DEPARTMENT OF PHYSICS MAY2021 Future of computing at the Large Hadron Collider_2

2025-05-06 0 0 4.22MB 26 页 10玖币

侵权投诉

DEPARTMENT OF PHYSICS

MAY 2021

Future of computing at the Large Hadron Collider

Author:

Dhananjay Saikumar

Academic Supervisors:

Dr Daniel O'Hanlon &

Prof Jonas Rademacker

FOR SUBMISSION IN 2021

arXiv:2210.13213v1 [hep-ph] 27 Sep 2022

Abstract

High energy physics (HEP) experiments at the LHC generate data at a rate of O(10) Terabits per second. This data rate

is expected to exponentially increase as experiments will be upgraded in the future to achieve higher collision energies. The

increasing size of particle physics datasets combined with the plateauing single-core CPU performance is expected to create

a four-fold shortage in computing power by 2030. This makes it necessary to investigate alternate computing architectures to

cope with the next generation of HEP experiments. This study provides an overview of different computing techniques used

in the LHCb experiment (trigger, track reconstruction, vertex reconstruction, particle identiﬁcation). Furthermore, this research

led to the creation of three event reconstruction algorithms for the LHCb experiment. These algorithms are benchmarked on

various computing architectures such as the CPU, GPU, and a new type of processor called the IPU, each roughly containing

O(10),O(1000), and O(1000) cores respectively. This research indicates that multi-core architectures such as GPUs and IPUs

are better suited for computationally intensive tasks within HEP experiments.

Contents

1. Introduction 4

2. Physics at the LHCb 5

2.1. CP violation 5

2.2. The Cabibbo-Kobayashi-Maskawa matrix 6

2.3. BSM signatures 6

2.4. Schematic of the LHCb detector 7

3. Computing architectures 7

3.1. Central Processing unit 7

3.2. Graphics Processing Unit 8

3.3. Intelligence Processing Unit 8

3.4. Tensor Processing Unit 8

4. Trigger system 8

5. Event reconstruction 9

5.1. Track reconstruction 9

5.1.1. Vertex locator (Velo) 10

5.1.2. Upstream tracker (UT) 11

5.1.3. Sci-Fi tracker 11

5.1.4. Muon detector 12

5.1.5. K´

alm´

an ﬁlter 12

5.2. Vertexing 12

5.3. Particle identiﬁcation 13

5.3.1. RICH 13

5.3.2. Calorimeter 13

6. Algorithms and Results 14

6.1. The predictive combinatorial seeding algorithm 14

6.1.1. Design 14

6.1.2. Results 15

6.2. The artiﬁcial retina algorithm 16

6.2.1. Design 16

6.2.2. Results 17

6.3. Convolutional neural network based track reconstruction 18

6.3.1. Supervised Machine Learning basics 18

6.3.2. Neural Networks 18

6.3.3. Design 19

6.3.4. Results 20

6.4. The trackless vertex ﬁnder 21

6.4.1. Design 21

6.4.2. Results 21

7. Discussion 22

7.1. Computing performance and limitations 22

7.1.1. ML optimization 22

7.1.2. Parallelism 22

7.1.3. TensorFlow (frontend vs backend) 23

7.2. Other computing avenues in HEP 23

8. Summary and Conclusions 24

References 25

1 Introduction

The Large Hadron Collider (LHC), located in Geneva

Switzerland, is the largest particle accelerator ever built to-

date, designed and constructed by the European Organization

for Nuclear Research (CERN). The LHC is made up of a 27-

kilometre underground tunnel ﬁlled with strong magnets that

guide and boost beams of protons up-to velocities 3.1 m/s

slower than the speed of light [1]. Protons beams travelling

in the opposite direction are designed to collide in four dis-

tinct regions corresponding to the four major particle physics

experiments at the LHC (see ﬁgure 1):

– A Toroidal LHC Apparatus (ATLAS): Largest General pur-

pose experiment at the LHC, designed to investigate a wide

spectrum of physics, ranging from: the Higgs boson, CP

violation1, BSM2, Dark matter, and more [2].

– Compact Muon Solenoid (CMS): Another General purpose

experiment at the LHC which works in conjunction with

the ATLAS Experiment (e.g: the joint discovery of the

Higgs boson in 2012). It has the same scientiﬁc goals as

ATLAS; however, the CMS detector operates on a com-

pletely different design approach [3].

– Large Hadron Collider beauty (LHCb): The LHCb is a spe-

cialised experiment designed to study the effects of CP vio-

lation in beauty-hadron systems (shedding light on the mat-

ter anti-matter asymmetry observed in the universe), mea-

suring forward-backward asymmetry in FCNC3decays,

searching for BSM phenomena, and more [4].

– A Large Ion Collider Experiment (ALICE): ALICE is de-

signed to study heavy-ion physics. Collisions of Pb-Pb nu-

clei allow ALICE to investigate the ﬁfth state of matter: the

quark-gluon plasma (QGP). In this state, quarks and gluons

are disentangled, resulting in conditions similar to a frac-

tion of a second after the Big bang [5].

Figure 1: The top of the ﬁgure represents the LHC and positions of

the four major experiments. The bottom of the ﬁgure depicts small

accelerators which inject hadrons into the 27KM LHC ring. Figure

taken from Ref [6].

1Charge conjugation Parity symmetry

2Physics beyond the Standard Model

3Flavor-changing neutral currents

The energy of the colliding particle beams is transformed

into matter via Einstein’s mass-–energy relation eq(1) in the

COM4frame, creating a wide shower of particles.

E=mc2(1)

Each collision is referred to as an event. In every event, the

position (hits) and momentum of newly created particles are

measured by the detectors which surround the interaction re-

gion. The raw data from the detector is used to reconstruct the

collision events in a process known as event reconstruction,

which involves identifying the particles, their energy, momen-

tum, trajectory, and their point of origin, hence measuring the

process that took place at the collision [7]. The ﬁrst run of

the LHC (2009) achieved 7 TeV of collision energy, subse-

quently increased to 13 TeV during Run 2 (2015). The LHC

was shutdown in 2018 for additional upgrades and is sched-

uled to resume operations in 2022 (Run 3) with an expected

collision energy of 14 TeV [8].

The amount of data generated at each event is expected to

increase as HEP5experiments are upgraded, since more mat-

ter is created at higher energies. The four-major experiments

at the LHC during Run 3 are expected to generate data at a

rate of O(10) Terabit/s [9]. Storing all of this data in real-

time for post-analysis is logistically infeasible, which is why

experiments at the LHC use a trigger system. By partially re-

constructing the events in real time, a trigger system selects

and stores interesting datasets for detailed post analysis. The

LHC is expected to generate 40 million bunch proton-proton

(pp) collisions per second (40 MHz). Hardware level triggers

in the ATLAS and CMS aim to reduce this throughput rate

down to O(1) KHz [7]. In contrast, the upgraded LHCb ex-

periment will use a software based trigger running entirely on

GPUs operating at the full bunch collision rate [10].

Figure 2: A pie chart of projected CPU resources (by computa-

tional task) required by ATLAS in 2030.

In addition to the immense computing resources required

by the trigger, HEP experiments also ubiquitously use Monte

Carlo (MC) simulations, mainly for modelling the collision

4Center of momentum

5High Energy Physics

events, detector response, detector calibration, and more (see

ﬁgure 2for CPU resource allocation by task). A steep rise in

computing resources is needed to keep up with the exponen-

tially increasing size of particle physics datasets. Future ex-

periments would not only face an engineering challenge, but

a computational one as well since these experiments would

be limited by how efﬁciently the computing resources can be

used. A four-fold shortage in computing power (ﬁgure 3) is

forecasted by 2030 [11]. This is because, although the transis-

tor density of a CPU6has been increasing over time; obeying

Moore’s law (albeit is decelerating), the single-core CPU per-

formance has plateaued since the mid-2000s due to the con-

straints on power density [12].

Figure 3: Estimated CPU resources needed (blue points) by the

ATLAS experiment from the years 2018 to 2028. The solid line rep-

resents the amount of resources expected to be available assuming

a ﬂat funding scenario of 20% per year. Other experiments at the

LHC follow similar resource-forecast trend. Figure taken from [11]

Due to these performance constraints, particle physics ex-

periments such as ATLAS, COMET, ALICE, CMS, and

LHCb [13–15] are investigating and implementing alternate

computing techniques and multi-core architectures (i.e soft-

ware based GPU trigger at LHCb) to cope up with the ever in-

creasing computing demands of HEP experiments. These in-

clude heterogeneous systems where computing architectures

such as the GPUs7and FPGAs8work in conjunction with the

CPUs. A modern CPU contains O(10) powerful cores built

on the MIMD9design, whereas the GPU contains O(1000)

simpler cores built on the SIMD10 design, making the GPU

ideal for running multiple identical computations in parallel.

On the other hand, FPGAs are integrated circuits that are not

hard etched; unlike CPUs and GPUs, this allows FPGAs to be

reprogrammed for speciﬁc computations, often resulting in far

superior performance compared to its traditional counterparts.

6Central Processing Unit

7Graphics Processing Unit

8Field programmable gate arrays

9Multiple Instruction, multiple Data

10Single instruction, multiple data

This urgency of coping up with an ever-increasing need for

computing resources in HEP experiments plays a central role

in this research project, which aims to:

– Give an overview of computing techniques used in HEP

experiments (with a focus on the LHCb experiment).

– Develop HEP algorithms, using traditional and machine

learning based techniques.

– Implement and benchmark these algorithms on multiple

computing architectures such as: CPUs, GPUs, and a new

kind of processor: the IPU11.

The report is organised as follows: section 2 aims to give an

overview of the physics at the LHCb. Section 3 describes

the computing architectures used in this research project, fol-

lowed by the description of the trigger system in section 4.

Section 5 outlines different stages of event reconstruction, fol-

lowed by section 6, which describes the algorithms developed

in this study and their performance. Finally, section 7 dis-

cusses the implications of these results, the limitations, fur-

ther work that can be carried out, and section 8 concludes this

study.

2 Physics at the LHCb

The LHCb experiment is designed to investigate decay

channels and oscillations of beauty and charm hadron systems

with a particular focus on CP violating phenomena, as well as

searches for anomalies in rare decays which indicate physics

beyond the standard model.

2.1 CP violation

The standard model of particle physics (SM) encapsulates

our current best understanding of three out of the four funda-

mental forces in the universe (weak, strong, electromagnetic,

excluding gravity). The SM classiﬁes elementary particles ac-

cording to their charges and describes their fundamental in-

teractions. In particle physics, the charge conjugation parity

symmetry states that the laws of physics must remain invari-

ant when a system’s spatial coordinates are ﬂipped (parity in-

version) eq(2) and the particle is replaced by its antiparticle

(charge conjugation) eq(3). This symmetry was ﬁrst observed

to be broken in 1964 in the decays of neutral kaons [16].

P(ˆx, ˆy, ˆz)→(−ˆx, −ˆy, −ˆz)(2)

C|Ψie−=|Ψie+(3)

The three generations of quarks in the SM naturally gener-

ate CP violating phenomena in both weak and strong inter-

actions. CP violation in weak interactions is described via

11Intelligence Processing Unit

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DEPARTMENTOFPHYSICSMAY2021FutureofcomputingattheLargeHadronColliderAuthor:DhananjaySaikumarAcademicSupervisors:DrDanielOHanlon&ProfJonasRademackerFORSUBMISSIONIN20212AbstractHighenergyphysics(HEP)experimentsattheLHCgeneratedataatarateofO(10)Terabitspersecond.Thisdatarateisexpectedtoexponentiallyinc...

展开>> 收起<<

DEPARTMENT OF PHYSICS MAY2021 Future of computing at the Large Hadron Collider_2.pdf

共26页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

DEPARTMENT OF PHYSICS MAY2021 Future of computing at the Large Hadron Collider_2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: