HSF Generator Meeting #3, 14 March 2019
Agenda: https://indico.cern.ch/event/802643
Present/Contributors: Simone Amoroso, Andrea Valassi, Efe Yazgan, Lesya Shchutska, Qiang Li, Taylor Childers, Zach Marshall, Olivier Mattelaer, Josh McFayden
Discussion on sharing SUSY samples across experiments
- ATLAS-CMS comparison recently started for process definitions in SUSY searches
- Detailed comparison of MadGraph setups (both ATLAS and CMS use MadGraph for SUSY)
- Run cards: matching configuration is very different in the two experiments, cuts instead differ only in minor ways. Difference in systematics setup partly depends on different MadGraph versions.
- Param cards, i.e. SUSY models: there are some differences. For
sbottoms, the physics that is tested is essentially the same. But
for electroweakinos, there may be quite some differences.
- Olivier: is b mass different? Josh: there is an agreed set for the experiments, there may still be something to harmonise. But it may be different from the MadGraph default. Olivier: please send me the values you use, so we can put it in MadGraph.
- Decays: ATLAS rely on MadSpin (both on-shell and off-shell), CMS
do not.
- Lesya: in CMS we reweight one sample to test different physics, e.g. for details of CMS look at Figure 8 in SUS-16-051.
- Pythia showering/merging: each experiment uses its own recommended
settings, this would be difficult to harmonise
- Josh: could this make a difference? Zach: probably not a big difference through the acceptance, but will try and run both schemes and compare.
- Qiang: which tuning is ATLAS using? CMS is using its own tune. Zach: working on new tunes at the moment, now using an ATLAS specific tune (A14). Zach: can also try the CMS tune to see if it makes a difference, is it available? Efe: not available in the pythia software, but parameters are public, will send them to you. [Info circulated by Efe after the meeting: CMS default tune parameters are in github (in most cases, for NNLO PDF CMS uses CP5 and for LO PDF they use CP2) - and there is a corresponding public note for this in cds, paper version is coming soon].
- Zach’s opinion on sharing
- Harmonizing does not mandate sharing, but sharing prevents later divergence
- If sharing: need to choose a format (LHE?) and mechanism, with authorizations (rucio?)
- Opinion: focus on harmonization for now, but then it would be nice to have the technical issues solved for sharing, can the HSF WG help?
- Early this can hit analysis is Moriond 2020 timescale, no pressure
- Long lived particles (LLPs): a bonus
- Discussion
- Josh on slide 9 (sharing): how advanced is rucio adoption? Efe: must ask computing guys. Josh: before we can do something technical, it would be nice to know how much rucio is actually used.
- Efe: for SUSY, what is the purpose of comparing/sharing? Zach: mainly physics, i.e. we want to ensure that when we talk of gluinos we talk of the same thing. If we do harmonise settings, then we also have the option of sharing.
- Zach: signal production is a very tiny fraction of CPU consumption, so CPU saving is certainly not the driving force. Simone: agree, CPU not relevant here, but this is a good first example for sharing also in other cases.
- Simone: one argument against sharing is that we reduce risk for instance of a bug in one generator used by one experiment. Having two independent samples reduces risks.
- Olivier: actually sharing does not require harmonisation, we can have two samples used by both experiments and use technique like re-weighting to morfe the other samples generated to be equivalent to the dedicated one.
- Andrea: very useful discussion both for physics and CPU consumption.
- Efe: so what is the long run model, a database of samples that anyone can use? Simone: yes something like this. Zach: one can even think of doing these LHE publicly available for theorists and phenomenologists. Simone: same way, theorists can put LHE files somewhere and then experiments can analyse them.
- Andrea: what’s the more general feeling in ATLAS and CMS, apart from the people here? Zach: actually we have generator and SUSY convenors from both ATLAS and CMS here today, so there seems to be quite some agreement that this is interesting!
ATLAS and CMS event generator accounting update
- Some slides are the same as last week
- Slide 3: more accurate info from one generation, generation is about 10% of GENSIM combined, so for a total of 30% GENSIM in Grid resources, generation is 3% of all resources, accurate to within a factor two
- Slide 6: new tests from Qiang, GEN is 1% to 4% of chain (including
miniAOD) for some samples
- Note that miniAOD is negligible
- Slide 7: filled in the table
- Still some clarifications needed on CMS and ATLAS definitions
- Josh: propose to have an approximate version only for the moment, to get a gross estimate, no need for details, but split at least LO and NLO. Andrea: agree. So you split 6600M into LO and NLO? Josh: yes, already done.
- Josh: agree however to have a textual description too, so we understand better similarities and differences.
- Josh: propose to split inclusive and merged samples.
- Andrea: is this for one year of data taking? Josh: yes, will update the numbers to more closely match what CMS has filled in.
- Josh: will fill in the values for ATLAS for the specific cases suggested.
- Andrea: are there public presentations showing the lists of big productions public? Josh/Efe: there are no presentations about this really, maybe the CRSG?
- Efe: can we find some infrastructure to run benchmarking tests? Zach: Openlab are great for that, we can get dedicated machines, even bare metal, for one day or longer, with exclusive access. Andrea: I can definitely help with getting these machines, but we would need possibly a plan first, how many machines we need for how long and to do what. If we (especially ATLAS/CMS) prepare this plan, then we can easily get some openlab machines from CERN IT.
AOB
- Next meeting will normally be in two weeks on March 28.