Event generators' and N(n)LO codes' acceleration

13 Nov 2023, 09:00 → 14 Nov 2023, 18:45 Europe/Zurich

4/3-006 - TH Conference Room (CERN)

The Workshop will address the challenges to improve the computing performance of Monte Carlo event generators and higher-order perturbative calculations, covering in particular the transition to heterogeneous hardware architectures beyond CPU (eg GPUs).  

The event, the first in what will likely become a series, will benefit from the participation and contributions of experts from TH, EP and IT communities. A general structure for the agenda will be announced shortly. Participants are invited to propose, through the "comments" area of the registration page, topics or discussion items, which we'll do our best to fit in the overall agenda.

The meeting will be in presence, with zoom connection available. A welcome drink will be offered on Monday, following the end of the afternoon sessions.

Organization committee:

• Bugra Bilin 
• Philip Ilten 
• Leif Lonnblad
• Michelangelo Mangano
• Olivier Mattelaer
• Josh McFayden
• Stefan Roiser
• Andrea Valassi 

Monday 13 November

09:00 → 09:15 Introduction

Speaker: Michelangelo Mangano (CERN)

introduction.mp4

Mangano_Intro.pdf

09:15 → 10:15 The experience of the experiments

09:15 ATLAS

Speaker: Tetiana Moskalets (Albert Ludwigs Universitaet Freiburg (DE))

ATLAS_EvGenAcc.pdf

atlas_intro.mp4

09:45 CMS

Speaker: Sitian Qian (Peking University (CN))

cms_intro.mp4

GenAcceWS_SQ_V4.pdf

10:15 → 14:45 The experience of the MC event generators

10:15 Herwig

Speaker: James Whitehead

herwig.mp4

Performance in H7

11:00 Coffee break

11:30 Pythia

Speaker: Leif Lönnblad (Lund University (SE))

AcceleratingPythia.pdf

pythia.mp4

12:15 Sherpa

Speaker: Enrico Bothmann (University of Göttingen)

bothmann-sherpa.pdf

sherpa.mp4

13:00 Lunch break

14:00 MG5_aMC@NLO

Speaker: Olivier Mattelaer (UCLouvain)

23_11_CERN_acceleration.pdf

mg5_amcNLO.mp4

14:45 → 15:30 Coding challenges/opportunties generated by new hardware

Speaker: Stephan Hageboeck (CERN)

23.11.13-Challenges_Opportunities_SIMD_GPUs.pdf

coding_challenges_Hageboeck.mp4

15:30 → 16:00 Capitalise on existing HEP software expertise to upgrade event generation

Speaker: Danilo Piparo (CERN)

231113_Generators Workshop.pdf

SFT_Piparo.mp4

16:00 → 16:30 Coffee break

16:30 → 17:30 The HPC landscape

16:30 ATLAS perspective

Speaker: Andrej Filipcic (Jozef Stefan Institute (SI))

ATLAS_HPC.mp4

Event Generators @ HPCs.pdf

17:00 CMS perspective

Speaker: Elizabeth Sexton-Kennedy (Fermi National Accelerator Lab. (US))

CMS_HPC.mp4

GEN-HPC_231113-OnC .pdf

HPC Landscape - CMS perspective

17:30 → 18:15 Accelerating Madgraph5_aMC@NLO using C++ vectorization and GPUs

Speaker: Andrea Valassi (CERN)

20231113-MG5aMConGPU-MCWS-AV-v007.pdf

20231113-MG5aMConGPU-MCWS-AV-v007.pptx

Mg5_amc_acceleration_Valassi.mp4

18:30 → 20:00 Welcome drink

Tuesday 14 November

09:00 → 10:30 Phase-space sampling

09:00 Event generation performance & new workflows

Details and recent findings on the performance of heavy-hitter Sherpa setups as used in production e.g. by ATLAS, and discuss new (HPC) workflows, such as HDF5 based parton-level event storage etc

Speaker: Christian Gutschow (UCL (UK))

cg_cern_performance_workshop.pdf

ChristianGutschow.mp4

09:30 The MadNIS Reloaded – Boosting MG5AMC with Neural Networks

We combine machine-learned multi-channel weights with a normalizing flow for importance sampling to improve classical methods for numerical integration. By integrating buffered training for potentially expensive integrands, VEGAS initialization, symmetry-aware channels, and stratified training, we elevate the performance in both efficiency and accuracy of the MadNIS framework. We empirically validate these enhancements through rigorous tests on diverse LHC processes, including VBS and W+jets

Speaker: Ramon Winterhalder (ITP Heidelberg)

madnis_rw.pdf

RamonWinterhalder.mp4

10:00 Towards a framework for GPU event generation

Speaker: Dr Juan M. Cruz Martinez (CERN)

jcm.pdf

JuanMCruzMartinez.mp4

10:30 → 11:00 Coffee break

11:00 → 13:00 LO & NLO issues

Matrix element acceleration, GPU porting, negative-weight reduction, ...

11:00 Portable event generation with PEPPER

Details and results on the portable parton-level event generation framework that we are developing, also with respect to the algorithmic choices for the tree-level calculation

Speaker: Max Knobbe (University of Göttingen)

knobbe_pepper.pdf

MaxKnobbe.mp4

11:30 Factorization-aware neural networks: NLO MEs and unweighting

Speaker: Mr Daniel Pierre Maitre

DanielPierreMaitre.mp4

NN_MEemulation.pdf

12:00 Optimising loop amplitude evaluations

I will discuss some of the current challenges for the evaluation of multi-loop amplitudes. Neural networks have had some success for the evaluation of tree-level and one-loop amplitudes that may also be useful in the context of higher order computations. I will show some simple applications to simulations of processes with no born level contributions, gg->di-photon+gluons, and discuss potential improvements

Speaker: Simon David Badger (Universita e INFN Torino (IT))

SB_CERNMC_2023.pdf

SimonDavidBadger.mp4

12:30 Removing negative weights in Monte Carlo event samples

Negative Monte Carlo event weights greatly hamper the statistical convergence, leading to large samples and incurring a huge increase in computational cost for subsequent steps in the simulation chain. I discuss the cell resampling method for cancelling negative weights inside small phase space regions and present recent improvements that render the application to large high-multiplicity samples feasible

Speakers: Andreas Maier (DESY), Dr Andreas Martin Maier (DESY)

AndreasMaier.mp4

cell_resampling.pdf

13:00 → 14:00 Lunch break

14:00 → 15:00 PDF and hadronization issues

14:00 PDFflow

Speaker: Stefano Carrazza (CERN)

pdfflow_carrazza.pdf

StefanoCarazza.mp4

14:30 Construction and Fitting of a Deep Generative Hadronization Model

Encouraged by the success of applying machine learning techniques to the non-perturbative PDF problem, we decided to try the same in the context of non-perturbative hadronisation. In this presentation, I will show the first steps we took to construct HADML a Deep Generative Hadronization Model.
Then, I will describe the protocol we created to fit the Deep Generative Hadronization Model in a realistic setting where we only have access to the set of final stat hadrons in the data.

Speaker: Andrzej Konrad Siodmok (Jagiellonian University (PL))

AndrzejKonradSiodmok.mp4

CERN_MC_workshop_Siodmok_2023-3.pdf

15:00 → 17:45 NNLO and beyond

15:00 Computer algebra challenges in precision calculations

Speakers: Thomas Gehrmann (Univ. Zurich), Thomas Kurt Gehrmann (University of Zurich (CH))

CERN_acceleration.pdf

ThomasGehrmann.mp4

15:20 Numerical challenges in precision calculations

Speaker: Alexander Yohei Huss (CERN)

AlexanderYoheiHuss.mp4

Numerics_Huss.pdf

toy-nlo

15:45 Coffee break

16:15 pySecDec: Experiences Evaluating Multi-loop Amplitudes on GPUs

In this talk, I will review some of our experiences implementing the evaluation of multi-loop integrals and multi-loop amplitudes on GPUs using the sector decomposition method. I will highlight recent developments and how they have impacted the performance of this method and I will summarise some of the main lessons learned. Additionally, I will briefly comment on potential avenues/obstacles to utilising heterogeneous computing platforms for other numerical/semi-numerical approaches to evaluating Feynman Integrals and integral reduction.

Speaker: Stephen Philip Jones (University of Durham (GB))

StephenPhilipJones.mp4

talkv1.pdf

16:45 Profiling NNLO+PS simulations: GENEVA as a case study

In this talk I will review the stages and parallelization strategies used in NNLO+PS simulations, focusing on the GENEVA Monte-Carlo code. Concentrating on three benchmark processes, Drell-Yan, Higgs production via gluon fusion and diboson production, I will report results of a profiling exercise aimed at identifying the current bottlenecks in each stage that will benefit the most from possible improvements connected with the usage of alternative computing architectures or different approaches based e.g. on machine-learning techniques.

Speaker: Simone Alioli (Universita & INFN, Milano-Bicocca (IT))

Alioli-CERN-WS.pdf

SimoneAlioli.mp4

17:15 Issues in the parallelization of physics algorithms on GPU

I will briefly report on two distinct attempts to port on GPU an algorithm commonly used in the evaluation of radiative corrections to particle scattering processes:
1) the reconstruction of a rational function, useful in the exact numerical solution of the linear systems typically present in multi-loop calculations;
2) the generation of events via a parton-shower algorithm.
In both cases some interesting indications emerge, related to the logical mistakes that one should avoid in the parallelization of algorithms which originally are designed for sequential evaluation.

Speaker: Alessandro Vicini (Università degli Studi e INFN Milano (IT))

AlessandroVicini.mp4

Vicini-Issues-GPU.pdf

17:45 → 18:30 Final discussion

17:45 Discussion of priorities, plans and resources needs

ClosingDiscussion.mp4

FInal_discussion.pdf