Copyright © 2010 Platform Computing Corporation. All Rights Reserved.1
The CERN Cloud Computing ProjectWilliam Lu, Ph.D.Platform Computing
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.2
LHC Computing Hierarchy
Markus Schulz, CERN 2Emerging Vision: A Richly Structured, Global Dynamic System
Tier 0 +1
Tier 1
Tier2 Center
Online System
CERN Center PBs of Disk; Tape Robot
FNAL CenterIN2P3 Center INFN Center RAL Center
InstituteInstituteInstituteInstitute
Workstations
~100-1500 MBytes/sec
2.5-10 Gbps
Tens of Petabytes by 2010.An Exabyte ~5-7 Years later.
~PByte/sec
10 Gbps
Tier2 CenterTier2 CenterTier2 Center
~2.5-10 GbpsTier 3
Tier 4
Tier2 Center Tier 2
Experiment
CERN/Outside Resource Ratio ~1:2Tier0/( Tier1)/( Tier2) ~1:1:1
0.1 to 10 GbpsPhysics data cache
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.3
Computers:• 40,000 CPU cores used by multiple experiments
Storage: • Disks + tapes• Storage management system (CASTOR) is tightly integrated
with workload management (Platform LSF)
Software:• Apps: Open source, home grown, • OS: Scientific Linux, other Linux• VMs: open source XEN, KVM
Environment
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.4
IT serves users manually• User requests of resource, OS, software stack etc. are
handled manually, which is slow
Users circumvent scheduling policies• Users are not satisfied with the centralized management
scheduling policies due to their unique needs• They submit a pilot job to occupy resources then run scripts
to prepare the application environment and schedule jobs within the resource block. This causes low resource utilization
Legacy application issues• Legacy applications need legacy OS, which does not run on
the latest hardware
Challenges
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.5
Batch VirtualizationRequirements
How too Insolate application
environmento Increased security
How to o Automate resource
provisioning and management
o Scalable management practice
Virtualization
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.6
Platform ISF + Platform ISF Adaptive Cluster Integration with Platform LSF to provision VMs based on
workload Integration with provisioning
system Quattor Each experiment is able to
schedule their own VMclusters with uniqueapplication environment
VM cluster capacity is elastic based on workload
Solution
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.7
How It Works?
Platform ISF
Shared pool of resources
HPC administrator sets up VM resource pools,
one for each experiment
1Platform LSF Platform LSF
User submits a workload that cannot be met by his
VM resource pool
3
Platform ISF AC interacts with Platform
ISF to adjust the size of the resource pool
4
External Provider
Platform LSF
Platform ISF AC Platform ISF AC
Platform ISF AC
HPC administrator also sets up minimum and
maximum number of VMs within each pool
2
Copyright © 2010 Platform Computing Corporation. All Rights Reserved.8
Increase user service level• Each experiment can control their own application stack and
resource allocation policies
Redeploy servers quickly and efficiently
Results
• Reduce cost and save power• Shares batch compute servers
with data management and database servers
Automated administration• Allow scalability
No hypervisor lock-in• Freedom of choosing multiple
VM hypervisors