Transcript Peer to Peer Grids and Collaboration Euresco Conference at Castelvecchio Pascoli Italy June 16-21 2001 EuroConference on Problem-Solving Environments for Numerical Mathematics, Science and.

Peer to Peer Grids
and Collaboration
Euresco Conference at Castelvecchio Pascoli Italy June 16-21 2001
EuroConference on Problem-Solving Environments for Numerical
Mathematics, Science and Engineering Applications
Geoffrey Fox
IPCRES Laboratory for Grid Technology
Computer Science, Informatics, Physics
Indiana University
Bloomington IN
[email protected]
11/6/2015
p2pgrideuropejune01
1
Some Possible Truths


Technology is rapidly changing and current projects
(both software and architecture) are ephemeral
– Some things like CORBA and Jini may die
– Grids will become P2P Networks
– Java may become C# (hope not)
– XML will become more important – especially for
objects not just documents
– Semantic Web good concept
Technologies are disruptive and cannot be ignored
– Industry not academia developing most important base
technologies
– Not so interesting to set standards – not at MPI stage
11/6/2015
p2pgrideuropejune01
2
Classic Grid Architecture
Database
Database
Composition
Neos
Netsolve
Security
Portal
11/6/2015
Resources
Clients
Middle Tier
Brokers
Service Providers
Portal
p2pgrideuropejune01 Users
and Devices
3
Peer to Peer Grids

A Computational Grid can be defined as a collection of
computers, on-line instruments, data archives and
networks that are all tied together by a shared set of
services which, when taken together, provide users with
transparent access through interface devices to the
entire set of resources.
– Initial focus to a power users accessing major resources

A P2P Network can be defined as a collection of users,
interface devices, computers, on-line instruments, data
archives and networks that are all tied together by a
shared set of services which, when taken together,
provide everything transparent access to everything
else.
– Initial focus to building communities and access to commodity
resources such as MP3 files
11/6/2015
p2pgrideuropejune01
4
Peers
Peer to Peer Network
User
Service
Resource
Routing
User
Service
Resource
Routing
User
Service
Resource
Routing
Peers are Jacks of all Trades linked to “all” peers in community
User
Service
Resource
Routing
11/6/2015
User
Service
Resource
Routing
p2pgrideuropejune01
User
Service
Resource
Routing
5
Peer to Peer Grid
User
Service
Resource
Routing
User
Service
Resource
Routing
GMS Routing
User
Service
Resource
Routing
Services
GMS or GES is Grid Message/Event Service
User
Service
Resource
Routing
11/6/2015
User
Service
Resource
Routing
p2pgrideuropejune01
Dynamic
Message or
Event
Routing from
Peers or
Servers
User
Service
Resource
Routing
6
.antinet

This is emerging approach to Internet Systems by
everybody except Microsoft
New distributed database?
Billions of ……
Java
Business
Logic
Discovery
Distributed XML
File systems
Or Databases
FS
XML
objects
Enterprise
Javabeans
XML Objects
Castor
XML > Java
11/6/2015
Database
Oracle 9i
p2pgrideuropejune01
7
Collaboratory Applications








Distance Education including advanced seminars and training
Help Desk including
– Microsoft helping user debug problem on home PC (connected to
Internet)
– Computer center consulting staff interacting over distance in real
time with a user with a program bug
– Yahoo staff asking in depth questions from users browsing either
their knowledge or Shopping sites
Scientists brainstorming difficult research issues in distributed
locations
Virtual communities around the world from children chatting to
each other or integration of distributed organizations (like ARL)
Indian Nation remaining in their homeland but participating
electronically in modern economy (digital.indigineousworld.org)
Implementing next round of PET activities
Crisis Management and Command and Control for Military
Peer to Peer Networks
11/6/2015
p2pgrideuropejune01
8
What Issues are we trying to address?

A framework to build custom collaboratories
– Education and Training – the Virtual University
– Custom Shared Objects – not just shared PowerPoint but
shared 3D visualizations (for collaborative visualization and
gaming school)
– Science and Engineering Research

Integrate Portals (WebCT) and Collaboration
(WebEx) as both built around accessing and sharing
objects
– We build laboratories (collaboratories) – not desks
(conventional portals)

Universal Access – CAVE, PC, Settop Box, PDA,
Braille Machine
– Separate specification and rendering of objects
11/6/2015
p2pgrideuropejune01
9
Goals and Features of Garnet

Integrate synchronous and asynchronous
collaboration
– Use same (Java) publish/subscribe Message Service to support
both forms of collaboration

Integrate concept of a portal (web interfaces to
applications) with collaboration
– Use same XML object metadata specification GXOS to
manage and share objects

Supports universal access including PDA’s
collaborating with desktops
– PDA Control of device or PDA alternative kiosk access gotten
as a collaborative session between device and PDA
– Object metadata includes rendering information for different
devices
11/6/2015
p2pgrideuropejune01
10
Different Ways of Sharing

All forms of Collaboration are Event based
– Different modes: Display, Export, “Event” correspond to events
generated at different places in object realization/rendering
pipeline



Shared Display – Events contain updates to frame
buffer
Shared “Event” – Events contain updates to state of
either original or transformed object
Shared Dynamic Export – Convert (rendering of)
object to some standard form that is more flexible than
bitmap of Shared Display. Build a custom sharing for
this exported form
– WebEx uses “patented sharing of virtual printer” which is
equivalent to sharing export to PDF
– I like shared HTML (web pages) or SVG described later
11/6/2015
p2pgrideuropejune01
11
Important Capabilities in Initial Garnet

Standard stuff: built in shared display, whiteboard, HearMe
Audio control, quizzes, annotations, chat/IM (Jabber.org)
– Desktop video will be special case of shared display

Record and replay all features of session (SMIL)
– A/V, Presentation, Annotations, Text Chat
Several Specialized Collaborative Shared Export Viewers: JSP,
HTML, Acrobat ..


Initial SVG (Scalable Vector Graphics) Shared Batik Viewer
– 2D Scientific Visualization/Whiteboard
– Macromedia (Flash~SVG) and Adobe (already “all” to SVG)

Initial source of SVG: Convert PowerPoint VML/WMF to SVG
– Gives shared export model for PowerPoint with each client able
to scale independently at high resolution
11/6/2015
p2pgrideuropejune01
12
Garnet Architecture Underpinnings





Most software written in Java (EJB); all data
structures defined in XML GXOS
Unified Event Model – assume all systems
communicate with XML based messages (possibly
wrapped); universal rendering – portalML
All entities – client and server devices, users,
programs -- defined in XML – resourceML
All events archived to enable fault tolerance and
replay
GMS -- Garnet(Grid?) Message Service -- extends
JMS to XML not text properties to allow server side
intelligence
– Dynamic Server Clusters
11/6/2015
p2pgrideuropejune01
13
Publish/Subscribe Collaboration





Integrate Asynchronous and Synchronous Collaboration
The web is full of objects – web pages sitting on web servers –
and these support asynchronous collaboration
– You post a web page and I later look at it in my own time
Replacing web document by a “CGI script” or servlet (web
interface to program, database etc.) gives general multi-tier
object sharing
This is Publish/Subscribe mechanism
– If add some mechanism (automatic email or word of mouth) to
tell viewing client when new information is posted
– We build on JMS (Java Message Service) as Industry standard for
publish/subscribe systems
Synchronous Collaboration provides “real-time” notification
and automatic update of changed objects
– Is JMS fast enough to do real-time? need “HP”JMS
11/6/2015
p2pgrideuropejune01
14
Everything communicates by Messages






What have we learnt from parallel computing,
clusters and the Internet?
Message Passing Rules …
Note an event is just a message with (an optional)
time stamp and typically (but not necessarily) an
active (interrupt) message processing paradigm
Message passing enables modularity, fault tolerance,
distribution
MPI for Parallel Computing – low latency etc.
JMS (or better “GMS” Grid Message Service”) high
functionality
11/6/2015
p2pgrideuropejune01
15
JMS (Java Message Service) Structure in Garnet
Basic primitive is a topic/property labeled queue = JXTA Pipe
JMS Global (distributed)
Event Receptor (Queue)
Subscribe
Subscribe
Publish
HHMS
Convert Events
to JMS
JavaScript
11/6/2015
Java C++
…..
HHMS (Hand Held Message
Service) Optimized for
Performance.
p2pgrideuropejune01
16
Performance of Commercial JMS I
Latency (Non-Persistent)
10000
JMQ
milli-seconds
1000
iBus
SonicMQ
100
FioranoMQ
10
1M
16
K
64
K
25
6K
4K
1K
25
6
64
16
4
0
1
Non-persistent as
We do database
backup outside JMS
Message Size (bytes)
11/6/2015
p2pgrideuropejune01
17
Performance of Commercial JMS II
Non-Persistent/Non-Durabl e
Number of Topic : 5
pub/sub per Topic : 10
Messages per Second
Message Size (bytes) : 100
1000
800
iBus
FioranoMQ
SonicMQ
600
400
200
0
Server:
Clients:
11/6/2015
pub
sub
Solaris
1 PC
pub
sub
Windows
1 PC
pub
sub
Solaris
2 PCs
p2pgrideuropejune01
Non-durable as
our database
copes with late
clients
18
Proposed GMS Model for Messages

All message publication labels and subscription profiles are
defined in XML
Subscribes to all events to get
Database persistence
Subscriber
Profile Objects
Specify Query to
Event Label
Message Queue
Labeled by (XML)
Topic Object
Subscribers
Publishers
11/6/2015
p2pgrideuropejune01
19
Peer to Peer P2P Networks



(Synchronous) Collaboration is critical service in P2P
networks
Publish/Subscribe is mechanism we use to establish
who gets what information
Gnutella and JXTA are different implementations
(from JMS) of P2P information propagation
– GMS can be built on top of JXTA or JMS architecture

JXTA like MyXoS identifies the implicit distributed
operating system
–
–
–
–
Both have message queues as primitives
Both have Shell
Both use XML based messages
JXTA Advertisements are similar to GXOS metaobjects
11/6/2015
p2pgrideuropejune01
20
Single Server P2P Illusion
Data
base
JMS/GMS
Server
11/6/2015
p2pgrideuropejune01
21
Multiple Server P2P Illusion
Data
base
JMS
Server
JMS
Server
Generate “Automatically”
JMS
Server
11/6/2015
p2pgrideuropejune01
22
P2P Grid Event Delivery Service





Dynamic Collection of some billion computers each
of which can either generate, route or consume events
Publisher labels events by an (XML) object which is
at simplest a URI but in general a collection of tagvalues
Subscribers issue some sort of XML Query e.g.
deliver all
p2pgrid://garnet/Education/Graduate/ComputerScience/
Syracuse/Spring2001/CPS616/Lecture3/*
Need Secure, High Performance, Efficient (don’t
propagate events further than they need), Fault
Tolerant delivery service
Shrideep Pallickara PhD June 1 2001
11/6/2015
p2pgrideuropejune01
23
Some Results – 22 Servers
Servers are
logically but
not necessarily
physically distinct
from clients
11/6/2015
p2pgrideuropejune01
24
Match Rates of 10%
11/6/2015
& Server Hop to client = 1
p2pgrideuropejune01
25
Match Rates of 50% & Server Hop to client = 4
11/6/2015
p2pgrideuropejune01
26
Integration of Hand Held Clients
Shared Export
PDF/HTML/
SVG Viewer
Part of
Jabber IM
JMS/GMS
Server
Adaptor
XML Processor
GMS Processor
11/6/2015
p2pgrideuropejune01
Shared Display
Processor
27
Integration of Hand Held Clients

Client Device (machine) Profile stored in GXOS
specifies O/S, default Screen Size modified by user
(person) preferences
– Dynamically updated with connection bandwidth, user updates


Application Profile stored in GXOS and modified by
event stream specifies data delivered by GMS
Adaptor (Personal Server) looks like a conventional
client to GMS and adapts data to specified client/user
specifications
– If PDA “small”, then SVG viewer on “adaptor” and ship
framebuffer to PDA
– “Resize” on PDA handled by adaptor, scrolling by PDA etc.
– Adaptor can process complex XML queries
11/6/2015
p2pgrideuropejune01
28
PowerPoint
Shared
Display
PC to PDA
11/6/2015
p2pgrideuropejune01
29
Instant Messenger
Sharing PC to PDA
11/6/2015
p2pgrideuropejune01
30
Batik Viewer on PC
PowerPoint can be converted to SVG
via Illustrator or Web export
11/6/2015
SVG Sharing PC to PDA
p2pgrideuropejune01
31
MetaObject Event-based Computing Paradigm




All entities are MetaObjects defined in universal GXOS
XML Syntax – Meta means GXOS doesn’t really want to
manage Object, just information required to find, render
and share it
– Rendering includes Palm devices as well as PC’s
– Entities are people, computers, data sources, information
from e-mail, scientific visualization to digitized Bible
All actions including object changes are events – all events
are GXOS objects
– e-mail, paging etc. are all GXOS events uniformly
routed/archived etc.
All action is instantiated or recorded in messages which are
events themselves
There is an MetaObject Shell MyXoS with basic Services
(copy, create, collaborate etc.)
11/6/2015
p2pgrideuropejune01
32
A Lesson from UNIX, Windows, Yahoo …

Structured information is always labeled as a hierarchy with
symbolic links e.g. xxx://A(home=root)/B/C/D ….
We have
search
engines for
unstructured
information
Symbolic Links
11/6/2015
p2pgrideuropejune01
33
Confirmed by ADL and W3C DOM






DoD’s advanced distributed learning effort (and IMS more
academic consortium) use a tree structure for educational
material in SCORM
– http://www.adlnet.org
Add educational specific tags like prerequisites and completion
requirements
– Node of tree is called block
ADL have an educational hierarchy with extensions for the rest
of the world
Rather one should have a general tree with extensions for
documents, people, education, computers, events,transactions
….
W3C DOM is at times a document object model but it is really
“general object instance” looked at as a tree
UNIX and Windows file systems also hierarchical
11/6/2015
p2pgrideuropejune01
34
XML DTD Structure for Block Element
? globalProperties
source ~
+ externalMetadata
model ~
location ~
Internal Node
* objectiveRef ~
title ~
identification
? description ~
? curricular ~
? labels
? developer ~
? prerequisites ~
? completionReq ~
course
source ~
* extensions
block ~
model ~
location ~
+ property
+
* au ~
* block ~
blockAlias ~
? objectives
11/6/2015
p2pgrideuropejune01
name ~
value ~
Leaf Node
35
Structured and Unstructured Data





Two important hierarchies – both labeled by URI
– Categories for Structured metadata – GXOS
resources are labeled with these (GNDI)
– Locations for Structured and unstructured data
protocol://root/x/y/ … CORBA, Java SOAP .. link
stored in metadata
Xlink/Xpath can identify an object component (value of
an element) by hierarchical external location (URL) or
hierarchical internal name (URI) )
Unstructured data discovered by “Web Search”
Structured data by Category look-up or by “Web
Search”
GXOS like P2P Networks (e.g. Napster) labels objects
with metadata (why it is useful) not just cryptic name
as in filesystems
11/6/2015
p2pgrideuropejune01
36
GXOS GNDI Structure I

Typically GXOS does NOT hold the object but rather
all the metadata for object including where to find it
– internalND (XML/EJB) defines where the metadata is (GNDI
Garnet Naming/Directory)
– externalND where the object is (JNDI Java Naming/Directory
Interface or CORBA or COM or SOAP)
– Idea used in WebFlow/Gateway portals – implies efficient as use
natural high performance mechanisms for “real objects”

GXOS uses classic UNIX style directory structure such
as //Education/University/FSU/CS/PhD/Course/Lecture
– Structure says can handle quadrillion meta-objects
//Jefferson/HallD/Run1/May132005/Tape5/Event1234/Detector3 …

Events are a special type of GXOS object defined with
GMS Extension to basic hierarchical object
11/6/2015
p2pgrideuropejune01
37
GXOS GNDI Structure II



The more you know about an object, the easier it is to build a
collaborative system
– It is hard to share a general program as properties are not explicit
– You put in GXOS metadata precisely those properties that need
to be shared
All GXOS objects can be containers as well as having full object
characteristics
– Unlike UNIX where directories (only containers) and files (never
explicit containers) are different
Can create a link (internalND pointer) with a profile to define
special properties of this object in a new context
– Intelligent symbolic links
– One links User, Client Device to VE (Virtual Environment) with XML
profiles mapping user to this session --> Universal Access
11/6/2015
p2pgrideuropejune01
38
Overall Structure of GXOS for a MegaMeeting
Event
Archive
Capabilities
Global Root
Users
Devices
Documents
MegaMeeting
Meeting
Admin
Multimedia
Have a hierarchy of MegaMeetings
(any collection of meetings)
Course, Degree .. Are MegaMeetings
Meeting
Meeting
gxos://Education/University/FSU/CS/PhD/Course/Lecture
11/6/2015
p2pgrideuropejune01
39
Structure of GXOS For a Collaborative Session
GNDI Links
Users
An Application
Supporting GMS
Meeting
Devices
GNDI Links
Capabilities
Sharedlets
Sharedlet1
Sharedlet3
Sharedlet2
VEa
VEb
VE is Virtual
Environment
11/6/2015
VEc
VEa
VEa
VEb
p2pgrideuropejune01
VEb
VEc
VEc
40
Semantic Web: URI and Topics etc.





The URI jxtapipe://fred/jim/agatha is
equivalent to
<label objecttype=“topic”
queuetype=“jxtapipe” field1=“fred”
field2=“jim” field3=“agatha” />
The <label /> syntax is more general
URI labeling allows consistency with W3C
Semantic Web and indeed current practice
Everything is then a “resource” in RDF
resource is a labeled object
11/6/2015
p2pgrideuropejune01
41
RDF Examples in GXOS I

These are sample MyXoS Shell Statements

Specify value for property in GXOS tree
<rdf:description
about="gndi://gxosroot/resourcename">
<gxos:property rdf:parseType="literal">
somevalue</gxos:property> </rdf:description>

Specify profile by linking between GXOS tree elements
<rdf:description about="gndi://gxosroot/sessionname">
<gxos:userprofile rdf:resource="gndi://uri_of_user"
gxos:customize="sessionspecificstuff" />
</rdf:description>
11/6/2015
p2pgrideuropejune01
42
RDF Examples in GXOS II


Specify MyXoS copy command for meta-objects
<rdf:description
rdf:about="gndi://gxosroot/system/bin/cp"
system:source="gxosobject1"
system:destination="gxosobject2" gxos:execute="true“
/>
Specify alternative locations to find all FSU users
<rdf:description
aboutEachPrefix="gndi://gxosroot/users/fsu">
<gxos:metaobjectlocation> <rdf:alt>
<rdf:li resource="http://main_fsuweblocation" />
<rdf:li resource="http://backup_fsuweblocation" />
</rdf:Alt> </ gxos:metaobjectlocation>
</rdf:description>
11/6/2015
p2pgrideuropejune01
43
Fragment of
GXOS Schema
11/6/2015
p2pgrideuropejune01
44
Interface of XML and Java I

How will we teach computing?
– K-4: Internet Access
– Middle School: (Simple) XML Schema interfaced to some
scripting language
– High School: Java as the programming model with Java
classes (for external data) generated

Probably don’t want to specify objects twice
– Start in Java; generate Schema


Or Start with Schema and generate Java
Seems a natural API of computer code to XML
– DOM or SAX XML Parser or
– JDBC like Interface of Java to Database (~XML)
11/6/2015
p2pgrideuropejune01
45
Interface of XML and Java II


Suppose we have a quadrillion (1015)
XML objects as say produced by a
physics accelerator per year
(Enterprise GXOS)
Need to combine:
– Search Interface to select nodes of XML
Tree
 Specify URI
 JDBC or Google like Interfaces
– Castor like Interface to map XML into
Java but need to control depth of
conversion from XML into
Database
XML
Choose
And Convert
Software
11/6/2015
p2pgrideuropejune01
46
Controlled Conversion XML to Java

Control expansion by either
– recursively descending tree
– Access other nodes not in current sub-tree
– Access “real object” in a possibly different object
model
Node
URI
Another
Node
Child1
Child2
…
ChildN
URI
11/6/2015
p2pgrideuropejune01
Another
Object Model
CORBA Java
SOAP …..
47
Current GXOS API Architecture

Initially implement “Personal” GXOS – Information
Repository small enough that we can afford to read all
possibly relevant information into memory and refine
this
– E.g. Support course data for individual faculty

File.xml  XML Object  Java Object and vice versa
– Use Castor to automate XML Schema to Java Object

Primitives Supported Initially
– Get a “leaf Object”
– Get a Collection (Internal Node) – “handle” and self.xml (the
GXOS properties associated with this node)
– List Contents of a collection (recursively)
– Get Contents of a collection (recursively)
11/6/2015
p2pgrideuropejune01
48
Current GXOS API Architecture
Lookup: XML
 Java
Client Applications
2. FS JNDI command,
lookup(“file:///root/object.xml”)
FS
JNDI
1. GNDI commands,
lookup(“root/object”)
5.File
object
SAX
parser
1b. lookup(“root/Users/username”)
GXOS Schema
API
SourceGenerator
org.gxos.schema
8. Java
object
Unmarshaller
p2pgrideuropejune01
Castor
Java XML classes +
Class descriptors
for validation
Castor
7. XML file
access & parsing
11/6/2015
org.gxos
9. Java
object
GNDI
ObjectFactory
6.FileReader
object
XML
objects
Architectural API
org.gxos.gndi
4. File object
alternatively
1a. getUser(“username”)
GXOS
GNDI
3. File find and
Return file handle
(file object)
FS
10. Java object
GXOS
XML
Schemas
49