1. Accelerating Science
with OpenStack
The CERN User Story
Tim Bell
Tim.Bell@cern.ch
@noggin143
EMEA OpenStack Day, London
5th December 2012
2. What is CERN ?
• Conseil Européen pour la
Recherche Nucléaire – aka
European Laboratory for
Particle Physics
• Between Geneva and the
Jura mountains, straddling
the Swiss-French border
• Founded in 1954 with an
international treaty
• Our business is fundamental
physics , what is the
universe made of and how
does it work
OpenStack London December 2012 Tim Bell, CERN 2
3. Answering fundamental questions…
• How to explain particles have mass?
We have theories and accumulating experimental evidence.. Getting close…
• What is 96% of the universe made of ?
We can only see 4% of its estimated mass!
• Why isn’t there anti-matter
in the universe?
Nature should be symmetric…
• What was the state of matter just
after the « Big Bang » ?
Travelling back to the earliest instants of
the universe would help…
OpenStack London December 2012 Tim Bell, CERN 3
13. Tier-0 (CERN):
•Data recording
•Initial data reconstruction
•Data distribution
Tier-1 (11 centres):
•Permanent storage
•Re-processing
•Analysis
Tier-2 (~200 centres):
• Simulation
• End-user analysis
• Data is recorded at CERN and Tier-1s and analysed in the Worldwide LHC Computing Grid
• In a normal day, the grid provides 100,000 CPU days executing over 2 million jobs
OpenStack London December 2012 Tim Bell, CERN 13
14. • Data Centre by Numbers
– Hardware installation & retirement
• ~7,000 hardware movements/year; ~1,800 disk failures/year
Racks 828 Disks 64,109 Tape Drives 160
Servers 11,728 Raw disk capacity (TiB) 63,289 Tape Cartridges 45,000
Processors 15,694 Memory modules 56,014 Tape slots 56,000
Cores 64,238 Memory capacity (TiB) 158 Tape Capacity (TiB) 73,000
HEPSpec06 482,507 RAID controllers 3,749
High Speed Routers
24
Xeon Xeon Xeon Other Fujitsu (640 Mbps → 2.4 Tbps)
3GHz 5150 5160 Xeon 0% 3%
Xeon 4% 2% 10% E5335
Ethernet Switches 350
L5520 7% Xeon Hitachi
33% 23% 10 Gbps ports 2,000
E5345
14% HP Switching Capacity 4.8 Tbps
Seagate
0%
15%
1 Gbps ports 16,939
Maxtor
Western 0% 10 Gbps ports 558
Xeon
Xeon Digital
E5405
Xeon 59%
L5420 6% IT Power Consumption 2,456 KW
8% E5410
16%
Total Power Consumption 3,890 KW
OpenStack London December 2012 Tim Bell, CERN 14
16. Our Challenges - Data storage
• >20 years retention
• 6GB/s average
• 25GB/s peaks
• 35PB/year recorded
OpenStack London December 2012 Tim Bell, CERN 16
17. 45,000 tapes holding 80PB of physics data
OpenStack London December 2012 Tim Bell, CERN 17
18. New data centre to expand capacity
• Data centre in Geneva
at the limit of
electrical capacity at
3.5MW
• New centre chosen in
Budapest, Hungary
• Additional 2.7MW of
usable power
• Hands off facility
• Deploying from 2013
with 200Gbit/s
OpenStack London December 2012 Tim Bell, CERN
network to CERN 18
19. Time to change strategy
• Rationale
– Need to manage twice the servers as today
– No increase in staff numbers
– Tools becoming increasingly brittle and will not scale as-is
• Approach
– CERN is no longer a special case for compute
– Adopt an open source tool chain model
– Our engineers rapidly iterate
• Evaluate solutions in the problem domain
• Identify functional gaps and challenge them
• Select first choice but be prepared to change in future
– Contribute new function back to the community
OpenStack London December 2012 Tim Bell, CERN 19
20. Building Blocks
mcollective, yum Bamboo
Puppet
AIMS/PXE
Foreman JIRA
OpenStack
Nova
git
Koji, Mock
Yum repo
Active Directory / Pulp
LDAP
Lemon /
Hardware
Hadoop
database
Puppet-DB
OpenStack London December 2012 Tim Bell, CERN 20
21. Training and Support
• Buy the book rather than guru mentoring
• Follow the mailing lists to learn
• Newcomers are rapidly productive (and often know more than us)
• Community and Enterprise support means we’re not on our own
OpenStack London December 2012 Tim Bell, CERN 21
22. Staff Motivation
• Skills valuable outside of CERN when an engineer’s contracts
end
OpenStack London December 2012 Tim Bell, CERN 22
23. Prepare the move to the clouds
• Improve operational efficiency
– Machine ordering, reception and testing
– Hardware interventions with long running programs
– Multiple operating system demand
• Improve resource efficiency
– Exploit idle resources, especially waiting for disk and tape I/O
– Highly variable load such as interactive or build machines
• Enable cloud architectures
– Gradual migration to cloud interfaces and workflows
– Autoscaling and scheduling
• Improve responsiveness
– Self-Service with coffee break response time
OpenStack London December 2012 Tim Bell, CERN 23
24. Public Procurement Purchase Model
Step Time (Days) Elapsed (Days)
User expresses requirement 0
Market Survey prepared 15 15
Market Survey for possible vendors 30 45
Specifications prepared 15 60
Vendor responses 30 90
Test systems evaluated 30 120
Offers adjudicated 10 130
Finance committee 30 160
Hardware delivered 90 250
Burn in and acceptance 30 days typical 280
380 worst case
Total 280+ Days
OpenStack London December 2012 Tim Bell, CERN 24
25. Service Model
• Pets are given names like
pussinboots.cern.ch
• They are unique, lovingly hand raised
and cared for
• When they get ill, you nurse them back
to health
• Cattle are given numbers like
vm0042.cern.ch
• They are almost identical to other cattle
• When they get ill, you get another one
• Future application architectures should use Cattle but Pets with
strong configuration management are viable and still needed
OpenStack London December 2012 Tim Bell, CERN 25
26. Current Status of OpenStack at CERN
• Working with Essex OpenStack release
– Excellent experience with Fedora/RedHat team using EPEL packages
– Started Folsom test environment in November
• Focusing on the compute side to start with
– Nova compute with KVM and Hyper-V
– Keystone identity now integrated with Active Directory
– Replaced network layer with CERN code for our legacy network
management system
• Current pre-production installation
– 170 Hypervisors
– 2,700 VMs
– 3 DevOps part time running and enhancing the service
– Running production ‘cattle’ workloads for stress testing
OpenStack London December 2012 Tim Bell, CERN 26
28. When communities combine…
• OpenStack’s many components and options make
configuration complex out of the box
• Puppet forge module from PuppetLabs does our configuration
• The Foreman adds OpenStack provisioning for user kiosk to a
configured machine in 15 minutes
OpenStack London December 2012 Tim Bell, CERN 28
29. Foreman to manage Puppetized VM
OpenStack London December 2012 Tim Bell, CERN 29
30. Opportunistic Clouds in online experiment farms
• The CERN experiments have farms of 1000s of Linux servers
close to the detectors to filter the 1PByte/s down to 6GByte/s
to be recorded to tape
• When the accelerator is not running, these machines are
currently idle
– Accelerator has regular maintenance slots of several days
– Long Shutdown due from March 2013-November 2014
• One of the experiments have deployed OpenStack on their
farm
– Simulation (low I/O, high CPU)
– Analysis (high I/O, high CPU, high network)
OpenStack London December 2012 Tim Bell, CERN 30
31. Federated European Clouds
• Two significant European projects around Federated Clouds
– European Grid Initiative Federated Cloud as a federation of grid sites providing
IaaS
– HELiX Nebula European Union funded project to create a scientific cloud
based on commercial providers
EGI Federated Cloud Sites
CESGA CESNET INFN SARA
Cyfronet FZ Jülich SZTAKI IPHC
GRIF GRNET KTH Oxford
GWDG IGI TCD IN2P3
STFC
OpenStack London December 2012 Tim Bell, CERN 31
32. Ongoing Projects
• Production Preparation
– Following High Availability recommendations from the community
– Integrate monitoring with CERN frameworks
• Quota management with Boson
– CERN does not have infinite compute resources or budget
– Experiments are allocated a quota to split between their projects
– Collaborating with the community to develop a distributed quota
manager
• Block Storage Evaluation
– Investigating Gluster, NetApp and Ceph integration for EBS
functionality
OpenStack London December 2012 Tim Bell, CERN 32
33. Going further forward
• Deployment to production
– Planned for Q1 2013 based on Folsom
– No legacy tools in the second data centre
• Exploit new functionality
– Ceilometer for metering
– Bare metal for non-virtualised use cases such as high I/O servers
– X.509 user certificate authentication
– Load balancing as a service
– Cells for scalability
Ramping to 15,000 hypervisors with
100,000 to 300,000 VMs by 2015
OpenStack London December 2012 Tim Bell, CERN 33
34. Final Thoughts • A small project to share documents at
CERN in the ‘90s created the massive
phenomenon that is today’s world wide
web
• Open Source
• Vibrant community and eco-system
• Working with the Puppet and OpenStack
communities has shown the power of
collaboration
• We have built a toolchain in one
year with part time resources
• Sharing with other organisations to
achieve scale is the only
economically feasible path
• CERN contributes and benefits from
contributions of others
OpenStack London December 2012 Tim Bell, CERN 34
35. Questions ?
OpenStack London December
Tim Bell, CERN 35
2012
36. References
CERN http://public.web.cern.ch/public/
Scientific Linux http://www.scientificlinux.org/
Worldwide LHC Computing Grid http://wlcg.web.cern.ch/
Jobs http://cern.ch/jobs
Detailed Report on Agile Infrastructure http://cern.ch/go/N8wp
HELiX Nebula http://helix-nebula.eu/
EGI Cloud Taskforce https://wiki.egi.eu/wiki/Fedcloud-tf
OpenStack London December 2012 Tim Bell, CERN 36
39. CERN’s tools
• The world’s most powerful accelerator: LHC
– A 27 km long tunnel filled with high-tech instruments
– Equipped with thousands of superconducting magnets
– Accelerates particles to energies never before obtained
– Produces particle collisions creating microscopic “big bangs”
• Very large sophisticated detectors
– Four experiments each the size of a cathedral
– Hundred million measurement channels each
– Data acquisition systems treating Petabytes per second
• Top level computing to distribute and analyse the data
– A Computing Grid linking ~200 computer centres around the globe
– Sufficient computing power and storage to handle 25 Petabytes per
year, making them available to thousands of physicists for analysis
OpenStack London December 2012 Tim Bell, CERN 39
40. Our Infrastructure
• Hardware is generally based on commodity, white-box servers
– Open tendering process based on SpecInt/CHF, CHF/Watt and GB/CHF
– Compute nodes typically dual processor, 2GB per core
– Bulk storage on 24x2TB disk storage-in-a-box with a RAID card
• Vast majority of servers run Scientific Linux, developed by
Fermilab and CERN, based on Redhat Enterprise
– Focus is on stability in view of the number of centres on the WLCG
OpenStack London December 2012 Tim Bell, CERN 40
42. 500
1500
2000
2500
3000
3500
1000
0
Mar-10
Apr-10
May-10
Jun-10
Jul-10
Aug-10
Sep-10
Oct-10
Nov-10
OpenStack London December 2012
Dec-10
Jan-11
Feb-11
Mar-11
Apr-11
May-11
Jun-11
Jul-11
Tim Bell, CERN
Aug-11
Sep-11
Oct-11
Nov-11
Dec-11
Jan-12
Feb-12
Mar-12
Apr-12
May-12
Virtualisation on SCVMM/Hyper-V
Jun-12
Jul-12
Aug-12
Sep-12
42
Linux
Oct-12
Windows
43. Scaling up with Puppet and OpenStack
• Use LHC@Home based on BOINC for simulating magnetics
guiding particles around the LHC
• Naturally, there is a puppet module puppet-boinc
• 1000 VMs spun up to stress test the hypervisors with Puppet,
Foreman and OpenStack
OpenStack London December 2012 Tim Bell, CERN 43
44. Federated Cloud Commonalities
• Basic building blocks
– Each site gives an IaaS endpoint with an API and common security
policy
• OCCI? CDMI ? Libcloud ? Jclouds ?
– Image stores available across the sites
– Federated identity management based on X.509 certificates
– Consolidation of accounting information to validate pledges and usage
• Multiple cloud technologies
– OpenStack
– OpenNebula
– Proprietary
OpenStack London December 2012 Tim Bell, CERN 44
45. Supporting the Pets with OpenStack
• Network
– CERN specific component to interface to our legacy network
environment
• Configuration
– Using Puppet with Puppetlabs modules for rapid deployment
• External Block Storage
– Currently using nova-volume with Gluster backing store
• Live migration to maximise availability
– KVM live migration using Gluster
– KVM and Hyper-V block migration
OpenStack London December 2012 Tim Bell, CERN 45
46. Active Directory Integration
• CERN’s Active Directory
– Unified identity management across the site
– 44,000 users
– 29,000 groups
– 200 arrivals/departures per month
• Full integration with Active Directory via LDAP
– Uses the OpenLDAP backend with some particular configuration
settings
– Aim for minimal changes to Active Directory
– 7 patches submitted around hard coded values and additional filtering
• Now in use in our pre-production instance
– Map project roles (admins, members) to groups
– Documentation in the OpenStack wiki
OpenStack London December 2012 Tim Bell, CERN 46
47. Welcome Back Hyper-V!
• We currently use Hyper-V/System Centre for our server
consolidation activities
– But need to scale to 100x current installation size
• Choice of hypervisors should be tactical
– Performance
– Compatibility/Support with integration components
– Image migration from legacy environments
• CERN is working closely with the Hyper-V OpenStack team and
Microsoft to arrive at parity
– Puppet to configure hypervisors on Windows
– Most functions work well but further work on Console, Ceilometer, …
OpenStack London December 2012 Tim Bell, CERN 47
48. Active Directory Integration
• CERN’s Active Directory
– Unified identity management across the site
– 44,000 users
– 29,000 groups
– 200 arrivals/departures per month
• Full integration with Active Directory via LDAP
– Uses the OpenLDAP backend with some particular configuration
settings
– Aim for minimal changes to Active Directory
– 7 patches submitted around hard coded values and additional filtering
• Now in use in our pre-production instance
– Map project roles (admins, members) to groups
– Documentation in the OpenStack wiki
OpenStack London December 2012 Tim Bell, CERN 48
Established by an international treaty at the end of 2nd world war as a place where scientists could work together for fundamental researchNuclear is part of the name but our world is particle physics
Our current understanding of the universe is incomplete. A theory, called the Standard Model, proposes particles and forces, many of which have been experimentally observed. However, there are open questions- Why do some particles have mass and others not ? The Higgs Boson is a theory but we need experimental evidence.Our theory of forces does not explain how Gravity worksCosmologists can only find 4% of the matter in the universe, we have lost the other 96%We should have 50% matter, 50% anti-matter… why is there an asymmetry (although it is a good thing that there is since the two anhialiate each other) ?When we go back through time 13 billion years towards the big bang, we move back through planets, stars, atoms, protons/electrons towards a soup like quark gluon plasma. What were the properties of this?
Biggest international scientific collaboration in the world, over 11,000 scientistsfrom 100 countriesAnnual Budget around 1.1 billion USDFunding for CERN, the laboratory, itselfcomesfrom the 20 member states, in ratio to the grossdomesticproduct… other countries contribute to experimentsincludingsubstantial US contribution towards the LHC experiments
The LHC is CERN’s largest accelerator. A 17 mile ring 100 meters underground where two beams of particles are sent in opposite directions and collided at the 4 experiments, Atlas, CMS, LHCb and ALICE. Lake Geneva and the airport are visible in the top to give a scale.
The ring consists of two beam pipes, with a vacuum pressure 10 times lower than on the moon which contain the beams of protons accelerated to just below the speed of light. These go round 11,000 times per second being bent by the superconducting magnets cooled to 2K by liquid helium (-450F), colder than outer space. The beams themselves have a total energy similar to a high speed train so care needs to be taken to make sure they turn the corners correctly and don’t bump into the walls of the pipe.
- At 4 points around the ring, the beams are made to cross at points where detectors, the size of cathedrals and weighing up to 12,500 tonnes surround the pipe. These are like digital camera, but they take 100 mega pixel photos 40 million times a second. This produces up to 1 petabyte/s.
- Collisions can be visualised by the tracks left in the various parts of the detectors. With many collisions, the statistics allows particle identification such as mass and charge. This is a simple one…
To improve the statistics, we send round beams of multiple bunches, as they cross there are multiple collisions as 100 billion protons per bunch pass through each otherSoftware close by the detector and later offline in the computer centre then has to examine the tracks to understand the particles involved
To get Quark Gluon plasma, the material closest to the big bang, we also collide lead ions which is much more intensive… the temperatures reach 100,000 times that in the sun.
- We cannot record 1PB/s so there are hardware filters to remove uninteresting collisions such as those whose physics we understand already. The data is then sent to the CERN computer centre for recording via 10Gbit optical connections.
The Worldwide LHC Computing grid is used to record and analyse this data. The grid currently runs over 2 million jobs/day, less than 10% of the work is done at CERN. There is an agreed set of protocols for running jobs, data distribution and accounting between all the sites which co-operate in order to support the physicists across the globe.
So, to the Tier-0 computer centre at CERN… we are unusual in that we are public with our environment as there is no competitive advantage for us. We have thousands of visitors a year coming for tours and education and the computer center is a popular visit.The data centre has around 2.9MW of usable power looking after 12,000 servers.. In comparison, the accelerator uses 120MW, like a small town.With 64,000 disks, we have around 1,800 failing each year… this is much higher than the manufacturers’ MTBFs which is consistent with results from Google.Servers are mainly Intel processors, some AMD with dual core Xeon being the most common configuration.
Upstairs in the computer centre, a high roof was the fashion in the 1980s for mainframes but now is very difficult to get cooled efficiently
Our data storage system has to record and preserve 35PB/year with an expected lifetime of 20 years. Keeping the old data is required to get the maximum statistics for discoveries. At times, physicists will want to skim this data looking for new physics. Data rates are around 6GB/s average, with peaks of 25GB/s.
Tape robots from IBM and OracleAround 60,000 tape mounts / week so the robots are kept busyData copied every two years to keep up with the latest media densities
Asked member states for offers200Gbit/s links connecting the centresExpect to double computing capacity compared to today by 2015
Double the capacity, same manpowerNeed to rethink how to solve the problem… look at how others approach itWe had our own tools in 2002 and as they become more sophisticated, it was not possible to take advantage of other developments elsewhere without a major break.Doing this while doing their ‘day’ jobs so it re-enforces the approach of taking what we can from the community
Model based on Google Toolchain, Puppet is key for many operations. We’ve only had to write one new significant custom CERN software component which is in the certificate authority. Other parts such as Lemon for monitoring are from our previous implementation as we did not want to change all at once and they scale.
We’ve been very pleased with our choices. Along with the obvious benefits of the functionality, there are soft benefits from the community model.
Many staff at CERN are short term contracts… good benefits for those staff to leave with skills in need.
Standardise hardware … buy in bulk and pile it up then work out what to use it forMemory, motherboards, cables or disks interventionsUsers waiting for I/O means wasted cycles. Build machines at night unused during the day. Interactive machines mainly during the dayMove to cloud APIs … need to support them but also maintain our existing applicationsDetails later on reception and testing
The concept of pets and cattle came from Cloudscaling.Puppet applies well to the cattle model but we’re also using it to handle the pet cases that can’t yet move over due to software limitations. So, they get cloud provisioning but flexible configuration management.
Communities integrating … when a new option is being used at CERN in OpenStack, we contribute the changes back to the puppet forge such as certificate handling. Even looking at Hyper-V/Windows openstack configuration…
The project’s success comes down to community. A vibrant community has momentum of its own. As the WWW showed, many contributors can change how we see the world.Looking forward, as we help improve Puppet, remember that you will also be helping achieve a clearer understanding of the universe and how it works.
CERN is more than just the LHCCNGS neutrinos to Gran SassoCLOUD demonstrating impacts of cosmic rays on weather patternsAnti-hydrogen atoms contained for minutes in a magnetic vesselHowever, for those of you who have read Dan Brown’s Angels and Demons or seen the film, there are no maniacal monks with pounds of anti-matter running around the campus
We purchase on an annual cycle, replacing around ¼ of the servers. This purchasing is based on performance metrics such as cost per SpecInt or cost/GBGenerally, we are seeing dual core computer servers with Intel or AMD processors and bulk storage servers with 24 or 36 2TB disksThe operating system is Redhatlinux based distribution called Scientific Linux. We share the development and maintenance with Fermilab in Chicago. The choice of a Redhat based distribution comes from the need for stability across the grid, where keeping the 200 centres running compatible Linux distributions.
LHC@Home is not an instruction on how to build your own accelerator but a magnet simulation tool to test multiple passes around the ring. We wanted to use it as a stress test tool and in ½ day, it was running on 1000 VMs.
