Final Presentation, Summer Internship Program, CERN

MARTINWILLSIMGRID Simulator

Martin Barisits Will Boyd

Supervised by Mario Lassnig and Vincent Garonne

August 13, 2009

Barisits, Boyd (Vienna UT, Georgia Tech) MARTINWILLSIM August 13, 2009 1 / 41

Content

1 Introduction

2 Approach

3 Topology Generator

4 Load Generator

5 Simulator

6 Simulation Results

7 Conclusion

8 Acknowledgements

9 References

Introduction

Introduction Authors

The Authors

Martin BarisitsVienna UT, Austria

• BSc: Medical ComputerScience

• MSc: ComputationalIntelligence

Will BoydGeorgia Tech, USA

• BSc: Physics & ComputerScience

Introduction Problem

The Problem

• Goal: Test data distribution strategies• Need: Simulator• Need: Ability to load the Simulator with the current GRID Topology• Need: Inject the Simulator with realistic workloads• Process Results

Approach

Approach Design

• Two basic challenges• Write a tool to get a snapshot of the whole GRID environment

(Topology, Usage) and to generate Loads• Write a Simulator which can execute this input

• Different Simulators for GRID analysis are available in theresearch community

• For time reasons we decided to use a Simulator package to baseour Simulator on

Approach Design

Approach Package Evaluation

Package Evaluation

• Evaluation of GRID/cloud computing simulation packages• SimGrid[2]

• Based on pure C

• Pros: Fast execution time; low memory consumption; scalable

• Cons: Lacking in some functionality; High level of abstraction• GridSim[3]

• Java-based

• Pros: Highly developed; internal logging of network traffic; easier touse; Packet-based

• Cons: Slow execution time; bad memory consumption; not scalable

Package Evaluation

• Based on pure C

• Java-based

Package Evaluation

• Based on pure C

• Java-based

Package Performance

• Attempted to simulate oneday on GRID (1.5 millionfile transfers)

• GridSim: exponential inCPU time with increasingtransfers

• SimGrid: linear in CPUTime with increasingtransfers

Approach Flow

Topology Generator

Topology Generator The GRID

GRID Sites

GRID sites across the world

ATLAS Computing Model

• Hierarchical computingnetwork

• Tier-0

• Tier-1

• Tier-2

• Tier-0 (CERN)generates data

• Tier-1s store data

• Tier-2s process data The Tier-0 and Tier-2network configuration[1]

• Tier-0

• Tier-1

• Tier-2

• Tier-0

• Tier-1

• Tier-2

• Tier-0

• Tier-1

• Tier-2

Topology Generator Simulator Topology

The Topology Generator

• TopologyGen.py• Script to construct GRID topology

• Parses TiersOfATLASCache.py

• Finds and associates Tier-1s and Tier-2s

• Queries the DQ2 database

• Total disk space capacity

• Used disk space

• Topology is written to two XML files

• Used disk space

Platform and Deployment Files

• Platform file• Node declarations

• Link declarations

• Route declarations

• Deployment file• Logfiles for each node

• Total and used disk space

• Used disk space by datatype

• Tier-0 loadfiles

• Associated Tier-1s and Tier-2s

Route declaration from a Tier-0 to Tier-1 in the platform file<route src= ’CERN_52 ’ ds t= ’RAL−LCG2_MCDISK ’>

< l i n k : c t n i d = ’RAL−LCG2_MCDisk_InternalLink ’>< l i n k : c t n i d = ’ RAL_OPNLinkInternal ’ / >< / l i n k : c t n i d = ’ CERN_52_InternalLink ’>

< / rou te>

Host definition in the platform file<process f u n c t i o n = ’ T ier1Storage ’ host= ’ INFN−T1_DATADISK ’>

< / process>

MARTINWILLSIM GRID Topology

The topology that is generated for simulation

Load Generator

Load Generator LoadGen.py

Generating a Load

• Loadfile given to each Tier-0

• Loadfiles define dataset transfers• Unique dataset ID

• Random (uniform) target Tier-1 storage node

• Random (uniform) filesize (0.5-6GB)

• Random (weekly distribution) inter-submission time

• Dataset datatype (i.e., RAW)

Load Generator LoadGen.py

Generating a Load

• Loadfile given to each Tier-0

• Loadfiles define dataset transfers• Unique dataset ID

• Random (uniform) target Tier-1 storage node

• Random (uniform) filesize (0.5-6GB)

• Random (weekly distribution) inter-submission time

• Dataset datatype (i.e., RAW)

Load Generator Simulating Real Loads

Load Distribution

• Dataset distribution• Uniform background

traffic

• Wednesday/Fridaypeak traffic

• Random "spikes" oftraffic

• Each component isweighted

• Distribution can easilybe adjusted

An example weekly dataset transfer distribution

Load Distribution

traffic

Load Distribution

traffic

Load Distribution

traffic

Load Distribution

traffic

Load Distribution

traffic

Simulator

Simulator Facts

• Based on SimGrid[2]• Implemented in C• Intent to implement an extensible Simulation-Framework rather

than a strict Simulator• Goals:

• Fast• Scalable• Representative

Simulator Facts

Features

• Build network topology according to the injected Topology File• Simulate the shipment and processing of DataSets• Give the user a framework to implement/change own behavior• Provide functions to write simulation output• Background Noise Generation (Traffic from other Experiments,

. . . )

Simulator Facts

Features

. . . )

Simulator Facts

Features

. . . )

Simulator Facts

Features

. . . )

Simulator Facts

Features

. . . )

Simulator Design

Major Entities

• Nodes (Tier0, Tier1, Tier2)• Tasks (Datatransfer)• DataSets• Links

Simulator Design

Major Entities

• Nodes (Tier0, Tier1, Tier2)• Tasks (Datatransfer)• DataSets• Links

Simulator Design

TaskTaskname Size (DataSet)

• Taskname• Command

• Size• Communication

Size• Execution Size

• DataSet• DataSet ID• DataSet Size• DataSet Type

Simulator Design

Command Language

• (PUSH, DataSet)• (PULL, DataSetTemplate)• (DELETE, DataSetTemplate)• (PROCESS, DataSet)• (NOISE)• (INITSHUTDOWN)• (FINALIZE)

Simulator Design

Command Language

Simulator Design

Command Language

Simulator Design

Command Language

Simulator Design

Command Language

Simulator Design

Command Language

Simulator Design

• Different types of nodes with different functions• Producer (Tier 0)• Storages (Tier 1, Tier 2)• Hosts (Tier 1, Tier 2)• FinalizeNode

• All nodes understand the Command Language• Different nodes execute commands differently

• It’s up to the user to define the semantics of a node• Node Features

• DataSet Store (Hashmap)• Ability to write simulation output• Queues• Simulation Functions (Execute, Sleep, . . . )

Simulator Design

Pseudo Code of a Tier-1 Storagewhile ( 1 ) {

task = receiveTask ( ) ;switch ( task )

case <PUSH, dataSet >:s to re ( dataSet ) ; / / S t o r e i n t h e F i l e S y s t e m

t i e r 2 = getNextT ier2 ( ) ;send ( t i e r 2 , PUSH, dataSet ) ; / / S e n d t h e t a s k

w r i t e S t a t i s t i c s ( ) ;case <DELETE, dataSet >:

. . .case <FINALIZE , dataSet >:

w r i t e S t a t i s t i c s ( ) ;break ;

. . .}

Simulator Design

The way of a DataSet (1/4)

Simulator Design

Simulation Results

Simulation Results Disk Space Evolution

Overloading the GRID

Tier-0 Dataset submissiondistribution

Disk space evolution withincreasing daily dataset transfers

Disk Space Evolution

Tier-1 storage node An associated Tier-2 storagenode

Data Storage by Datatype

Uniform dataset transfer distribution Simulated dataset transfer distribution

Simulation Results Scalability

Scalability of MARTINWILLSIM

• MartinWillSim run tosimulate increasingnumber of days

• 250,000 tasks/day(800TB/day)

• Simulated one month ofdataset transfers in 40mins.

• CPU time linear withnumber of simulated tasks

Conclusion

Conclusion Conclusion

• Evaluation of different Simulator Packages [2 Weeks]• Design and Implementation of:

• Topology & Load Generator [6 Weeks]• MartinWillSim Simulator [6 Weeks]

• Result Analysis• Documentation

Conclusion Future Work

Future Work

• Add LISP hooks for Decision Making• Add more detail of the ATLAS Computing model to the simulator• Add functions for random errors (node failures, link failures)• Add recording of other statistics for result analysis

• Link Throughput• Usage of Processing Capacities• Replication Factors• User behavior

• For higher detail: Add Files as smallest Entity to the Simulator• Further Validation

Future Work

Acknowledgements

Thank you!

• Mario Lassnig

• Vincent Garonne

• Angelos Molfetas

• Ingrid Schmid

• All those who helped make the 2009 Summer Student Programmepossible!

References

[1] "https://twiki.cern.ch/twiki/pub/LHCOPN/ImplementationDetails/map-lhcopn.png"

[2] Casanova, Legrand, Quinson: SimGrid: a Generic Frameworkfor Large-Scale Distributed Experiments, 2008

[3] Buyya, Murshed: GridSim: A Toolkit for the Modeling andSimulation of Distributed Resource Management and Schedulingfor Grid Computing, 2002

Final Presentation, Summer Internship Program, CERN

Education

Transcript of Final Presentation, Summer Internship Program, CERN

Arlington/Roe Summer Internship Program 2018 Summer ... · Arlington/Roe Summer Internship Program 2015 L to 2018 Summer Internship Program The Arlington/Roe internship gave me a

CERN summer school 2013

SUMMER INTERNSHIP PROJECT REPORT Cover page …kudumbashree.org/storage/files/gbom8_aneez s - summer internship... · summer internship report submitted by ... college of engineering

Pnb Summer Internship

Biomedical Informatics Summer Internship Program · PDF fileBiomedical Informatics Summer Internship Program ... Java, or ability to learn ... Biomedical Informatics Summer Internship

CERN openlab Overview CERN openlab Summer Students 2015 Fons Rademakers.

Summer @CERN

Summer Internship

Entrepreneurship@Cornell Summer Internship …eship.cornell.edu/downloads/Internship-Employer-Handbook.pdfEntrepreneurship@Cornell Summer Internship Program Employer Handbook This

CERN summer presentation

Helge Meinhard CERN CERN Openlab Summer Student Lectures 2014

CERN Summer Student

Summer Internship Brochure

ITC SUMMER INTERNSHIP

Summer Internship Report

CERN Summer Study 2011

CERN Summer Research Experience for Undergraduatesum-cern-semester-abroad.org/CERN_Brochure.pdf · The University of Michigan CERN Summer Research Experience for Undergraduates (REU)

Summer Internship Project