ProductionQuality Grid Environments with UNICORE

1
Production-Quality Grid Environments with UNICORE

• Sep 23, 2005
• Sogang University
• Distributed Computing Communication Laboratory
• Eunseok, Kim

2
Outline

• What is UNICORE?
• Usage of UNICORE
• How does it provide production-quality grid
  service?
• Conclusion

3
What is UNICORE?

• A project between 1997 and 1999
• Funded by BMBF the German Ministry for
  Education and Research
• The goal of UNICORE Create
• Seamless, Secure, Intuitive
• access to distributed German HPC resources

4
Core Functions

• Security
• Certificate/PKI Based Security Model
• Interoperable with Globus Certificates
• Support for stronger trust model than Globus
• or Globus Trust Model using proxies.
• Workflow
• Local, remote, and nested task graphs
• File management, transfer, and streaming
• If and loop constructs
• Flow control based on task status, time events,
  and file-state events
• GRID Management
• Hard/soft fail recovery.
• System administrators control interface.
• Extensive logging support
• Built in certificate management in the GUI client.

5
Core Functions

• Resource Broker
• Multi-site resource check prior to submit.
• Estimate of time until execution.
• Ticket generation and checking.
• Dynamic brokering at execution time.
• Interactive Access Extension
• Allows standard terminal style interaction.
• Unicore Single Sign-on
• Complete multi-site authentication
  authorization.
• Includes Interactive Batch.

6
Architecture
7
Distributed Job
8
User tier

• Abstract Job Object (AJO)
• Realization of UNICOREs job model
• Contains
• platform and site independent descriptions of
  computational and data-related tasks
• Resource information and workflow specifications
  along with user and security information
• Sent to the UNICORE gateway
• In form of serialized and signed Java objects

9
Server tier

• Gateway
• Controls the access to a Usite
• Acts as the secure entry point
• Usite identifies a participating organization
  which is resolved by gateway URL
• Vsite identifies a particular set of resources at
  a Usite
• Controlled by a NJS (Network Job Supervisor)
• UNICORE User DB (UUDB)
• Maps user certificates to login

10
Server tier

• Network Job Supervisor (NJS)
• Virtualizes underlying resources
• By mapping AJO on a specific target system
• Called incarnation
• Using Incarnation DB (IDB)
• Processes workflow descriptions
• Perform pre- and post-staging of files
• Authorizes the user via UNICORE User DB (UUDB)
• Resource Broker
• Support for simple QoS bids
• Turnaround time and price
• Ticket based
• Sites retain control of ticket validity.
• Ticket checking at execution time
• Advanced reservation possible (if site supported)

11
Target tier

• Target System Interface (TSI)
• Implements the interface to the local resource
  management system
• Perl interface to host platform
• Ex) PBS, GRAM etc

12
More detailed Architecture
13
What Unicore Can Do (summary)

• Run control remote
• Shell scripts
• Applications (without modifications)
• Files
• Transfer client lt-gt servers
• Transfer servers lt-gt servers
• Remote management
• Simple VOs
• Resource
• Discovery
• Brokering
• Reservation
• Provide site autonomy
• Data archive access
• Integration with local admin.

• Workflow
• Nested remote jobs
• Nested local jobs
• Loops Conditionals
• Time based events
• File state events
• Rerun on failure
• Seamless Computing
• Filer streaming
• Application to application
• Application to Client
• CORBA Integration
• Built in cert. management
• Grid administration
• Tools
• Failsafe restart
• Dynamic hosting

14
Interfaces

• Client Plugins
• Application specific support (CPMD, Gausian,
  NASTRAN, ...)
• Unicore Protocol Layer
• Resource discovery, job submission, job
  management
• Incarnation Data Base and TSI Interface
• Incarnation support
• Unicore User Data Base
• User mapping and authorization
• Broker/Scheduler Interface
• Interface defined and implemented in NJS.
• Independent File Transfer Interface
• Interface defined

15
The Unicore Client
16
WorkFlow
Do Construct
If Construct
Complex Dependencies
17
Plugins
CPMD Plugin Wizard
18
DEISA

• Current
• Four partners are connected at 1 Gbit/s
• Consists of over 4000 IBM Power4 and 416 SGI
• Peak performance - 22 TFlops
• LoadLeveler
• Goal (by 2006)
• 11sites connected at 10 Gbit/s
• Peak performance - 140 TFlops
• Metascheduler

19
DEISA configuration
20
Lessons Learned

• Taught by end-users and software developers from
• Industry pharmaceutical, petrochemical,
  bio-molecular, weather prediction, automotive,
  engineering
• Research astrophysics, quantum physics, material
  science, medicine, biology, chemistry
• Deployment of new production software has to
  offer added value
• Ease usage, increase effectiveness, decrease
  cost,
• Users have to be stimulated and encourage to
• use Grid technology for applications,
  computations, data transfer and access to
  resources
• adapt/integrate their applications to/into Grids
• Operation of production environments is costly
• Certification authority, administrative tools,
  integration into site management, licenses,
• Common production environment difficult to
  maintain

21
More Lessons Learned

• Fulfillment of functional requirements is not
  enough
• Users want
• software of high quality, especially high
  reliability and resilience
• help to overcome initial hurdles like
• obtaining certificates, adapt applications, ...
• 24/7 availability of the Grid infrastructure
• 24/7 availability of the Grid experts
• support hotline, help desk, mailing lists,
• long-term commitment for continuous development
  and support
• workshops, hands-on training, ...

22
Conclusion

• Production Grids are possible
• Easy/unified usage, integration of legacy
  applications, cost reduction,
• When will we see the WS- impact?
• Users demand a fully-fledged product
• Functions, but also support, support, support
• Continuity is crucial
• Open Source distribution is the right way
• Source for bug reports, requirements,
• Higher visibility community building