Title: 2001%20Summer%20Student%20Lectures%20Computing%20at%20CERN%20Lecture%203%20
12001 Summer Student LecturesComputing at
CERNLecture 3 Looking ForwardsTony Cass
Tony.Cass_at_cern.ch
2Data and Computation for Physics Analysis
event filter (selection reconstruction)
detector
processed data
event summary data
raw data
batch physics analysis
event reconstruction
analysis objects (extracted by physics topic)
event simulation
interactive physics analysis
3LEP and LHC Parameters Compared
4Evolution of CERN Computing NeedsCPU Capacity
1997-2002
5Evolution of CERN Computing NeedsTape Storage
1995-2000
6Evolution of CERN Computing NeedsCPU Capacity
1997-2006
10'000'000
Infrastructure
9'000'000
8'000'000
Engineering
7'000'000
Others
6'000'000
LEP
5'000'000
NA48
CERN Units
4'000'000
NA45
3'000'000
2'000'000
COMPASS
1'000'000
LHC
0
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
year
7Evolution of CERNs CPU requirementsA different
view
Estimated CPU Capacity required at CERN
K SI95
5,000
Moores law some measure of the capacity
technology advances provide for a constant number
of processors or investment
4,000
LHC
3,000
2,000
Other experiments
1,000
0
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
Jan 20003.5K SI95
8Evolution of CERN Computing NeedsTape Storage
1995-2006
9HELP!
- The previous slides show that we have to cope
with a dramatic increase in computing capacity
before the start of LHC. - Can we afford it?
- How many boxes are needed (i.e. can we manage the
equipment)? - Fortunately, the price of computing equipment
falls each year. Will this help us?
10CPU and Disk Cost Predictions
CPU Costs
Disk Costs
11CPU and Disk Cost Predictions
CPU Costs
Disk Costs
year
12CPU and Disk Cost Predictions
CPU Costs
Disk Costs
13CPU and Disk Cost Predictions
CPU Costs
Disk Costs
14Disk Storage The Bad News
1996
2000
4GB 10MB/s
50GB 20MB/s
?
1TB
?
250x10MB/s2,500MB/s
20x20MB/s400MB/s
I/O
15Tape Storage Estimates
- Although CPU and disk costs are expected to
decrease dramatically, current estimates are that
the cost of tape storage and tape devices will
fall by less than a factor of 2 over the next 8
years. - These are not commodity items!
- Most tape use is for archive storage (write once,
read never), not HEP like usage. - Who backs up their home PC?
- Tape storage and tape devices are expected to
represent a significant fraction of the cost of
computing for the LHC experiments, particularly
for ALICE.
!
16Which system architecture for LHC?
Which of the different system architectures SMP
Scalable Distributedis appropriate for the LHC
experiments?
17Networks and CPU load
- High bandwidth commodity networks carry the
baggage of their low speed commodity origins. - The MTU1 for Gigabit Ethernet is still the 1.5KB
of Ethernet. - c.f. 64KB for HiPPI.
- processing packets takes timeand, with Gigabit
Ethernet, the packets come thick and fast.
1MTU MaximumTransmission Unit
18Can we build LHC Computing Farms?
- Probably or almost certainly, depending on your
level of optimism. - On the positive side
- CPU and disk price/performance trends suggest
that the raw processing and disk storage
capacities will be affordable, and - raw data rates and volumes look manageable
- perhaps not today for ALICE.
- But this does not mean it will be easy.
- Many, many boxes will be needed compared to
todays systems. - Building and managing coherent systems from such
large numbers of boxes will be a challenge.
1999 CDR _at_ 45MB/s for NA48!
2000 CDR _at_ 90MB/s for Alice!
19LHC Computing Worldwide
This picture, from the CMS CTP, shows how a
regional centre, here Fermilab, fits into the
computing environment between CERN and
universities. It is assumed here that high
bandwidth networks are available between CERN and
this, US based, regional centre. However, the
possibility of an Air Freight link for data
transfers is also indicated.
Although regional centres, in the US and
elsewhere, will certainly exist, we do not yet
know how best to make use of the facilities they
will offer. Can we link CERN and all the regional
centres into one global facility, usable from
everywhere? Or do the regional centres just
provide resources for their local clients?
20LHC Computing Worldwide - MONARC
- The MONARC Project has been set up to study these
issues. - Models Of Networked Analysis at Regional Centres
- More input on the practicalities of global
analysis is needed for - the Computing Progress Reports to be produced
this year by ATLAS and CMS and maybe other
experiments - Funding Agencies, especially in the US, and
- Planning!
21The Grid
- Over the past year, the Grid metaphor for
providing access to remote computing resource has
become popular. - Will the Grid bind Regional centres together?
- Studies are underway in Europe and the US.
22The Globus Toolkit
- Providing transparent access to different
computing resources requires an interface layer
which hides details of - batch systems (LSF, LoadLeveler, Condor),
- security and authentication
-
- The Globus Toolkit has been developed as just
such an interface layer and is being tested at
CERN and other HEP labs. - Theres still a long way to go, though!
23LHC Computing Grid
The LHC Computing Centre
24Authentication Kerberos vs PKI
- Kerberos is a popular authentication and access
control system. I prove I know something (my
password) and a central server gives me a ticket
to access resources. - I have a ticket, so I just need to type my
password once, - But a central server is needed at each site.
- In a Public Key system, I have a certificate
signed by some trusted body which I need to show
to prove who I am. - My certificate will be accepted by anybody who
trusts the organisation that signed my
certificate, - but I must protect it so you dont steal it and
use it instead! So I have to type a password or
passphrase whenever I need to use the
certificate.
25Software Concerns for LHC
- Software will throw LHC data away.
- software (human!) errors will lose data forever.
- Would you take this responsibility? Can you write
bug free code? - What would you do if you were managing the
worldwide effort? - Object Oriented techniques are todays industry
standard and LHC experiments must impose best
practice. - There are also secondary considerations
- widespread use of OO techniques outside HEP
implies widespread availability of support tools
and software, and - OO trained (ex) physicists will find more
employment opportunities.
26Software Concerns for LHC II
- Everything will change between now and 2005.
- The computing environment
- Unix vs NT.
- The programming language
- C vs Java.
- The in things
- OO vs ? Java vs ? what will computers look
like in 2005? - These all changed for LEP and those planning for
LHC must take this into account. - But maybe were being too worried. LEP was
planned at a time when IBM mainframes and DEC
minis looked invincible. The LEP experiments
still coped with change.
27Data and Computation for Physics Analysis
event simulation
Storage Solutions Zebra, Objectivity/DB, ROOT
event summary data
Simulation Packages GEANT3, GEANT4, FLUKA
raw data
analysis objects (extracted by physics topic)
batch physics analysis
event reconstruction
event filter (selection reconstruction)
interactive Physics analysis
Experiment frameworks provide interfaces to
storage and common services. HEP toolkits and
packages provided to meet common needs.
Analysis and visualisation packages HBOOK, PAW,
ROOT, Lizard, Iguana, JAS
Everything built using language standards, e.g.
STL
28OO Techniques and Data Storage/Management
- HEP has added many Data Storage and Management
systems on top of Fortran - e.g. Zebra for data structures, FATMEN for
event/file management - With the move to OO, can HEP use OO databases for
event storage and management? - Can it be done?
- Is it efficient?
- It seems the answer is yes. How do we really
switch to this model? - LHC software designers have to embrace this model
of working now and work to provide optimised
storage/processing environments.
29Why use an Object Database?
Raw data is reconstructed to produce ESD/AOD and
then interesting eventsare selected for further
study.
Hits
Tracker
In the traditional schemethis produces different
data sets and going back from a high to a low
level is difficult.
?dst/ntuple
Tracks
Raw Data
ESD
Event Header
AOD
Particles
Bookkeeping Database
AOD
Event Tags
With an object model and an object database it is
much easier to navigate between the different
levels of description of an event.
ESD
Event Header
Event Header
Raw
30Why use an Object Database?
- Hiding the details of the file storage is done by
the database manager. An RDBMS (e.g. Oracle) also
hides details of file storage, so why use an
ODBMS? - With an ODBMS, the underlying details of the I/O
are hidden. The program variables are the storage
variables, there is no need for explicit copying
by the programmer. - Physicists dont set out to select all tracks of
a given event. They might want to access some
tracks of an event, though. This sort of access
maps better onto an object database. - An ODBMS allows applications to suggest that
parts of an event should be stored close to each
otherrather than storing all tracks close
together. - But these dont seem to be general
requirementsthe ODBMS market has not taken off.
We need to be careful!
31Data Databases in 2001
- RDBMS vendors have been moving towards the ODBMS
market for some timeintroducing
Object-Relational DBMS. - Oracle 9i, with the recently announced C
interface, provides all the ODBMS features of the
previous slide. - You can now navigate between objects in the
database. - We are now actively testing the use of Oracle 9i
for physics data. Particular aspects being
investigated are - Scalability Storage overhead
- Mass Storage System Integration Data
Import/Export - Initial results are promising.
32Toolkits versus Frameworks
Toolkits Sets of generic procedures1 that can be invoked to perform related tasks. Do not constrain users (apart from parameter lists!). Can be provided by experiments but also by others, e.g. IT or 3rd parties.
Frameworks Systems to decide the order of execution, invoke procedures to do necessary work in determined order including user procedures. Constrain users to work within the overall architecture. Are experiment specific.
1 Note that the word procedure is used here in
a general sense. In terms of procedural
languages, procedures are subroutines and
functions. For an Object Oriented language, a
procedure is a class.
33Toolkit Design
- A toolkit should be
- generic so it can be used in more than one
framework - independent i.e. not forcing the use of other
toolkits - well defined with clear interfaces so it can be
replaced.
34Data Analysis Toolkits for LHC
- Just as for the data storage/management, HEP has
developed a specialised, Fortran based, analysis
environmentHBOOK, PAW and CERNLIB as a whole. - These needed to be rewritten/reinvented as HEP
moved to OO techniques. - Can we instead profit from commercial data
analysis tools? - OO based simulation packages are needed nowand
GEANT4 is becoming a reality. - The GEANT4 project, launched in 1994, is also a
demonstration of effective worldwide
collaboration on a major software project, - and there is much interest in GEANT4 beyond HEP.
35(LHC) Framework Design Choices
- From the user point of view, an experiment
computing framework ensures that they can write
code for a specific purpose (e.g. analysis or
detector reconstruction) without having to worry
about anything else - The framework ensures that
- objects and services they need are made
available, and - any objects they create will be stored if
required. - The three LHC frameworks are best distinguished
by the choices they have made in two areas. - Exposure of the persistency model for storage. Do
users work with transient or persistent objects?
Do users see the inheritance from the base
persistence class? - Procedure invocation. Do users themselves decide
the order of invocation of a set of procedures to
produce a given object? Or do they demand the
object and leave the framework to decide which
procedures must be invoked to produce it?
36 GAUDI (after the Catalan architect)
Converter
Converter
Application Manager
Converter
Transient Event Store
Data Files
Persistency Service
Message Service
Event Data Service
JobOptions Service
Algorithm
Algorithm
Algorithm
Data Files
Transient Detector Store
Persistency Service
Particle Prop. Service
Detec. Data Service
Other Services
Data Files
Transient Histogram Store
Persistency Service
Histogram Service
37CARF (CMS Analysis and Reconstruction Framework)
Application Framework
Physics modules
Reconstruction Algorithms
Event Filter
Physics Analysis
Data Monitoring
Calibration Objects
Event Objects
Visualization Objects
Utility Toolkit
ODBMS
C standard library Extension toolkit
Geant4
CLHEP
PAW Successor
LHC
38 AliROOT (Alice and ROOT)
Transport Engine selected at run time
Fast MC
Generators
FLUKA
Geant4
Geant3.21
Virtual MC
Geometry Database
39LHC Frameworks Another Comparison
- In Object Solutions, Booch says that there are
three basic types of object oriented
applications. -
- If they focus on they are
- direct visualization and manipulation of
the user-centric - objects that define a certain domain
- preserving the integrity of the persistent
objects data-centric - in a system
- the transformation of objects that are
computation-centric interesting to the system
- Using this categorisation, we could say that
- AliROOT is user-centric
- CARF is data-centric
- GAUDI is computation-centric
40Non-event data
- To make sense of an event, the raw detector data
is not enough. Non-event data is needed - for the overall geometry and structure of the
detector, including information about magnetic
fields, and - as they are not perfectly still, to understand
the real positions of the subdetectors at the
moment of the collision - to have the correct detector calibration at the
time of the collision as detector response also
changes (e.g. with temperature) and - about the run conditions of the accelerator, e.g.
beam energy, at the time of the collision. - All of these non-event data must also be stored
and managed.
41The Overall Picture
Globally, then, tags point to a collection of
events which are in a collection of runseach of
which has certain properties such as energy or
calibration constants.
How do these different collections fit together?
Or event data can be kept in one database with
non event data kept in a different
databaseeither object or relational.
42When are objects created?
- Once
- as part of some standard processing step (e.g.
reconstruction) run - for all (interesting) events in batch mode, or
- when needed for any individual event.
- Many times
- Once at least! See above
- But also as necessary if recomputing using local
data is faster than fetching the existing objects
from some remote system.
43Computer Supported Collaborative Working
- As we have seen, LHC collaborations are huge,
with people distributed around the globe. A
notable change from previous CERN experiments is
the significant contribution expected from US
institutes. Computers aid wide spread
collaboration in a number of ways.
44Video Conferencing
- There are two varieties of Video Conferencing.
- CODEC based video conferencing works well and is
much used commercially, but - it is expensive, and conferencies with 3 or more
sites require special equipment. - IP based video conferencing is cheapconnections
already existand many people can participate,
but - network links, especially those to the US, are
already overloaded, and we cant yet reserve
bandwidth for video conferencing. - An LHC project is trying to make the two systems
interoperate. - Use of Video Conferencing is growing and the LHC
collaborations will benefit more if network
bandwith increases - or if we can manage to reserve bandwidth solely
for conferences.
45Looking ForwardsSummary
- LHC demands for CPU and I/O capacity
significantly exceed those of the LEP
experiments. - Fortunately, experiments such as COMPASS have
intermediate requirements and allow us to study
the problems before LHC startup. - CPU cost trends suggest we can afford distributed
computing farms which provide adequate resources - but we have to start installing these in
2003/2004. - Software quality is a major concern for the LHC
experiments. - Object Oriented techniques are being adopted.
- This allows us to consider the use of Object
Oriented Databases for data management and other
commercial packages for analysis work.
46Computing at CERNConclusions
- Computing at CERN is interesting! Computing at
CERN is about Data! - Computing facilities at CERN are essential for
designing, building and operating both
accelerators and detectors. - Computers, of course, play a key role in the
reconstruction and analysis of the raw data
collected by experiments. - There are many interesting challenges as we look
forward to high data rate experiments in the next
couple of years and beyond to the LHC.