HSF Generator Meeting #6, 27 June 2019
Agenda:
https://indico.cern.ch/event/821205
Present/Contributors: Andrea Valassi, Stephen Jiggins, Josh McFayden,
Zach Marshall, Qiang Li, Kyle Cormier, Javi Fernandez Menendez, Simone
Amoroso, Alexander Grohsjean, Graeme Stewart, Steve Mrenna
General news / Josh
- Low attendance initially due to clash with Parton Shower
Uncertainties workshop
(https://indico.cern.ch/event/827858/)
- Propose to move from biweekly to monthly meetings
- Ideas for longer term activities after the ATLAS/CMS accounting?
- Andrea: preliminary to GPU/HPC port it would be nice to get a
reproducible setup
ATLAS and CMS event generator update / Qiang and Josh
- Copied from the
note
attached to the indico agenda:
- CMS numbers in Table1 (Updates
marked in red):
- (1) Sample
Statistics:
9.4->12.3B (we missed to count some samples)
- (2) Reminder: Time/Evt numbers from Computing Experts
- G+S = 50 sec /ev/core, D = 10 sec/ev/core, R = 25
sec/ev/core : Tot:
85sec/ev/core
- (3) GEN fraction: updates from
https://indico.cern.ch/event/804266/
- Old GEN fraction ~7.3%
(1.41*0.05+1.96*0.17+0.43*0.05+1.94*0.04+0.712*0.05+0.154*0.05+0.876*0.05+1.95*0.05)/9.4
- New GEN fraction ~9.3%
(1.62*0.05+4.58*0.17+0.43*0.05+1.94*0.04+0.712*0.05+0.154*0.05+0.876*0.05+1.95*0.05)/12.2
- Where 0.05, 0.17 and 0.04 ….are from numbers achieved from local tests.
- Thus for GEN we have 85*9.3%~7.9 sec/ev/core.
- And the total GEN CPUs is 7.9*12.26B~96.9B
- (4) In Table 1, we also provided GEN CPUs for each sample category:
- For example, DY MG MLM~1.9s/event for GEN step based on local
tests, thus the total CPUs is 1.9*1618M=3.07B.
- Summing over the numbers we get in Table1 for each sample
category, we get 95.5B, which is quite close to 96.9B
- ATLAS numbers (Josh)
- (1) Changes to numbers since last time came from two places:
- Double counting of a small number of samples
- Accidental removal of samples that should be included
- --> Relatively small effect
- (2) HS06s vs s
- All the ATLAS numbers are raw seconds. Would be preferable
to have in HS06 seconds instead. Technically this is
possible but will require relatively major update to the
code, in the end it’s just a matter of finding the time to
do it.
- For reference: current tables in
https://www.overleaf.com/1326158343ftxgrxxcspxg:
- Table 1
- Table 2
- Qiang/CMS
- After update of statistics and reconstruction/digitisation/sim
times, fraction of time for generation is now 9.3% for CMS
compared to full MC production chain
- One relatively large sample is missing, will add this and make
few other updates later
- Note that 9.3% for CMS and 17% for ATLAS is relative to full
MC production chain only, not relative to the full Grid time.
Andrea: should multiply both by around 50%? Josh: MC
production is more like 70% for ATLAS.
- CMS has only one s/evt column in table 2 while ATLAS has two,
because CMS does not have the B/C filters that ATLAS uses,
which need an extra efficiency factor.
- Josh/ATLAS
- ATLAS numbers are in raw seconds, will try to get HS06 seconds
but it will take time. For CMS these are probably also in real
seconds, not HS06 seconds now. Andrea: it only makes sense to
ask ATLAS to move to HS06 if also CMS can do it, we should
have the same units at least. Qiang: will follow up with CMS
computing.
- Andrea: why ttbar+jets LO missing for ATLAS even if so big for
CMS? Josh: ATLAS chose not to generate these so far as they are
largely covered by the Powheg production, but might generate these
in the future using sherpa, which may also take quite some CPU
time. Javi: for CMS this is used for SUSY analyses only for
consistency, but for all other top analysis the ttbar inclusive is
used.
Les Houches / Josh
- A lot of discussion on negative event weights in Les Houches
(details in Josh’s slides)
- Also discussed standard test configurations (using docker), output
formats and GPU.
- There was not a huge amount of positivity about porting
generators to GPUs…
ATLAS/CMS ttbar sample sharing / Kyle and Javi
- Slide 9: differences between ATLAS/CMS come mainly from showering
(different Pythia settings), rather than from ME (different Powheg
settings)
- Josh: would it be difficult to set up an EOS area for this?
Andrea: probably not, should investigate. Zach: there is also a
public EOS area that I have already used to share data with
theorists.
- Andrea: will follow up in CERN IT
- Andrea: are LHE files actually produced already? Josh: in ATLAS we
do separate the two steps, first we produce LHE files and then we
shower to full HepMC, but CMS probably does it differently (with
LHE an internal format in memory). The timeshare is probably 50-50
% for a simple process like this.
Proposal for PDGID extension / Zach
- Zach: forwarded request to the CMS exotica conveners. Qiang: will
do the same again.
AOB
- Next meeting in four weeks, to be confirmed (or even earlier if we
have things to discuss)