1. WLCG Outlook Ian Bird, CERN GridPP Meeting 24 th September 2013 Accelerating Science and Innovation Accelerating Science and Innovation 24-Sep-2013 WLCG@GridPP311

2. 24-Sep-2013 2 A success story!

3. From the 2013 update to the European Strategy for Particle Physics g. Theory is a strong driver of particle physics and provides essential input to experiments, witness the major role played by theory in the recent discovery of the Higgs boson, from the foundations of the Standard Model to detailed calculations guiding the experimental searches. Europe should support a diverse, vibrant theoretical physics programme, ranging from abstract to applied topics, in close collaboration with experiments and extending to neighbouring fields such as astroparticle physics and cosmology. Such support should extend also to high- performance computing and software development. 24-Sep-2013 WLCG@GridPP313 i. The success of particle physics experiments, such as those required for the high-luminosity LHC, relies on innovative instrumentation, state-of-the- art infrastructures and large-scale data- intensive computing. Detector R&D programmes should be supported strongly at CERN, national institutes, laboratories and universities. Infrastructure and engineering capabilities for the R&D programme and construction of large detectors, as well as infrastructures for data analysis, data preservation and distributed data-intensive computing should be maintained and further developed. High Performance Computing

4. The Worldwide LHC Computing Grid WLCG: An International collaboration to distribute and analyse LHC data Integrates computer centres worldwide that provide computing and storage resource into a single infrastructure accessible by all LHC physicists WLCG: An International collaboration to distribute and analyse LHC data Integrates computer centres worldwide that provide computing and storage resource into a single infrastructure accessible by all LHC physicists WLCG@GridPP314

5. 5 200-400 MB/sec Data flow to permanent storage: 4-6 GB/sec ~ 4 GB/sec 1-2 GB/sec

6. Relies on – OPN, GEANT, US-LHCNet – NRENs & other national & international providers WLCG@GridPP316 LHC Networking

7. A lot more to come … WLCG@GridPP317 24-Sep-2013

8. Upgrade schedule Run1 Run2 Run3 ALICE + LHCb Run4 ATLAS + CMS CPU needs (per event) will grow with track multiplicity (pileup) and energy Storage needs are proportional to accumulated luminosity

9. Evolution of requirements 24-Sep-2013 WLCG@GridPP319 Estimated evolution of requirements 2015-2017 (NB. Not yet reviewed by LHCC or RRB) 2008-2013: Actual deployed capacity Line: extrapolation of 2008-2012 actual resources Curves: expected potential growth of technology with a constant budget (see next) CPU: 20% yearly growth Disk: 15% yearly growth

10. Technology outlook Effective yearly growth: CPU 20%, Disk 15%, Tape 15% Assumes: -75% budget additional capacity, 25% replacement -Other factors: infrastructure, network & increasing power costs WLCG@GridPP3110 24-Sep-2013

11. CPU Performance Exploiting new CPUs in scalable fashion requires changes to programming models, modifications to algorithms to use parallelism and reengineering of software. About 10 years ago processors hit power limits which brought to an end the era of "Moore's Law" scaling of the performance of single, sequential applications. Performance gap is developing between sequential applications and those tuned to utilize parallel capabilities of modern CPUs and continue to benefit from Moore’s law

12. Clock frequency Vectors Instruction Pipelining Instruction Level Parallelism (ILP) Hardware threading Multi-core Multi-socket Multi-node Running different jobs as we do now is still the best solution for High Throughput Computing (Grid/Cloud) } Gain in memory footprint and time-to-finish but not in throughput Very little gain to be expected and no action to be taken Micro-parallelism: gain in throughput and in time-to-finish 8 “dimensions of performance” SOFTWARE >>

13. HEP Software Challenge Must make more efficient use of modern cores, accelerators, etc -And better use of the memory Implies: -Multi-threading, parallelism at all levels, optimisation of libraries, redesign of data structures, etc All this requires significant re-engineering of frameworks, data structures, algorithms, … HEP must develop expertise in concurrent programming Requires investment of effort Initiative started: concurrency forum Strengthen this to a more formal HEP-software collaboration -Enable recognition for contributions -Clear plan – areas where people can contribute, etc

14. Grids: what did we achieve? And fail to achieve? Solved our problem of making effective use of distributed resources Made it work at huge scale Effective to ensure all collaborators have access to the data Networks are a significant resource Federation of trust and policies – important for future Cluster computing/grids not suitable/needed for many sciences Operational cost is high Very complex middleware was not (all) necessary Many tools were too HEP-specific 24-Sep-2013 WLCG@GridPP3114

15. Some lessons for HEP 24-Sep-2013 WLCG@GridPP3115

16. And the world has moved on Today we all use distributed computing services all the time -Dropbox, google drive, … -Streaming video, catch-up TV, … -Streaming music -Amazon, Google, Microsoft, etc web/cloud services – compute and storage -… 24-Sep-2013 WLCG@GridPP3116

17. Networks a problem? Global traffic within data centres is around 2000 EB/year -Global HEP traffic is ~2 EB/year; Global traffic between data centres is some 200 EB/year, -Global HEP traffic ~0.3 EB/year 24-Sep-2013 WLCG@GridPP3117 By 2015-16: global IP traffic will be ~1000 EB/year (75% video) - And we are 0.3… BUT, many areas where connectivity is a real problem

18. Industry involvement Clouds? 24-Sep-2013 WLCG@GridPP3118 Today’s grids – evolving technology Private clouds (federated?) Public clouds for science Public- Private partnerships Commercial clouds

19. Cloud characteristics 10 times more CPU cores @ 10 times fewer sites Reduces complexity Reduces management and maintenance cost Does not pretend to be a unified resource User has to select a particular zone to connect and stay in the same zone Data access across the zones is possible But not for free Data storage offers high availability At the expense of lower performance Provides means to communicate and move data asynchronously Does not prevent users to setup their own arbitrary infrastructure on top of the basic Cloud services

20. Evolution of today’s grids Grid sites are already deploying cloud software and using virtualisation -Many already offer cloud services Cloud software could replace parts of grid middleware -Even some encouragement to do this Huge support community compared to grid middleware -More sustainable support opportunities 24-Sep-2013 WLCG@GridPP3120

21. What would be needed? Open Source Cloud middleware OpenStack, CloudStack, OpenNebula… VM building tools and infrastructure CernVM+CernVM/FS, boxgrinder.. Common API EC2 is de facto standard but proprietary Common Authentication/Authorization Lots of experience with Grids High performance global data federation This is where we have a lot of experience HEP wide Content Delivery Network To support software distribution, conditions data Cloud Federation To unify access and provide cross Cloud scheduling

22. Evolution of the Grid Reduce operational effort so that WLCG Tiers can be self supporting (no need for external funds for operations) The experiments should be able to use pledged and opportunistic resources with ~zero configuration (Grid) clusters, clouds, HPC, … Implications: Must simplify the grid model (middleware) to as thin a layer as possible Make service management lightweight Centralize key services at a few large centers Make it look like a Cloud

23. Commercial clouds USA and Europe (and rest of world) are very different markets – and costs Outside of HEP, data often has intrinsic value (IP and/or commercial value) -E.g. genomics, satellite imagery, … -Concerns over data location, privacy, data access for many sciences -Several policy issues related to this European market is fragmented -No large (European) cloud providers 24-Sep-2013 WLCG@GridPP3123

24. Pricing… Costs are often higher than incremental costs of in-house clusters- some exceptions: Spot markets -Eg used by BNL to submit to Amazon “Backfill” -Use idle capacity for non-critical workloads – e.g. MC Also eventually may see other “value”: -Hosting data sets – get free CPU (because the data attracts other users) 24-Sep-2013 WLCG@GridPP3124

25. Scaling CERN Data Centre(s) to anticipated Physics needs WLCG@GridPP3125 CERN Data Centre dates back to the 70’s Upgraded in 2005 to support LHC (2.9 MW) Still optimizing the current facility (cooling automation, temperatures, infrastructure) Exploitation of 100 KW of remote facility down town Understanding costs, remote dynamic management, improve business continuity Exploitation of a remote Data centre in Hungary Max. 2.7 MW (N+1 redundancy) - Improve business continuity 100 Gbps connections Renovation of the “barn” for accommodating 450 KW of “critical” IT loads (increasing DC total to 3.5 MW) A second networking hub at CERN scheduled for 2014 24-Sep-2013

26. Connectivity (100 Gbps) WLCG@GridPP3126 24-Sep-2013

27. CERN CC – new infrastructure Replace (almost) entire toolchain Deploy as a private cloud Rationale -Support operations at scale Same staffing levels with new data centre capacity -HEP is not a special case for data centres -Improve IT efficiency, e.g. Use hardware before final allocation Small virtual machines onto large physical hardware Flexible migration between operating systems Run existing applications on top of the cloud -Enable cloud interfaces for physics Support new APIs, CLIs and workflows 24-Sep-2013 WLCG@GridPP3127

28. 24-Sep-2013 Bamboo Koji, Mock AIMS/PXE Foreman AIMS/PXE Foreman Yum repo Pulp Yum repo Pulp Puppet-DB mcollective, yum JIRA Lemon / Hadoop / LogStash / Kibana Lemon / Hadoop / LogStash / Kibana git OpenStack Nova OpenStack Nova Hardware database Puppet Active Directory / LDAP Active Directory / LDAP WLCG@GridPP3128

29. CERN Private Cloud Computing Resources on Demand -Ask for a server through a web page -Get the server in 2 to 15 minutes Flexible -Windows, Linux or roll-your-own -Various #cores, disk space options Amazon-like Infrastructure as a Service -Programmable through APIs 24-Sep-2013 WLCG@GridPP3129

30. Private Cloud Software 24-Sep-2013 WLCG@GridPP3130 We use OpenStack, an open source cloud project http://openstack.orghttp://openstack.org The same project is used for ATLAS and CMS High Level Trigger clouds HEP Clouds at BNL, IN2P3, NECTaR, FutureGrid, … Clouds at HP, IBM, Rackspace, eBay, PayPal, Yahoo!, Comcast, Bloomberg, Fidelity, NSA, CloudWatt, Numergy, Intel, Cisco …

31. Status Toolchain implemented in 18 months with enhancements and bug fixes submitted back to the community CERN IT cloud Hypervisors: 1300 + 100/week Cores: 24 000 + 1200/week Now in production in 3 OpenStack clouds (over 50,000 cores in total) in Geneva and Budapest managed by Puppet 24-Sep-2013 WLCG@GridPP3131

32. Initial Service Level Basic – like Amazon -Estimate 99.9% available (8 hours/year) -Each user has a 10 VM quota (Personal Project) -Experiments can request new projects and quotas from their pledges -You can upload your own images -Availability zones for load balancing services 24-Sep-2013 WLCG@GridPP3132

33. Production using Basic SLA Applications need to be ‘cloud enabled’ for production use (if need >99.9%) Use IT reliable backing stores such as -AFS, DataBase on Demand (MySQL), Oracle Use an automated configuration system -Puppet/Foreman -Contextualisation -CERNVMFS Backup if needed by the client (e.g. TSM) 24-Sep-2013 WLCG@GridPP3133

34. Coming … Deployment to new data centre in Budapest -Additional capacity and disaster recovery More flexibility and availability -Kerberos and X.509 support -E-groups for project members -Larger disk capacity VMs (like Amazon EBS) -Higher Availability VMs (CVI-like) -Other OpenStack functions as released Aim is 90% CERN IT capacity in the private cloud by 2015 -Around 15,000 hypervisors, 150,000 – 300,000 virtual machines 24-Sep-2013 WLCG@GridPP3134

35. 24-Sep-2013 WLCG@GridPP3135

36. What is needed Clear sustainable model (i.e. funding) essential to get buy-in of large research infrastructures currently in construction -FAIR, XFEL, ELIXIR, EPOS, ESS, SKA, ITER and upgrades to ILL and ESRF etc. Must support the needs of the whole research community, including the “long tail of science” Cannot be a one-size-fits-all solution Focus on solid set of reliable core services of general utility -But provide a way to share experience and knowledge (and higher level solutions The user community should have a strong voice in the governance of the e-Infrastructure Essential that European industry engage with the scientific community in building and providing such services 24-Sep-2013 WLCG@GridPP3136

37. What do we have already? Experience, lessons, or products from: Existing European e-infrastructure long-term projects -GEANT, EGI, PRACE Many “pathfinder” initiatives have prototyped aspects of what will be needed in the future -Includes much of the work in the existing e-Infrastructure projects but also projects such as EUDAT, Helix Nebula, OpenAIRE+, etc -Thematic projects such as WLCG, BioMedBridges/ CRISP/ DASISH/ ENVRI, as well as Transplant, VERCE, Genesi-DEC and many others 24-Sep-2013 WLCG@GridPP3137

38. What does an e-Infrastructure look like? Common platform with 3 integrated areas -International network, authorization & authentication, persistent digital identifiers -Small number of facilities to provide cloud and data services of general and widespread usage -Software services and tools to provide value-added abilities to the research communities, in a managed repository Address fragmentation of users (big science vs. long tail) -Make services attractive and relevant to individuals and communities Evolution must respond directly to user feedback and need 24-Sep-2013 WLCG@GridPP3138

39. An e-Infrastructure system 24-Sep-2013 WLCG@GridPP3139 Networks, Federated ID management, etc. Grid for comm unity CCS for comm unity Application software tools and services Cloud Resource(s) Data Archives HPC Facilities Collaborative tools and services Software investment Managed services – operated for research communities Individual science community operated services Key principles: Governed & driven by science/research communities Business model: Operations should be self-sustaining: -Managed services are paid by use (e.g. Cloud services, data archive services, …) -Community services operated by the community at their own cost using their own resources (e.g. grids, citizen cyberscience) Software support – open source, funded by collaborating developer institutions

40. Prototype public cloud for science “Centre of Excellence” CERN proposes a prototype to focus on data-centric services on which more sophisticated services can later be developed Use the resources installed by CERN at the Wigner Research Centre for Physics in Budapest, Hungary Accessible via federated identity (EDUGAIN): -Multi-tenant compute environment to provision/manage networks of VMs on-demand -‘dropbox’ style service for secure file sharing over the internet -Point-to-point reliable, automated file transfer service for bulk data transfers -Open access repository for publications and supporting data allowing users to create and control their own digital libraries (see www.zenodo.org) www.zenodo.org -Long-term archiving service -Integrated Digital Conferencing tools allowing users to manage their conferences, workshops and meetings -Oline training material for the services 24-Sep-2013 WLCG@GridPP3140

41. Prototype: Based on open source software: Openstack, owncloud, CERN storage services, FTS3, zenodo, Indico Services not offered commercially but run on a cost recovery basis All services will be free at the point of use -i.e. the end user does not have to pay to access the service All stakeholders participate in the funding model which will evolve over time CERN will: -Operate the services at the Wigner data centre -Not exert any ownership or IP rights over deposited material -Cover the operating costs during the first year -Make formal agreements with partners that wish to jointly develop/use the services -Negotiate/Procure services from commercial suppliers on-behalf of all partners 24-Sep-2013 WLCG@GridPP3141

42. Beyond the initial prototype Learn from the prototype to build similar structures around Europe -Not identical: each has its own portfolio of services and funding model -All interconnected: to offer a continuum of services -All integrated with public e-infrastructures: GEANT network (commercial networks are not excluded!) PRACE capability HPC centres EGI ? Determine whether this is: -Useful -Sustainable Understand the costs and determine what could be commercially provided 24-Sep-2013 WLCG@GridPP3142

43. Conclusions WLCG has successfully supported the first LHC run, at unprecedented scale -Will evolve to make the best use of technology and lessons learned HEP must make a major investment in software Proposal for a series of workshops to rethink the outdated HEP computing models – -10-year outlook will not possible to continue to do things in the “old” way For the future we see a need for basic e- infrastructures for science, that support community-specific needs -Propose a prototype of a few basic services to understand the utility of such a model 24-Sep-2013 WLCG@GridPP3143