Grid Computing
DCS861A Emerging Computing II
Spring 2005
DPS Team 2
29/11/2011
What is “Grid Computing”?
“…a type of parallel and distributed system
that enables the sharing, selection, and
aggregation of geographically distributed
"autonomous" resources dynamically at
runtime depending on their availability,
capability, performance, cost, and users'
quality-of-service requirements.”
11/29/2011 Source: Grid Computing Info Centre (www.gridcomputing.com) 2
What is “Grid Computing”?
“…a type of parallel and distributed system
that enables the sharing, selection, and
aggregation of geographically distributed
"autonomous" resources dynamically at
runtime depending on their availability,
capability, performance, cost, and users'
quality-of-service requirements.”
11/29/2011 Source: Grid Computing Info Centre (www.gridcomputing.com) 3
What is “Grid Computing”?
“…a type of parallel and distributed system
that enables the sharing, selection, and
aggregation of geographically distributed
"autonomous" resources dynamically at
runtime depending on their availability,
capability, performance, cost, and users'
quality-of-service requirements.”
11/29/2011 Source: Grid Computing Info Centre (www.gridcomputing.com) 4
Where Are These Resources?
Mainframes are idle
about 35% of the time
UNIX servers are
actually "serving"
something less than
15% of the time
And most PCs do
nothing for 95% of a
typical day
Imagine an airline
with 85% of its fleet
on the ground, an
automaker with 35%
of its assembly
plants idle, a hotel
chain with 95% of its
rooms unoccupied!
11/29/2011 5
“Computing Grid As Utility”
A common metaphor in the literature:
“…[a computing grid is] analogous to electric
power network (grid) where power
generators are distributed, but the users
are able to access electric power without
bothering about the source of energy and
its location.”
― Grid Computing Info Centre
11/29/2011 6
“Grid as Utility” Origins
Early on in 1969, Len Kleinrock, one of the original
Arpanet designers, wrote…
“We will probably see the spread of
„computer utilities‟, which, like present
electric and telephone utilities, will
service individual homes and offices
across the country.”
11/29/2011 7
On-demand, Dispersed Resources
Quality, economies of scale
Decouples production &
consumption, enabling…
On-demand access
Economies of scale
Consumer flexibility
New devices
11/29/2011 8
Source: Ian Foster, U. of Chicago
Time
Grid Computing Scales
Cluster Grids Enterprise Grids Global Grids
11/29/2011 9
“But Computing isn’t Electricity”
Usually users only consume electricity,
they don’t also produce it ― software
applications both consume and produce
data
“Computing” is not a homogenous “thing”,
but is highly heterogeneous: data,
sensors, services, software, computing
hardware, …
11/29/2011 10
“But Computing isn’t Electricity”
This complicates things; but, it means that the
result can be greater than the sum of the parts
Also it raises some fundamental questions…
Building applications that exploit the infrastructure?
Operating such a complex environment?
Managing heterogeneous resources not centrally
owned?
Ensuring QoS across these distributed services?
11/29/2011 11
Another Way of Looking at Grids
From a less technical viewpoint:
“Grid computing has emerged as an important
new field, distinguished from conventional
distributed computing by its focus on large-scale
resource sharing, innovative applications, and,
in some cases, high-performance
orientation...we [define] the "Grid problem”…as
flexible, secure, coordinated resource sharing
among dynamic collections of individuals,
institutions, and resources - what we refer to as
virtual organizations.”
The Anatomy of the Grid
Enabling Scalable Virtual Organizations
11/29/2011 Ian Foster, Carl Kesselman, Steven Tuecke 12
Intl. Journal Supercomputer Applications, 2001
Virtual Organizations (VOs)
In VOs a grid infrastructure is more a means to an end:
Enables integration & sharing of distributed resources
Removes geographical constraints on teams
Creates consistent qualities of service via fault-
11/29/2011 tolerance, dynamic workload balancing, etc. 13
Grid History: I-WAY ― A Seminal Event
Experiment led by researchers at the University of Illinois
at Chicago and Argonne National Laboratory
For a week in Nov 95, it linked 11 research networks to
create one high-speed network infrastructure
Connected 17 sites across the US and Canada
Demonstrated 60 applications, from distributed
computing to virtual reality collaboration
Attempted to construct a unified software infrastructure
providing scheduling, single sign-on, and other grid-
enabled services
11/29/2011 14
Early Grids: Govt.-funded Science
GUSTO (1998): 80 global research sites
3,000+ host grid software testbed
NASA Information Power Grid (since 1999)
Production grid linking NASA laboratories
INFN Grid, EU DataGrid, iVDGL, … (2001+)
Grids for data-intensive science
TeraGrid, DOE Science Grid (2002+)
Production grids linking supercomputer centers
U.S. GRIDS Center
Software packaging, deployment, support
11/29/2011 15
Why are Grids Hot Now?
Hardware performance improving exponentially
Computer speed doubles every 18 months
Network speed doubles every 9 months
Difference = order of magnitude every 5 years
1986 to 2000…
Computers: x 500
Networks: x 340,000
2001 to 2010…
Computers: x 60
Networks: x 4,000
11/29/2011 16
Moore’s Law vs. storage improvements vs. optical improvements. Graph from Scientific American (Jan-2001) by Cleo Vilett, source Vined Khoslan,
Kleiner, Caufield and Perkins.
Why are Grids Hot Now?
Grids begin to address some real world IT issues:
Low overall utilization of enterprise resources
High cost of provisioning for peak demand
Lack of information integration
Physical distribution of teams is increasing
Inability to apply available resources to advanced
computation & data-intensive applications when and
where they are needed
However, the marketing hype is outrageous; every
possible SW & HW product has been “gridified”
11/29/2011 17
Early Commercial Adopters
Aerospace and Automotive (for collaborative
design and modelling)
Architecture (engineering and construction)
Electronics (design and testing)
Energy (for oil and gas for exploration)
Finance/insurance/real estate (securities and
brokerage especially for stock/portfolio analysis
and risk management)
11/29/2011 18
Early Commercial Adopters
Life sciences (particularly in pharmaceuticals)
Manufacturing (inter/intra-team collaborative
design, process management)
Media/entertainment (to generate digital
animation)
Utilities (to improve efficiency while dealing with
peaks and valleys in utilization)
11/29/2011 19
Grid Market Projections
Leading adopters (Oct 2003)…
•Financial services: 31%
Grid Services Market Opportunities 2005
•Life sciences: 26%
•Manufacturing: 18%
Manufacturing
Financial
Services Mechanical/ LS /
Electronic Bioinformatics Other
Design
Energy Derivatives
Analysis Process Cancer Entertainment Web
Seismic Simulation Research Applications
Statistical
Analysis Analysis Finite Drug Digital Weather
Reservoir Element Discovery Rendering Analysis
Portfolio Analysis
Analysis Risk
Protein Massive
Analysis Multi-Player Code Breaking/
Failure Folding Simulation
Analysis Games
Protein Academic
Sequencing Streaming
Media
“Gridified” Infrastructure
11/29/2011 20
Sources: IDC, 2000 and Bear Stearns- Internet 3.0 - 5/01 Analysis by SAI
Example Adopter: Novartis
PC-based grid of
3,700 desktop systems
“We have projects we
R&D pharmaceutical
calculate would take 6
applications
years on a single
Potentially mainstream
business computing
supercomputer.
> 5 teraflop/s
Today, the run time is
computing power 12 hours.”
Estimated savings of
$200M over 3 years ― Peter Sany, Novartis CIO
11/29/2011 21
Grid Application Attributes
Computational complexity
Genome research
Financial product creation
Geophysical studies
Digital animation creation
Massive data requirements
Digital mammography diagnostics
Particle physics research
Astronomical observation analysis
11/29/2011 22
Computational Complexity:
Protein Analysis
Example: Determining
the structure of a
complex molecule,
such as the cholera
toxin shown here, is
the kind of
computationally
intense operation that
grids are intended to
tackle
(Adapted from G. von Laszewski et al., Cluster
Computing, volume 3(3), page 187, 2000)
11/29/2011 23
Massive Data Requirements
Storage density doubling every 12 months
Dramatic growth in online data (1 petabyte
= 1000 terabytes = 1,000,000 gigabytes)
2000 ≈ 0.5 petabyte
2005 ≈ 10 petabytes
2010 ≈ 100 petabytes
2015 ≈ 1000 petabytes?
These are sometimes called “data grids”
11/29/2011 24
Massive Data Requirements:
Digital Mammography
Digital Radiology (hospital digital data)
Mammogram X-rays
MRI / CAT scans
Endoscopies
Very large data sources
7 terabytes per hospital per year
Dominated by digital images
11/29/2011 25
Massive Data Requirements:
Digital Mammography
Why target
mammography?
Increasing need for film
recall & computer analysis
Large volumes (4,000
GB/year ― 57% of total)
Storage and records
standards exist
Great clinical value
11/29/2011 26
Grid Management Challenges
Scale of data and compute resources is huge
QoS and performance criteria are severe
Platform must be scalable, able to evolve, fault-
tolerant, robust, persistent and reliable
It should work seamlessly, and transparently –
the user might not know or care where their
calculation is done using how many machines,
or where data is actually held
11/29/2011 27
Grid Management Challenges
Resource configurations are transient, dynamic
and volatile as services (databases, sensors,
compute servers) are switched in and out
They are ad-hoc as service consortia have no
central location or control and no existing trust
relationships
They may be large, with hundreds of services
orchestrated at any time
They may be long-lived, for example a protein
folding simulation could take weeks
11/29/2011 28
Technical Challenges
How does a grid infrastructure, in a dynamic, multi-
institutional, physically distributed setting,…
Locate suitable computers?
Authenticate & authorize user requests?
Allocate resources on those computers?
Select appropriate communication methods?
Configure the computations?
Initiate these computations on those computers?
Access data files and return output?
Respond appropriately to resource changes?
11/29/2011 29
Grid Software Sources
Academic & Scientific Researchers
U. of Chicago & USC (Globus Toolkit)
UC Berkeley (BOINC)
Public consortium-based organizations
Global Grid Forum (OGSA)
Commercial Vendors
IBM, Entropia, United Devices, etc.
11/29/2011 30
Globus Toolkit (www.globus.org)
Includes software for…
Early open-source
security
grid infrastructure
toolkit information
infrastructure
Set of protocols, resource management
services & software data management
libraries that supports communication
grids and grid
fault detection
applications
portability
11/29/2011 31
Evolving Open Grid Standards
Research Managed shared
virtual systems
Increased functionality,
Open Grid
Web services, etc.
standardization
Services Arch
Real standards
Multiple implementations
Internet
standards Globus Toolkit
Defacto standard
Custom Single implementation
solutions
1990 1995 2000 2005 2010
11/29/2011 32
OGSA (www.gridforum.org)
Grid technologies ― including the Globus
Toolkit ― are evolving toward the Open
Grid Services Architecture (OGSA)
OGSA provides an extensible set of
services that virtual organizations can
aggregate in various ways
Built on concepts and technologies from
both the Grid and Web services
communities
11/29/2011 33
OGSA
OGSA defines:
Grid service semantics (like Web services)
Standard mechanisms for creating, naming, &
discovering transient grid service instances
Location transparency and multiple protocol
bindings for service instances
Support for integration with underlying native
platform facilities
11/29/2011 34
OGSA
OGSA also supports (via WSDL):
creating/composing complex distributed systems
lifetime management
change management
notification
reliable invocation
authentication & authorization
11/29/2011 35
Grid Standards: Summary
Grid Services and Web Services are merging
Web Services standards landscape is in flux
OGSA will need to evolve with it
Fuzzy security & policy standards are a concern
W3C, OASIS, GGF are key standards orgs
Open source software important for adoption
11/29/2011 36
Some Commercial
Grid Software Vendors
IBM (www.ibm.com/grid)
Avaki (www.avaki.com)
GridIron Software (www.gridironsoftware.com)
United Devices (www.ud.com)
Platform Computing (www.platform.com)
DataSynapse (www.datasynapse.com)
Entropia (www.entropia.com)
Oracle 10g (www.oracle.com/technologies/grid)
11/29/2011 37
“Wait a second! What about…”
SETI@home (extra-terrestrial signal search)
GIMPS (Great Internet Mersenne Prime Search)
folding@home (protein manipulation)
Distributed.net (brute force decryption)
…and all those other Internet “grid” projects
I’ve been reading about?
11/29/2011 38
“Public Resource” Computing
These are all examples of what Dave Anderson
of Berkeley calls “public resource computing”
Most of the world's computing power is no longer in
supercomputer centers or institutional machine rooms
Instead, it is now distributed in the hundreds of
millions of personal computers, game consoles, and
TV set-top boxes
“If all this computing power could be made
available to researchers somehow…”
11/29/2011 39
Hallmarks of Public Resource Computing
Public resource computing shares some traits
with grid computing, but is qualitatively different
“Open” vs. “closed” society of resources
“Asymmetric usage”: more suppliers of resources
than consumers, e.g., millions of PC screensavers vs.
small team of researchers
Must be able to attract “altruistic” participants
Often some “reward” mechanisms will exist for
resource suppliers
11/29/2011 40
Public Resource Application Profile
High computing to data ratio is typical
Computation independence & parallelism
is crucial
Must be tolerant to errors and outages
Must be able to handle “malicious” users
Sporadic connectedness is the norm
11/29/2011 41
Public Resource vs. Grid Computing
Public-Resource Grid
Managed resources? no yes
Secure resources? no yes
Always on? no yes
Always connected? no yes
Network bandwidth Expensive, scarce abundant
Network connection 1 way (pull) 2 way (pull or push)
Must be unobtrusive? yes no
Credit system? yes maybe
How to get resources complex complex
Public education/outreach? yes no
Self-upgrading? yes no
11/29/2011 Source: David Anderson, BOINC project (UC Berkeley) 42
Example: SETI@home
SETI = “Search for Extraterrestrial Intelligence”
Goal: detect intelligent life outside the Earth
Uses radio telescopes to listen for narrow-
bandwidth radio signals (not known to occur
naturally) from space
Initial version used hand-crafted server
architecture and workstation clients
11/29/2011 43
SETI Computational Model
Signal data is divided into fixed-size work units
that are distributed, via the Internet, to a client
program running on numerous computers
Client program computes a result (a set of
candidate signals), returns it to the server, and
gets another work unit
Each work unit is processed multiple times to
detect and discard results from faulty processors
and from malicious users
11/29/2011 44
SETI@home at Work
11/29/2011 45
SETI@home Technical Specs
SETI@home client program is written in C++
Platform-independent framework with platform-
specific implementations
graphics library
SETI-specific data analysis code
SETI-specific graphics code
Client ported to 175 different platforms using the
GNU toolset
Client can run as a background process, as a
GUI application, or as a screensaver
11/29/2011 46
SETI@home Results to Date
Totals Last 24 Hours
(as of 03/31/2005)
Users 5,388,068 784
Results received 1,811,656,328 1,339,532
Total CPU time 2,251,657.404 years 925.204 years
Floating Point 5.224175e+18
6.649645e+21
Operations (60.46 TeraFLOPs/sec)
Average CPU time
10 hr 53 min 15.2 sec 6 hr 03 min 01.6 sec
per work unit
11/29/2011 47
Lessons from SETI@home
Public resource computing concept does
work, but…
How do you make it easy for researchers to
access the public’s resources & good will?
How do you make it easy for the public to
contribute their resources to multiple projects?
One answer: the BOINC public resource
computing platform from UC Berkeley
11/29/2011 48
BOINC Goals
For computing projects
easy/cheap to create and operate projects
support a wide range of applications
no central authority
For participants
easy to participate in multiple projects
resource allocation among projects
invisible use of disk, CPU, network
11/29/2011 Source: David Anderson, BOINC project (UC Berkeley) 49
BOINC Architecture
11/29/2011 50
Some BOINC-based Projects
SETI@home (updated for BOINC support)
Predictor@home (protein-related disease)
Einstein@home (gravity waves, LIGO)
CERN (particle physics)
UCB/Intel network performance study
climateprediction.net (future climate impact)
11/29/2011 51
Example: climateprediction.net
The Earth is likely to warm over the coming
century. Question is by how much?
climateprediction.net is the world’s largest
climate modelling experiment to try and
answer this question
62,000 participants in 130 countries (8/04)
11/29/2011 52
climateprediction.net Summary
1. Each user downloads and runs a unique
simulation model of the Earth's climate
2. Models undergo an initial calibration
3. Each model is tested by simulating 20th century
climate
4. Models which cannot reproduce present and
past climate are discarded
5. All remaining models are run to predict the 21st
century climate
6. These results create the probabilistic forecast
for the 21st century climate
11/29/2011 54
For More Information
Globus Alliance
www.globus.org
Globus Consortium
www.globusconsortium.com
Global Grid Forum
www.ggf.org
Open Science Grid
www.opensciencegrid.org
Grid Today newsletter
www.gridtoday.com
Grid Blog
www.gridblog.com
BOINC
boinc.berkeley.edu
11/29/2011 55