HSF Generator Meeting #5, 09 May 2019
Agenda:
https://indico.cern.ch/event/813620
Present/Contributors: Andrea Valassi, Simone Amoroso, Efe Yazgan, Qiang
Li, Frank Siegert, Alexander Grohsjean, Lukasz Kreczko, Stefan Hoeche,
Steve Mrenna, Marek Schoenherr, Markus Schulz, Stephen Jiggins, Josh
McFayden
General news
- Efe: Alexander new convener in CMS replacing Efe after August
ATLAS and CMS event generator update
- Simone: slides with an update about generators
- This is an update on the numbers presented at the November
workshop.
- These numbers might soon become an ATLAS public document.
Eventually also the configurations used to produce them might
be shared.
- Slide 3: Powheg plus Herwig faster than Powheg plus Pythia.
- Slide 5: yellow is sherpa with FxFx scale (i.e. the same scale
used by MG). This makes sherpa faster by a factor. We plan to
exercise this in ATLAS. It may even become the default in
sherpa, the authors are looking at that. StefanH: yes we get
encouraging results, but still need more tests.
- Qiang: does FxFx affect the negative weight fraction? Simone:
did not check. Marek? If anything the negative weight fraction
should get better.
- StefanH: note that MG + Pythia for NLO can have MG unweighted,
but then when you add Pythia this also gives you Pythia
weights, and you have to unweight again. You could store all
events and then unweight. There is some similarity to what
alpgen does.
- Slide 7: something a bit unrelated to the previous slides,
strategies for reducing negative weights in MG aMC@NLO.
- Andrea: rule of thumb for impact of negative weights?
Simone/Frank: some rules of thumb plus some refinement.
Andrea: please send them to me or put them on the minutes.
- Slide 7: Qiang, is there a similar way to reduce the negative
weights in sherpa in the same way as in MG? Stefan: most
negative weights come from h events rather than s events, and
a folding strategy is not possible. Stefan: have a feeling
that going to a full powheg implementation may be a better
strategy. Efe: MG are working on something similar to the
powheg method, which may be more effective.
- Simone: would be useful if Josh/Andrea ask the MG authors to
report when this becomes available.
- Efe and Qiang: updated the numbers in
https://www.overleaf.com/1326158343ftxgrxxcspxg
- Table 1 is now updated. Josh: one thing that changed is that
we removed the “Others” section, so the ATLAS number is
reduced and it is more apples to apples.
- Take away: ATLAS 10x slower in generation, 5x slower in
detector simulation, 2x slower in reconstruction for the
same number of events!
- Alexander: is ATLAS fast sim or full sim? Josh: it’s a
mixture of fast sim and full sim.
- Andrea: is this table complete now? Josh: yes, apart from
simulation time pointed out by Alexander, this is now
pretty complete for generators, nothing missing or nothing
too much.
- Qiang: Table 3 is a full apples to apples for LO.
- Efe: here there is a good agreement. There is 30 seconds per
event in both cases.
- Frank: however we must make sure we are not biased by filters.
For instance there is a difference between the first two
columns for timing, still something to understand. In
particular, it is becoming apparent that ATLAS makes stricter
filters and then produces more expensive, but fewer, events.
- Table 2 then is apples to apples for NLO. Here sherpa is a factor
20x slower!
- Josh: will write down in text some comments to help
interpretation of some details.
- Andrea: totals are 80M events ATLAS and 4B CMS. Ideally it
would be nice to get total CPU in Table 1 for each row if
possible. Josh: should be easy for ATLAS. Efe; more complex
for CMS.
- Andrea: very good progress. We should eventually converge on
making this a writeup of the November workshop. Andy Buckley
volunteered to help with the writeup of the workshop proceedings,
now he has more time free from other commitments
Proposed CHEP 2019 abstract
“Addressing the software and computing challenges of physics event
generators”
The HL-LHC physics program will require unprecedented computing
resources for simulated collisions and therefore Monte Carlo (MC) event
generation. The number of MC events to generate and simulate, which
scales as the integrated luminosity of real collisions, is expected to
increase at a much faster rate than the available computing resources.
The fraction of CPU consumed by MC event generators will also increase
dramatically, as the LHC experiments are expected to use primarily fast
detector simulation. In addition, the availability and reliance upon
higher precision theoretical calculations is also expected to increase,
resulting in increased computing resource requirements for MC generation
compared to today. Consequently, to ensure physics results are not
restricted by a limited number of MC events that can be generated on the
available resources, significant optimisation upon the current model is
required.
To achieve this, first a detailed analysis of the current computing
model for MC event generators is required, followed by optimisation to
improve generator software performance. In addition, developing
generator software to be able to exploit the evolution of computing
architectures in the HL-LHC era, which will likely take the form of
accelerator-based High Performance Computing clusters, will be vital.
Such architectures are better suited to the smaller code-bases of MC
generators rather than the vast code bases of the LHC experiments
simulation and reconstruction infrastructures. Finally, more precise
projections of the generator-level physics requirements for HL-LHC are
also required from theorists and experimentalists.
The HEP Software Foundation has recently created an Event Generators
Working Group that aims to bring together MC generator authors, LHC
experiment users and software experts to address these issues. The first
results of this working group are presented here. This includes a
detailed study of the ATLAS and CMS generator usage during Run 2,
benchmarking of different MC generator CPU performance, and first
studies on using MC generator code on new architectures.
- Andrea: propose to send this abstract to CHEP, with an initial
list of authors including WG conveners and those that have so far
been most active with presentations (experiment WG convenors and
Zach), unless they want to opt-out, and myself as initial
candidate speaker. However all WG members can also choose to
opt-in and ask to be included as authors if they feel this
appropriate. Then we still have time to decide about the actual
speaker and even more time for the actual list of authors.
- Efe/Qiang: sounds good.
AOB
- Andrea: would be nice to resume work and discussion on GPUs in
future meetings.
- Simone: Taylor made a presentation on VEGAS for GPUs in an
ATLAS presentation, we could ask him to give more details.
Andrea: yes, I have also discussed with him his work on alpgen
on PowerPC.
- Frank: note this was work from a few years ago, ATLAS is not
really using alpgen now. Stefan: true, alpgen less used now,
but the techniques Taylor and colleagues had used are still
useful. We are now working on an integration of sherpa on
HPCs, reusing some of the technicalities that Taylor had used
for alpgen. This is using MPI. Even going to many jets, you
can cut memory and fit on existing machines, so you can fit in
a KNL based HPC.
- Andrea: very good to know this also fits on KNL, but am really
more worried about GPUs on HPCs in the future. Stefan: yes we
definitely must work on GPUs, anyway KNL based HPCs exist and
are there to stay with us for some time.
- Josh: speaking of new topics, we should go back to the question of
how to get more people involved in the coding effort.
- Stephen: agree, getting the grants to do software porting is
important, I see this now as a postdoc working on code.