Collaboration with LPCC has been very fruitful - should be joint activity
    • Get in touch with Michelangelo again
  • Over to the convenors to now set the pace!
• Timescales
  • Target a conference like CHEP 2019 (Adelaide). Abstracts in the Spring.