State of the Grid 2002

1
State of the Grid 2002
GGF, July 2002

• Dr. Francine Berman
• Director, NPACI and SDSC
• Professor, Department of Computer Science and
  Engineering, UCSD

2
Grid Computing in the News
3
Has the Grid been oversold?
4
State of the Grid 2002

• How did we get here?
• Short history
• Trends
• Grids Today -- Where are we now?
• Drivers
• Applications
• Technology
• Has the Grid been oversold?
• The Grid the next 10 years
• New application paradigms
• New devices
• Policy and social dynamics
• New research

5
A short history of the Grid in 2 slides

• Science as a team sport
• Grand Challenge Problems of the 80s
• Parallel computation
• First serious study of program coordination
• Gigabit Testbed program
• Focus on applications for the local to wide area
• I-Way at SC 95
• First large-scale grid experiment
• Provided the basis for modern grid infrastructure
  efforts

6
1995 2000 Maturation of Grid Computing

• Grid book gave a comprehensive view of the
  state of the art
• Important infrastructure and middleware efforts
  initiated
• Globus, Legion, Condor, NWS, SRB, NetSolve,
  AppLes, etc.
• 2000 Beginnings of a Global Grid
• Evolution of the Global Grid Forum
• Some projects evolving to de facto standards
  (e.g. Globus, Condor, NWS)

7
Current Trends 1

• Proliferation of resources
• Everyone has computers
• Multiple IP addresses per person
• Increasing Application Complexity
• Multi-scale
• Multi-disciplinary
• Immense amounts of data

Arpanet1969
Internet2002
8
Current Trends 2

• Coordination/collaboration is default mode of
  interaction
• The Internet
• Globalization, virtualization
• Open source movement
• At scale, heterogeneity is a fact of life

9
Grid Computing Today
DISCOM SinRG APGrid IPG
10
It Can Be Done Real World Distributed
Applications

• Walmart Inventory Control
• Satellite technology used to track every item
• Bar code information sent to remote data centers
  to update inventory database and cash flow
  estimates
• Satellite networking used to coordinate vast
  operations
• Inventory adjusted in real time to avoid
  shortages and predictdemand
• Data management,prediction, real-time,wide-area
  synchronization

11
Real World Distributed Applications

• Everquest
• 45 communal world servers (26 high-end PCs per
  server) supporting 430,000 players
• Real-time interaction, individualized database
  management, back channel communication between
  players
• Data management adapted to span both client PC
  and server to mitigate communication delays
• Game masters interact with players for real-time
  game management

12
Real World Distributed Applications

• SETI_at_home
• 3.8M users in 226 countries
• 1200 CPU years/day
• 38 TF sustained (Japanese Earth Simulator is 40
  TF peak)
• 1.7 ZETTAflop over last 3 years (1021, beyond
  peta and exa )
• Highly heterogeneous gt77 different processor
  types

13
From Distributed Applications to Grid
applications

• Real world applications demonstrate that it is
  technically, commercially, and economically
  viable to deploy robust, large-scale distributed
  applications
• The Grid should accelerate progress
• Current applications currently developed as
  stand-alone entities
• Availability of Grid services should allow
  designers to build on existing infrastructure
  and evolving technologies
• Applications developed for the Grid will likely
  contribute to community infrastructure,
  standards, progress
• Stability/performance of multiple applications
  currently not addressed
• Grid applications will help evolve policies for a
  scalable Grid

14
Unifying the InfrastructureA Community Grid
Model

• Roll your own SW but agree on interfaces, service
  architecture, standards

NPACI, TeraGrid Grid Applications
Grid Applications
NPACI Grid Middleware
User-focused and targeted grid middleware,
tools, and services
Common Infrastructure layer (NMI, GGF standards,
OGSA etc.)
Grid Resources
15
Common Grid Application Paradigms

• Minimal Communication applications
• Includes embarrassingly parallel apps, parameter
  sweeps
• Staged/linked applications (do part A then do
  part B)
• Includes remote instrument applications (get
  input from instrument at site A, compute/analyze
  data at site B)
• Access to resources (get stuff from/do something
  at site A)
• Portals, access mechanisms

16
Minimal Communication Applications

• MCell -- Simulation of neuromuscular synaptic
  transmission
• Uses Monte Carlo diffusion and chemical
  reaction algorithm in 3D to simulate complex
  biochemical interactions of molecules
• Molecular environment represented as 3D space
  in which trajectories of ligands against cell
  membranes tracked
• Ultimate Goal A complete molecular model of
  neuro-transmission at level of entire cell

MCell Animation
17
Grid Software Provides the Distribution Mechanism

• MCell is a parameter sweep application
• Parameter Sweeps class of applications that are
  structured as multiple instances of an
  experiment (task) with distinct parameter sets
• APST middleware developed to schedule/deploy Grid
  parameter sweeps and promote application
  performance

SharedInput files
CLIENT
APSTMiddleware
experiments
Shared output files
MCell Software agent
Grid of clusters and MPPs
18
Linking Applications

• Telescience -- Derivation 3D information about a
  sample from a series of 2D projections
• Links computation and data management to
  unique, expensive instrumentation
• Requires advanced visualization tools for
  segmentation and analysis of the data
• Provides critical database of biological
  structure info for neuroscientists

3D Model of the Node of Ranvier
19
Grid Software used to coordinate resources
PORTALS
SECURITY, RESOURCE MANAGEMENT, DATA
MANAGEMENT SERVICES
20
Access to Resources

• NASA Information Power Grid
• Computing resources ?800 CPU nodes in half a
  dozen SGI Origin 2000s and several workstation
  clusters at Ames, Glenn, and Langley, ?200 nodes
  in a Condor pool
• Storage resources 50-100 Terabytes of archival
  information/data storage uniformly and securely
  accessible from all IPG systems via MCAT/SRB and
  GSIftp / Gridftp
• Globus provides Grid common services
• Stable and supported operational environment
• (help desk, consulting, training)

21
Grid portals used to provide resource access and
information

• Grid portals provide
• submission, tracking, and management of jobs
  running on IPG resources
• integration of IPG services with the user desktop
  environment,
• access to persistent user profiles.

22
Community Grid Model
23
Resources and Infrastructure Driver

• TeraGrid will provide in aggregate
• 13.6 trillion calculations per second
• Over 600 trillion bytes of immediately accessible
  data
• 40 gigabit per second network speed
• TeraGrid will provide a new paradigm for
  data-oriented computing
• Critical for disaster response, genomics,
  environmental modeling,

24
TeraGrid
574p IA-32 Chiba City
256p HP X-Class
128p Origin
128p HP V2500
HR Display VR Facilities
Caltech Data collection and analysis applications
92p IA-32
HPSS
HPSS
ANL Visualization
SDSC Data-orientedcomputing
Myrinet
UniTree
HPSS
Myrinet
1024p IA-32 320p IA-64
1176p IBM SP Blue Horizon
1500p Origin
Sun E10K
NCSA Compute-Intensive
25
TeraGrid Common Infrastructure Environment

• Linux Operating
• Environment
• Basic and Core Globus
• Services
• GSI (Grid Security Infrastructure)
• GSI-enabled SSH and GSIFTP
• GRAM (Grid Resource Allocation Management)
• GridFTP
• Information Service
• Distributed accounting
• MPICH-G2
• Science Portals

• Advanced and Data Services
• Replica Management Tools
• GRAM-2 (GRAM extensions)
• CAS (Community Authorization Service)
• Condor-G (as brokering super scheduler)
• SDSC SRB (Storage Resource Broker)
• APST user middleware, etc.

26
Measures of Success

• Use a single node on TeraGrid
• Portals, SW, scheduling should allow access to
  designated individual resources
• Use as a wide-area cluster computer
• Use multiple designated resources of the same
  type for a single computation
• Use as a simple grid
• Use multiple resources of different types in a
  staged or concurrent computation
• Use as a full grid
• Use multiple nodes as an ensemble via advanced SW
  environment

27
Scaling TeraGrid -- ETF

• 4 TeraGrid sites PSC have just responded to NSF
  Dear Colleague letter for Extensible Terascale
  Facility (ETF)
• ETF will contain
• More networking
• More data
• Larger nodes
• Heterogeneity

ETF Team Fran Berman (SDSC) Charlie Catlett
(ANL) Ian Foster (ANL) Paul Messina (CalTech,
ANL)Mike Levine (PSC) Dan Reed (NCSA) Ralph
Roskies (PSC) Rick Stevens (ANL)
28
Heterogeneity

• What does it mean to add heterogeneous nodes to
  TeraGrid?
• Ensure that basic services supported on all
  architectures
• GRAM, GridFTP, etc.
• Ensure that core services supported on all
  architectures
• GIIS, NWS, SRB, MPICH-G, etc.
• Develop mechanism for scheduling between
  architectures
• sophisticated techniques will require research
• Continually monitor system to ensure SW
  compatibility
• Ensure that SW mitigates differences
• data formats, byte ordering, etc.
• Deploy consistent user interfaces and portals
• Help support a growing application community

29
Has the Grid been Grid oversold?

• The promise of the Grid has been not been
  oversold but the difficulty of developing the
  requisite Grid infrastructure has been
  underestimated.

30
The Grid is more than just a development and
integration project

• E.g. TeraGrid was developed as a vision for the
  future, which needs to accomplished
• Within a short time frame (3 years)
• Using current and emerging products
• Leveraging current research
• Targeting a current set of cutting edge
  applications
• There are many questions not addressed by
  TeraGrid and other projects that must be
  addressed to develop a usable and useful Grid
  information infrastructure

31

• We have barely scratched the surface on
• Program development environments (e.g. compiling
  for the grid)
• Debugging
• Fault tolerance
• Modeling of dynamic, unpredictable environments
• Grid market economy (allocation, accounting, cost
  models)
• Extreme heterogeneity (sensors, supercomputers,
  cell phones, cars, etc.)

32
Grids The Next 10 Years
Picture ofdigital sky
33
Ultimate Goal A useful, usable, stable Grid
that is

• High-capacity (rich in resources)
• High capability (rich in options)
• Persistent (promoting stable infra
  knowledgeable workforce)
• Evolutionary (able to adapt to new technologies
  and uses)
• Usable (accessible, robust, easy-to-use)
• Scalable (growth must be a part of the design)
• Adequately supported (both in funding and
  commitment)
• Useful, able to support/promote new science
• More cooperative than competitive

34
Applications are key to the Grids success

• Applications will use whatever parts of the
  infrastructure that can really deliver
• Apps developers are willing to be dedicated and
  creative but it has to be worth their while
• Goal is for Grid infrastructure to some day be as
  natural a part of the picture as the OS
• Grid will be considered oversold if the only
  people who can productively use it are the
  techies

35
New Application Paradigms for the Grid

• Next generation Grid applications
• Adaptive applications (run where you can find
  resources satisfying criteria X)
• Real-time applications (do something right now)
• Coordinated applications (dynamic programming,
  branch and bound)
• Poly-applications (choice of resources for
  different components)
• We still cant throw any application at the
  grid and have SW determine where and how it will
  run

36
Focus on Adaptive Applications
Everyware -- a highly adaptive Grid application
which investigated solutions to the Ramsey Number
Problem

• Everyware Wolski, SC98 ran on

• Berkeley NOW
• Convex Exemplar
• Cray T3E
• HPVM/NT Supercluster
• IBM SP2
• Intel X86
• SGI
• Sun SPARC
• Tera MTA
• Laptops

• Batch Systems
• Condor
• Globus
• Java
• Legion
• Netsolve
• Unix
• Windows NT

all at the same time
37
Everyware adapted to whatever resources were
available

• Application was
• Ubiquitous -- able to run everywhere
• Resource Aware capable of managing
  heterogeneity
• Adaptive -- able to dynamically tailor its
  behavior to optimize performance
• NOT embarrassingly parallel -- Branch-and-Bound
  and Simulated Annealing used

38
Focus on Performance Grid Programming
Environments

• Grid-friendly libraries, compilers, schedulers,
  performance tools
• Program performance through adaptation
• Contract-based grid performance economy

• The GrADS (Grid Application Development Software)
  Project
• Design and development of a Grid program
  development and execution environment

39
Focus on Data A Killer App for the Grid

• Over the next decade, data will come from
  everywhere
• Scientific instruments
• Experiments
• Sensors and sensornets
• New devices (personal digital devices,
  computer-enabled clothing, cars, )
• And be used by everyone
• Scientists
• Consumers
• Educators
• General public
• SW environment will need to support unprecedented
  diversity, globalization, integration, scale, and
  use

Data from instruments
Data from simulations
Data from analysis
40
From Data to Information to Knowledge
High speed networking
Networked Storage (SAN)
instruments
Storage hardware
sensornets
SDSC Data and Knowledge Systems Program
41
Developing a Data Condominium

• Well-defined interfaces for tools, services
• Services to be added, swapped in as they evolve

A Data Condominium Grid Data
Tools/ServicesFramework
42
Next generation Grids will include new
technologies

• New devices
• PDAs, sensors, cars, clothes, smart dust,smart
  bandaids,

• Wired and Wireless
• HPWREN (Hans-Werner Braun, Frank Vernon et al.)
• 45 Mbps between Mount Laguna telescope and SDSU
• Wireless access to Pala, Rincon, La Jolla Indian
  Reservations, etc.

43
Fiber, Wireless, Compute, Data, Software

• Campus Wireless

UCSD campus GridFrom Sensors to Supercomputers
44
Global Information InfrastructureA Grid of Grids
45
The Global Information Infrastructure must cross
technical, political, social boundaries
46
Policy and Social Dyanmics

• Policy issues must be considered up front
• Social engineering will be at least as important
  as software engineering
• Well-defined interfaces will be critical for
  successful software development
• Application communities will need to participate
  from the beginning

47
A New Model of Interaction is Needed

• The Grid is about cooperation
• Process of building and using the Grid predicated
  on shared resources, agreement, coordination
• Will need to identify/adopt systems that
  incentivize the individual to contribute to the
  success of the group
• Examples highway driving, EOT PACI, single line
  bank queues
• Cooperation must bridge technological, political,
  social boundaries
• Will need to settle issues of turf, credit,
  resources

48
Weve made a great start but there is much
farther to go

• New Research
• Fault tolerance
• Compilers, performance prediction, scheduling
• Agent-based computing
• Location-independence
• Extreme heterogeneity
• Applications which push the envelope
• Applications with dependences
• Adaptivity, poly-algorithms, commercial
  applications

• Policy and economics for grid environments
• Sharing as a default mode of interaction
• Trust, policy, negotiation, payment
• Usability and performance
• Programming environments for the Grid, portals
• Adaptivity as the prevalent mode for performance

Drivers Wanted We should be developing a new
generation of scientists, technologists and
solutions to address the challenges of a Global
Grid Infrastructure
49
Thank You