Subject Area D Cost- and energy-efficient use of computing resources - Indico
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Subject Area D
Cost- and energy-efficient use of computing resources
Florian Bernlochner Kick-Off Meeting, Feb 2019
Direct Searches
Overview of Area D
1. Obvious path to NP:
`
New p
Introduction: . . . (1/2)
Int
i.g Direct
Physics? versus indirect searches
os
ens
m
SM p
os
ct Searches ity
C
2
Indirect Searches
E = mc New
NewPhysics
Physics
AAAB8HicbVBNSwMxEJ2tX7V+VT16CRbBU9ktgl6EoggeK9gPadeSTbNtaJJdkqxQlv4KLx4U8erP8ea/Md3uQVsfDDzem2FmXhBzpo3rfjuFldW19Y3iZmlre2d3r7x/0NJRoghtkohHqhNgTTmTtGmY4bQTK4pFwGk7GF/P/PYTVZpF8t5MYuoLPJQsZAQbKz3coEskEHms9csVt+pmQMvEy0kFcjT65a/eICKJoNIQjrXuem5s/BQrwwin01Iv0TTGZIyHtGupxIJqP80OnqITqwxQGClb0qBM/T2RYqH1RAS2U2Az0oveTPzP6yYmvPBTJuPEUEnmi8KEIxOh2fdowBQlhk8swUQxeysiI6wwMTajkg3BW3x5mbRqVc+tendnlfpVHkcRjuAYTsGDc6jDLTSgCQQEPMMrvDnKeXHenY95a8HJZw7hD5zPH8xDjxY=
Energy
Obvious path to NP: Search for New Physics at the Energy Frontier
1. Less
2. Needobvious
a targetpath to NP
region:
bility Challenge (SC)
`
New Region
New Physics
Physics of phase space,
Common
. . . Problem:
process, .. (e.g. NP ! dijets)
p
Moore’s Law
New Physics 3. Models can have large
Scalability
CPU Performance
Big Data
revolution Challenge
parameter space that need to
Data Volume
Need a target region: beNew
Search for
2. Need scanned
aPhysics
processat the Intensity Frontier
that can be
Power wall for
one processor
Region of phase space, predicted at high precision for
Modelthe
driven
SM or ’Bump’ hunting
log
process, .. (e.g. NP ! dijets)
Models can have Time
large 3. Model independent searches
n parameter space that need to Power wall
Both based on general
approaches operators
complement each other
Florian Bernlochner Kick-Off Meeting, Feb 2019 !2
weder durch höhere Effizienz oder durch restriktivere Trigger-Entscheidungen, die allerdings
das phyikalische Potential der Experimente reduzieren, das Bedarfswachstum zu reduzieren.
Overview of Area D
Eine weitere Herausforderung, insbesondere für Experimente an Hadron-Beschleunigern, ist
die begrenzte Verfügbarkeit von großen Arbeitspeichern bei wachsender Komplexität der Ein-
zelereignisse.
Lösungsansätze können auf verschiedenen Ebenen gefunden werden. Im Rahmen dieses Teil-
projekts stehen insbesondere die Entwicklung alternativer Algorithmen (z.B. Cellular Automa-
ton-basierte Ansätze statt Kombinatorischer Kalman Filter), bessere Software Codes, die z.B.
größere Datenlokalität und single instruction, multiple data (SIMD) Optimierungen möglich ma-
chen, und alternative Software-Architekturen (z.B. GPUs) zusammen mit einem grösseren Maß
an Parallelisierung im Mittelpunkt. Um angesichts des erforderlichen Expertentums auch die
Personalressourcen optimal zu nutzen, sollten die Entwicklungen möglichst in gemeinsam ge-
nutzten Bibliotheken gebündelt werden und Beispielanwendungen zum besseren Verständnis
zur Verfügung stehen. Der größte Teil der Entwicklungen, die im Rahmen dieses Projekts an-
gestrebt werden, wird daher mit Hilfe von ACTS (A Common Tracking Software) angebunden.
4.2 Förderziele und Arbeitsstruktur
Three work packages:
Ziel der Entwicklungen ist es experimentübergreifende Bibliotheken weiterzuentwickeln, die den
1.Kosten-
NovelundAlgorithms for track
Energie-Aufwand für dasreconstruction
Computing im Rahmen der Ereignisrekonstruktion in ver-
2.tretbarem
Common Rahmen halten,
tools ohne übermoptimal
to determine äßig an Sensitivit
track ät zu verlieren. Die Arbeitspakete, die
auf die Rekonstruktion von Spuren geladener Teilchen zielen, können grob in Algorithmen zum
parameters
Auffinden der Spuren und zur Bestimmung der optimalen Parameter aufgeteilt werden. Neutri-
3.noexperimente unterliegen speziellen
Event reconstruction at neutrinoAnforderungen durch das große Detektorvolumen, wel-
experiments
ches abgedeckt werden muss. Die wichtigsten Elemente der Arbeitspakete können in Tabelle
7 gefunden werden.
Florian Bernlochner Kick-Off Meeting, Feb 2019 !3
AP D1: Novel Track reconstruction algorithms
• Computing needs scale roughly as CPU Power ≃ ℒ3
IC Particle Physics
2026
inst. Luminosity
37
• HL-LHC will need need smarter algorithms to
Today
The AESC Team
identify charged particle trajectories
ATLAS proton-proton event CMS lead-lead collision event
Florian Bernlochner
Florian Bernlochner
‣ 50%(!) of event reconstruction time used for track reconstruction.
• Entwicklung effizienterer Algorithmen zur Spurfindung, die beispielsweise auf zellulären
Automaten aufbauen. CA can be easily parallelised; first promising results using this at CMS
(pixel-track seeding)
The AESC Team
• Anpassung der Algorithmen zur Nutzbarmachung moderner massiv paralleler Rechner-
architekturen (GPUs).
Today
37
Florian Bernlochner Kick-Off Meeting, Feb 2019 !4
AP D2: Common tools for tracking
• Instead of each experiment writing their own
tracking algorithms
Often algorithms are not optimised for CPU usage
and small runtimes
http://acts.web.cern.ch
→ pool resources and expertise
• ACTS (“A common tracking software”) is an attempt to create such a
common library, which implements many of the often used
algorithms;
many implemented algorithms much faster than what experiments currently are using
• Goal: Help to extend ACTS to become fully usable by energy and intensity
frontier experiments
Add missing features, evaluate how aspects of the library can be used to improve the
current track finding and fitting.
Florian Bernlochner Kick-Off Meeting, Feb 2019 !5
AP D3: Event reconstruction at neutrino experiments
• Neutrino experiments have unique
challenges:
• Huge detector volumes; sparse sensors
that observe signals from a distance
• Inclusion of time-information important
to make sense of detector signals
• Signatures typically have multiple
components: e.g. Muon and hadronic
recoil in CC-scattering of muon
neutrinos
• Immense amount of information that
needs to be processed: e.g. JUNO uses
O(17k) PMT signals to reconstruct a
single muon track. A single fit to do this
can take O(1h)
Florian Bernlochner Kick-Off Meeting, Feb 2019 !6
Involved Groups
Thomas Kuhr Computing Verbund Kickoff Meeting, 21.02.2019 Page 4
Standort PI FTE Experiment AP D1 AP D2 AP D3
Aachen Schmidt 0,5 CMS X
Aachen Stahl 0,5 ⌫-Exp. X
Frankfurt Kisel 1 CBM X
KIT Bernlochner 1 Belle II X
Assoziiert
CERN(ATL/SFT) Elsing/Hegner - ATLAS X X
DESY Gaede - ILC X
FZJ Prencipe - Panda X
Tabelle 8: Beteiligte Gruppen. Die assoziierten Gruppen sind nicht-universitäre Partner die ke
Florian
ne Bernlochner
Mittel beantragen. Kick-Off Meeting, Feb 2019 !7
Activities Overview
D (& C)
Florian Bernlochner Kick-Off Meeting, Feb 2019
AP D2: KIT
Belle
https://indico.physics.lbl.gov/indico/event/712/
Florian Bernlochner Kick-Off Meeting, Feb 2019 !9
ACTS Workshop Overview
• 22 Participants from various experiments
• 2 ACTS experts: Dr. Moritz Kiehn und Paul Gessinger
• Members from ATLAS, Belle II, and Mu2e
• Diverse 5 day program
• Mix of tutorials from Moritz and Paul;
• Talks from Belle II (Dr. Nils Braun), ATLAS (Dr. Nick Styles), Mu2e (Dr.
Dave Brown) tracking challenges
Hackathon @ KIT
• A lot of Hackathon time to kick-start projects
ackathon
ogether and use the
open tasks to improve
analysis
Florian Bernlochner Kick-Off Meeting, Feb 2019 !10Example Projects
• Full list available here:
https://docs.google.com/presentation/d/1agWKZN0SiIwxptRfx6kIwx9GimhbvHdNTwCk2Afef9Q/edit#slide=id.p
Florian Bernlochner Kick-Off Meeting, Feb 2019 !11First results: Belle II Geometry in ACTS
TRY GEOMETRY
Belle II
Belle II Drift Chamber
VXD
ACTS Workshop Berkeley - Nils Braun
y - Nils Braun January 26, 2019 17/25
Florian Bernlochner Kick-Off Meeting, Feb 2019 !12First results:
MAGNETIC FIELD
Belle II Magnetic field
ACTS Workshop Berkeley - Nils Braun Janua
Florian Bernlochner Kick-Off Meeting, Feb 2019 !13First results: Track extrapolation
Abstract
The extrapolation of track parameters and their associ-
ated covariance matrices to arbitrarily oriented surfaces of
different types inside a non-uniform magnetic field is a fun-
damental element of any tracking software. It has to take pB
multiple scattering and energy loss effects along the prop- mB
agated trajectories into account. A good performance in
eB
respect of computing time consumption is crucial due to fac
Sur
hit and track multiplicity in high luminosity events at the mA pA
eA
fac
EXTRAPOLATION LHC and the small time window of the ATLAS high level r
Su
trigger. Therefore stable and fast algorithms for the trans-
port of the track parameters and their associated covariance Figure 1: Schematic representation of the transportation of
matrices in specific representations to different surfaces in track parameters and their associated errors from Surface
the detector are required. The recently developed track ex- A to Surface B. The track parameters on surface are il-
trapolation package inside the new ATLAS offline tracking lustrated by a local position and its error ellipse, such
software is presented in this document. as a momentum vector and a cone representing the error
on the momentum direction. The error on the magnitude of
Belle II INTRODUCTION the momentum is omitted in this illustration.
algorithm During the recent redesign of the ATLAS offline recon-
be retrieved from this service at any point in the program
versus struction software, a new track extrapolation package has
been developed within the C++ based software framework
flow.
ACTS ATHENA [1]. The transportation of track parameters and
The steering of the propagation setup, including the
setup of the magnetic field, the propagation algorithm and
their associated covariance matrices to a given detector sur-
surface finding logics is done by dedicated python classes.
face is a fundamental and frequently performed process in
The extrapolation process can be driven in two differ-
most track fitting algorithms. In general, the extrapolation
ent modes, a preconfigured and an unconfigured mode. In
process can be divided into two parts. The first step is the
the preconfigured mode, the underlying propagation setup
geometrical transport of the track parameters respectively
(type of propagation, magnetic field setup) is fully deter-
covariance matrices to given surfaces and will be in the fol-
mined at startup of the program and should be used for ex-
lowing referred to as propagation, Fig. 1 shows a simplified
pectable situations in the program flow.
illustration of such a propagation. The second procedure is
the update of the propagated parameters and errors, taking The unconfigured mode is characterized by the fact that
multiple Coulomb scattering and energy loss effects dur- the Propagator AlgTool itself is passed to the Extrapolator
ME = Magnetic effects ing the (turned
propagation on orinto
process off), IF =This
account. interpolated field of
note covers AlgTool, Bellea strategy
following II pattern design. This allows
mainly the first part of the extrapolation process. an optimization of the extrapolation process by dynami-
(instead of constant 1.5 T) cally choosing the propagation algorithms depending on
the situation specific parameters, such as the magnetic field,
ACTS Workshop Berkeley - Nils Braun THE EXTRAPOLATION PACKAGE an estimated propagation January
distance26,or2019
even starting track
20/25pa-
Florian Bernlochner Kick-Off Meeting, DESIGN
Feb 2019
rameters characteristics. Various instances of Extrapolator
AlgTools in different configurations can be used in parallel. !14There are not that many conceptual show-stoppers for just ”plugging the
Conclusion ACTS extrapolation into Belle II instead of genfit”:
No DAF (should be solvable easily)
No slanted parts (but not a conceptional problem with it)
Their Kalman is working a bit differently from ours, but should be no problem
Their CDC hit handling is different (maybe this will make a difference,
maybe not)
However, if we want to replace genfit with ACTS, there is a lot more to do:
The EDM is different
The measurement model is different
No CKF, no T0
How about extrapolations in ECL, (B/E)KLM etc.? Dr. Nils Braun
• Could carry out first preliminary studies;
ACTS Workshop Berkeley - Nils Braun Januar
• KIT group could connect with the ACTS developers
• Involved people at KIT: me, Dr. Nils Braun, Patrick Ecker, Lu Cao, Pablo
Goldenzweig, Michael Eliachevitch
• Plan to organize follow up workshop in August or September in
Germany with a similar style of agenda
• Tutorials, some overview talks but ample time to work on dedicated
projects with ACTS experts in the room to help.
Florian Bernlochner Kick-Off Meeting, Feb 2019 !15AP C3: KIT
• Simulation of extensive air showers or hadronic /
electromagnetic showers from first principles are
time intensive tasks
• Can we delegate the entire simulation (or parts of
it) to a neural network?
Dr. Markus Roth
Key challenge: sparse input features, e.g. particle type,
energy, mass, etc. —> misses randomness of MC simulation
Florian Bernlochner Kick-Off Meeting, Feb 2019 !16You can also read
