Service

Large-scale distributed
computing systems
Lecture 2:
Remote services
December 2014
Johan Montagnat
CNRS, I3S, MODALIS
http://www.i3s.unice.fr/~johan/
Course overview
►
►
►
►
►
►
►
►
1. Distributed computing and models
2. Remote services and cloud computing
3. Infrastructures and deployment
4. Workload and performance modeling
5. Workflows
6. Authentication, authorization, security
7. Data management
8. Evaluation
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
2
Course content
►
2. Remote Services
▸
▸
▸
Services and interoperability
Web Services, WS-RF, WS-*
Platform as a service, Cloud computing
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
3
Service-Oriented Architectures
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
4
Service architectures
►
Middleware
▸
▸
▸
►
is distributed software
requires interoperability, despite heterogeneity
requires platform independence (e.g. MS Windows interfaces,
Unix computing platforms)
Service architectures help in achieving these goals
▸
▸
▸
▸
Earlier initiatives (e.g. GT2) based on more static technologies
led to difficult middleware/application integration
Web Services widely developed during the late 90s
OGSA standards emerging in 2003
WS-RF is an evolution of WS and OGSA towards convergence
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
5
►
Modular design
▸
►
Object-oriented software (C++/Java)
▸
►
Reuse functions: classes
Component-based software
▸
►
Reuse code: subroutines and functions
Reuse a functionality
Service Oriented Architecture (Web-Services)
▸
Reuse across platforms and protocols
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
Increasing code abstraction level
Increasing tooling and code generation
Code reusability
6
Software composition
►
Software engineering development
▸
►
From assembly and static binaries to dynamic object oriented
programming, dynamic libraries, kernel modules, etc
Software components
▸
▸
▸
▸
▸
Define and publish a standard interface
Interact by message exchanges
Ease (dynamic) composition
Modularity (no compilation time dependency)
High level message exchange protocols (format, types...)
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
7
Remote Procedure Calls
►
RPC
▸
▸
▸
►
Interface Description Language (IDL)
Communication protocol
Data exchange format
RPC system examples
▸
▸
RMI (java description, java serialization)
CORBA (Common Object Request Broker Architecture)
standard from OMG (CORBA IDL, CORBA bus)
∙
▸
GridRPC (no standard IDL, different communication
implementations)
∙
▸
http://www.corba.org
http://en.wikipedia.org/wiki/GridRPC
Web Services standard from the W3C (WSDL, SOAP, XML)
∙
http://www.w3.org/TR/{wsdl,soap}
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
8
Service-Oriented-Architectures (SOA)
in a nutshell
►
Basic principles
Messages
►
A service is an exposed piece of functionality with 3
properties:
(1) The interface contract to the service is platform-independent
(2) The service can be dynamically located and invoked
(3) Services maintain a relationship that minimizes dependencies
and only require that they maintain an awareness of each other
(loosely coupling)
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
9
SOA principles
►
Reuse, modularization
▸
▸
▸
▸
▸
Language independence
Loose coupling
Black-box: Hide internal logic from the outside world
Autonomy: Control over internal logic
Granularity:
∙
∙
►
Logic divided into services to promote reusability
Yet, functionality is presented at a granularity recognized by the user
as a meaningful service
Interoperability
▸
▸
▸
Compliance to standards (interface, protocols)
Contract: services adhere to a communication and quality of
service agreement
Composability: Collections of services can be coordinated and
assembled to form composite services
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
10
SOA principles
►
Usability
▸
Identification and categorization
∙
▸
▸
►
Possibly alternate services, providing different QoS for a given
functionality
Provisioning and delivery
Monitoring and tracking
Service framework
▸
▸
Encapsulation: Many services are consolidated to be used
under the SOA
Discoverability: Services are designed to be outwardly
descriptive so that they can be found and assessed via
available discovery mechanisms
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
11
Use-case scenario
X
►
►
►
Y
X wants to use the algorithm of Y on his data
Y wants to test his algorithm on X's data
Underlying issues:
▸
▸
How to share algorithms ?
How to share data ?
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
12
Usual solutions
►
Share codes
▸
▸
▸
Y creates a code repository (or a .tgz) with its C code
X checks it out, compiles it and tests the algorithm
Technical problems:
∙
∙
►
Share binaries
▸
▸
►
Y's code does not compile on X's machine and X does not know how (or
does not want to) to debug C
Y created a new version of the algorithm and X still uses the old one
X and Y need the same execution platform
X need to properly configure and deploy the software
License problems (source availability, ...)
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
13
Use case revisited in SOA
(2) write a client
Interface
description
(1) publish
Client
X
►
►
►
(3) invoke
Service location
Y
Y has published the interface of its algorithm (1)
X wrote a client to invoke it (2)
X invokes Web-Service (3)
▸
All data is transferred as part of the WS parameters
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
14
Web Services basics
►
Client/Server interaction
►
Platform independent and language independent
▸
►
Self-describing
▸
►
►
Client and server program can be written in different languages,
run in different environments
Once located you can ask it how to use it
Standard
Loosely coupled
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
15
Web-Services standard
►
►
Originally from IBM and Microsoft
Standardized by the W3C:
▸
Contract / Interface format
Web Services Description Language
▸
Messages format
Simple Object Access Protocol
▸
Discovery format
Universal Description Discovery & Integration
►
Based on XML
▸
▸
Text format
Platform/language independence
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
16
Web services
►
Typical use case
discover
describe
invoke
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
17
Service discovery: UDDI
►
►
►
Current version: 3.0 (feb. 2005) - OASIS
UDDI is a set of Web-Services
White pages:
▸
►
Yellow pages
▸
▸
►
Web-Services classified w.r.t service provider
Web-Services classified w.r.t service categories
There might be several yellow pages per provider
Green pages
▸
▸
Web-Services classified w.r.t technical interfaces
Typically one green page per binding of a WS
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
18
Service description: WSDL
http://w3.org/TR/wsdl
►
►
►
Describes the interface of the Web-Service
Current version: 1.1 (15th march 2001)
7 basic XML tags:
▸
Wha t ?
▸
▸
▸
How ?
▸
▸
▸
►
Types: complex types made from basic types
Message: set of parts (Input/Output parameters)
Operation: set of messages (≃method)
Port type: set of operations (≃class)
Binding: protocol used to invoke the service (e.g: SOAP/HTTP)
Port: internet address and port
Service: set of ports
Following example on an image registration Web-Service
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
19
WSDL structure: <types>
Types definition, based on XSD primitive types
<types>
<complexType name="image">
XML standard types
<sequence>
<element name="file-name" type="xsd:string" />
<element name="modality" type="xsd:string"/>
</sequence>
</complexType>
</types>
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
20
WSDL structure: <message>
Messages to be exchanged (as request or response)
<message name="registrationRequest">
<part name="reference-image" type="ns:image"/>
<part name="floating-image" type="ns:image"/>
</message>
<message name="transformation">
...
</message>
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
21
WSDL structure: <portType> and <operation>
Operations (request + response message) and collections of
operations (port type)
<portType name="registrationPortType">
<operation name="register">
<input message="registrationRequest"/>
<output message="transformation"/>
</operation>
</portType>
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
22
WSDL structure: <binding>
Protocol binding
<binding name="SOAP11Binding" type="registrationPortType">
<SOAP:binding style="rpc" transport="http"/>
</binding>
HTTP, SMTP...
“rpc”
(Remote Procedure Call)
or
“document”
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
23
WSDL structure: <port> and <service>
Service definition and service endpoint (port)
<service name="registration_service">
<port name="registration1" binding="registrationB">
<address location="http://rigid.registration.com:1234"/>
</port>
<port name="registration2" binding="registrationB">
<address location="http://bestAlgos.fr/registration"/>
</port>
</service>
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
24
Invocation: SOAP
http://w3.org/TR/soap
►
►
►
►
Describes the format of the messages
Current version 1.2 - 24thJune 2003
Various underlying protocols (HTTP, SMTP, FTP...)
SOAP envelope:
▸
▸
►
►
Header
Body: method invocation
Faults mechanism
Attachment of binary data
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
25
SOAP (over HTTP) example
POST / HTTP/1.1
Host: localhost:18006
User-Agent: gSOAP/2.7
Content-Type: text/xml; charset=utf-8
Content-Length: 544
Connection: close
SOAPAction: ""
<?xml version="1.0" encoding="UTF-8"?>
<SOAP-ENV:Envelope>
<SOAP-ENV:Body>
<ns:register>
<reference-image>
<file-name>image1</file-name>
<modality>MRI</modality>
</reference-image>
<floating-image>
<file-name>image2</file-name>
<modality>MRI</modality>
</floating-image>
</ns:register>
</SOAP-ENV:Body>
</SOAP-ENV:Envelope>
Master Ubinet: Large-Scale Distributed Computing (2)
SOAP Message
HTTP Headers
SOAP Envelope
SOAP Header
Headers
SOAP Body
Method Call & Data
Johan Montagnat
26
Interoperability issues in practice
►
Problems with specific data structures
▸
▸
Complex types
Alternatives in choosing data types
∙
∙
►
Problems with protocols supported
▸
►
Strong typing (more control, less interoperability)
Weak typing: “everything is a string” (eases interoperability, no control)
SOAP 1.1 / SOAP 1.2 / HTTP POST
Problems with encoding styles
▸
RPC / RPC literal / document literal
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
27
Other limitations of Web-Services
►
Statelessness
▸
►
Low performance (text protocols)
▸
►
Alternatives: GridRPC (DIET [http://graal.ens-lyon.fr/DIET],
NetSolve, Ninf-G) and other binary protocols
Low scalability
▸
▸
►
Services as complete black-box is a desirable property in a pure
SOA but it may bring limitations
Static endpoint declaration
No scheduling / load-balancing framework
OGSA / WS-RF extensions
▸
▸
▸
▸
Name and do bindings
Start and end services
Query, subscription, and notification
Share error messages
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
28
RESTful services
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
29
REpresentational State Transfer
►
REST is an architectural style
▸
▸
High-performance, scalable client-server interaction
Based on Web standards
∙
∙
∙
▸
Create/Read/Update/Delete (CRUD) actions only
∙
∙
►
HTTP1.1 protocol
XML and JSON data formats
URI identifiers
Binding HTTP POST/GET/PUT/DELETE messages to CRUD actions
Must conform to correct message use
RESTful services
▸
Are completely stateless
∙
▸
▸
Use HTTP methods explicitely
Expose structured data as URI trees
∙
▸
Minimal load on server to improve scalability
Make data self-explanatory
Transfer XML and/or JSON
∙
Independently of the internal client / server data structures
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
30
RESTful vs Web Services
►
RESTful services
▸
▸
▸
▸
Are typically lighter, simpler and minimize impact on server
Can only use HTTP protocol
Convert all data to XML/JSON
Require client and server to have common data structures and
context
∙
▸
Do not benefit from message extensions
∙
►
No self-explanatory WSDL description
No SOAP messaging
Web Services
▸
▸
▸
▸
Are typically heavier and leverage client / server complete
decoupling
Define WSDL contracts
Embed higher-level functionality in SOAP messages (security,
addressing, etc)
Support multiple protocols
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
31
WS-* specifications
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
32
Open Grid Service Architecture
►
OGSA aims at delivering improved services for compute
intensive applications
▸
▸
▸
►
Defines a higher level service management system
Defines a standard interface for grid services (extended WSDL)
Obsoleted and now integrated in Web Services Resources
Framework (WSRF)
OGSA services vs regular Web Services
▸
▸
▸
Transient (destroyable) grid services vs persistent web services
Factory services to manage other service instances
Stateful services: information may be stored on server side and
made available among multiple clients
∙
▸
Getting away from the “service as a black box” model
Service information metadata
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
33
Services lifecycle management
►
Services management
▸
▸
▸
▸
▸
▸
►
Define service instantiation to support decentralization
▸
►
Service creation
Life time
Fault management
State management
Publication
…
Multiple instances of a given service
But no real support for load distribution or fault tolerance
▸
It is up to services to define their load management policy and
reactions to faults
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
34
Stateful vs Stateless services
►
Discoverable and accessible state
►
A state is anything the service needs to expose
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
35
Web Services Resource Framework
►
Resources: keep web services and state separate
▸
▸
▸
States are stored externally in “resources”
Each resource has a unique key
Resources can be anything
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
36
Resources Framework
►
►
Web Service + Resource = WS-Resource
Address of a WS-resource is called an endpoint
reference
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
37
Standard interface
►
►
Service information
State representation


GetRP
►
GetMultRPs
Client
SetRP
QueryRPs
Web
Service
►

►

Master Ubinet: Large-Scale Distributed Computing (2)
SetTerminationTime
ImmediateDestruction
Notification Interfaces

►
GetRP, SetRP, QueryRPs,
GetMultipleRPs
Lifetime Interfaces

Destroy
Endpoint Reference
State Interfaces

Subscribe
SetTerm
Time
State identification

►
Resource
Resource Property
Subscribe
Notify
ServiceGroups
Johan Montagnat
38
WS-Notification
►
Event notifications (asynchronous communications)
▸
►
Naming
▸
►
Resources created following factory pattern / destroyed
Information model (monitoring & discovery)
▸
▸
►
Every resource can be uniquely referenced
Lifecycle
▸
►
Publication/Subscription model for WS
Properties associated with resources (with queriers and setters)
Asynchronous notification of changes to properties
Service Groups (registries & collective services)
▸
Group membership rules & membership management
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
39
Other frameworks
►
WS core led to a family of frameworks (WS-*) grouped in
WS-RF
Applications of the framework
(Compute, network, storage provisioning,
job reservation & submission, data management,
application service QoS, …)
WS-Agreement
(Agreement negotiation)
WS Distributed Management
(Lifecycle, monitoring, …)
WS-Resource Framework & WS-Notification
(Resource identity, lifetime, inspection, subscription, …)
Web services
(WSDL, SOAP, WS-Security, WS-ReliableMessaging, …)
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
40
Practical use of Services in largescale systems
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
41
Web-Services in practice
►
What is needed to build a Web-Service (server side)
▸
▸
▸
►
What is needed to invoke a WS (client side)
▸
▸
►
Generate WSDL from a method implementation
Receive HTTP requests (web server)
Interpret SOAP requests (stub)
Generate method header from WSDL
Write SOAP requests (stub)
Language-specific tooling for generating Web Services
Client
Business code
Native data types
data bindings
serialization
Server
stub
transport
layer
Master Ubinet: Large-Scale Distributed Computing (2)
skeleton
data bindings Service code
serialization
Johan Montagnat
Native data types
42
Web Services container
►
►
Web services: software exposing a
set of operations
Container
▸
▸
▸
►
SOAP Engine: handles SOAP requests
and responses (e.g. Apache Axis)
Application Server: provides “living
space” for applications that must be
accessed by different clients (e.g.
Tomcat)
HTTP server: also called a Web server,
handles HTTP messages
Managing contained services
▸
▸
▸
Services deployment (static / hot)
Network connections manager / firewall
Secure communications (HTTPS)
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
43
Web-Services development toolkits
►
Apache axis - http://ws.apache.org/axis2
▸
▸
►
▸
Web server: Tomcat container / Apache server
▸
Stubs and bindings: wsdl2java, java2wsdl
Sun metro - https://metro.dev.java.net/
▸
▸
▸
▸
►
Platform: UNIX, Windows
Language: Java
Platform: SUN OS, UNIX, Windows
Language: Java
Web server: Glassfish / Tomcat container
Stubs and bindings: wsgen (annotated classes), wsimport
gSOAP - http://gsoap2.sourceforge.net/
▸
▸
▸
▸
Platforms: UNIX, Windows
Language: C/C++
Web server: stand-alone / Apache module
Stubs: wsdl2h
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
44
WS development toolkits (2)
►
Microsoft .Net - http://www.microsoft.com/NET/
▸
▸
▸
►
Platform: Windows
Languages: C, C++, C#, Visual Basic
Web server: Microsoft IIS
SOAP::Lite - http://www.soaplite.com
▸
▸
▸
Platforms: UNIX, Windows
Languages: Perl, PHP
A Web-Service client in 3 lines
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
45
Globus Software: dev.globus.org
Globus Projects
MPICHG2
GridWay
Incubator
Mgmt
Common
Runtime
GT4
Java
Runtime
Delegation
MyProxy
Data
Rep
Replica
Location
C
Runtime
CAS
GSIOpenSSH
GridFTP
MDS4
GRAM
Reliable
File
Transfer
GT4 Docs
Python
Runtime
Security
OGSA-DAI
C Sec
Execution
Mgmt
Master Ubinet: Large-Scale Distributed Computing (2)
Data Mgmt
Info
Services
Johan Montagnat
Other
46
Globus Toolkit
►
►
A collection of services to address common requirements
Software for grid infrastructure
▸
►
Tools to build applications that exploit grid infrastructures
▸
▸
►
Extensible framework, WS-RF compliant
Services container
Open source & open standards
▸
►
Strong emphasis on security and distributed resources
http://www.globus.org
Enabler of a rich tool and service ecosystem
▸
▸
With large community effort and incubated services
It is “only” a toolkit: a grid middleware needs to be composed
out of it
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
47
GLOBUS Toolkit and applications
►
GOBUS supports WSRF and legacy WS services
Custom
Web
Services
Custom
WSRF
Services
Globus
WSRF Web
Services
Registry
and Admin
Globus Container
(e.g., Apache Axis)
User Applications
WS-A, WSRF, WS-Notification
WSDL, SOAP, WS-Security
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
48
Data exchanges
►
Serialization
▸
▸
▸
►
Attachment
▸
▸
▸
►
Data embedded in SOAP messages
Converted into XML (text)
Performance killer for large data sets
Binary data attached to SOAP messages
Interoperability issues
Still not efficient (in a distributed environment, the client may
not host the data)
References
▸
▸
SOAP messages only contain references to the data
The grid provides a common reference space
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
49
Sharing data on a distributed
system
►
Each file is identified by a Logical File Name (LFN)
LFNresult
Storage
Resources
LFN2
(1) register
images
(4) put result
LFN1
(5) reply(LFNresult)
X
►
►
client
(3) get
images
(2) invoke (LFN1, LFN2)
Service
location
Y
X sends LFNs of its images to Y's Web-Service (2)
Y's Web-Service sends back LFNs to the client (5)
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
50
Distributing computations
►
X1...X1000 want to access Y’s algorithm
Storage
Resources
LFNresult
X1
X2
X3
X4
LFN2
LFN1
(1) register
images
(4) get images
(3) submits/monitors
grid job
Computing
resources
...
(6) reply(LFNresult)
X10^6
►
(5) put result
(2) invoke (LFN1, LFN2)
Service
location
Y
Y submits jobs to answer computing demand
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
51
Service wrapper
►
Embedding legacy codes (non-service oriented)
▸
▸
▸
►
Common in scientific areas
Heavy-weight, difficult to recompile, modify and validate
Provide a standard interface, description and remote invocation
capability
Hiding data transfers
▸
Including grid data access protocol and data referencing
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
52
Generic Service Wrapper
►
The remote job handling can be decoupled from Y
▸
Independent from middleware evolutions
Storage Elements
(1) LFN-algo
(2) algo description
(1) register
algorithm
(2) publish
description
(3) get algorithm
description
Y
(4) register
images
(10) reply(LFNresult)
X
(5) invoke (LFN1, LFN2,algo desc.)
Master Ubinet: Large-Scale Distributed Computing (2)
LFN2
LFNresult
(9) put result
LFNalgo
(7) get algo
LFN1
(8) get images
(6) submits/monitors
grid job
Computing
resources
Generic
Service
location
Johan Montagnat
53
Generic algorithm descriptor
►
Executable access method:
▸
▸
►
Input/Output
▸
▸
►
URL
Grid file
Command-line options
Access methods (for files)
Sandbox files
▸
▸
Dynamic libraries
Config files,...
Master Ubinet: Large-Scale Distributed Computing (2)
<description>
<executable name="ExecThisCode.pl">
<access type="URL">
<path value="http://somewhere.eu/"/>
</access>
<value value="ExecThisCode.pl"/>
<input name="image" option="-im1">
<access type="LFN" />
</input>
<input name="scale" option="-s"/>
<output name="crest_lines" option="-c2">
<access type="LFN" />
</output>
<sandbox name="UtilityProgram.pl">
<access type="URL">
<path value="http://elsewhere.dk/"/>
</access>
<value value="UtilityProgram.pl"/>
</sandbox>
</executable>
</description>
Johan Montagnat
54
Generic Application Service Wrapper
►
►
Provide service wrapper to non instrumented code
Handle data transfers (references to grid data)
Algorithm parameters
description
Input 0 Input 1
Generic service
Output 0
Legacy code 1
Algorithm parameters
description
Input 0 Input 1
Generic service
Output 0
Legacy code 2
Master Ubinet: Large-Scale Distributed Computing (2)
<description>
<executable name="ExecThisCode.pl">
<access type="URL">
<path value="http://somewhere.eu/"/>
</access>
<value value="ExecThisCode.pl"/>
<input name="image" option="-im1">
<access type="LFN" />
</input>
<input name="scale" option="-s"/>
<output name="crest_lines" option="-c2">
<access type="LFN" />
</output>
<sandbox name="UtilityProgram.pl">
<access type="URL">
<path value="http://elsewhere.dk/"/>
</access>
<value value="UtilityProgram.pl"/>
</sandbox>
</executable>
</description>
Johan Montagnat
55
GASW setup
►
Generic Application Service Wrapper
▸
▸
►
Provide service wrapper to non instrumented code
Handle data transfer (references to grid data)
Invocation
User Web Server
WSDL Contract
Code descriptor
Storage Resource
Executable
Executable
Input file
Required library
Wraper script
Code descriptor
GASW host
Input files transfer
Command-line
Output files transfer
Remote Job
Master Ubinet: Large-Scale Distributed Computing (2)
Job submission
service
Johan Montagnat
Computing
resource
56
GASW invocation
Storage Resource
User Web Server
WSDL Contract
Executable
Executable
Code descriptor
Input file
Required library
Input file
Wraper script
Input files transfer
Command-line
Output files transfer
Code descriptor
Required library
GASW host
Executable
Execute
Job submission
service
Master Ubinet: Large-Scale Distributed Computing (2)
Computing
resource
Johan Montagnat
Output file
57
GASW results
Storage Resource
User Web Server
WSDL Contract
Executable
Executable
Code descriptor
Input file
Required library
Output file
Wraper script
Input files transfer
Command-line
Output files transfer
Code descriptor
GASW host
Output file
Job submission
service
Master Ubinet: Large-Scale Distributed Computing (2)
Computing
resource
Johan Montagnat
58
Service wrapper
►
Possibly supporting other non-functional concerns
▸
▸
▸
►
Security (credentials, delegation, encryption of data...)
Optimization (requests grouping, bulk submission...)
...
Bridging service-oriented environment (metacomputing
flavor) with global computing systems
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
59
Delegation service example
►
A service needs to
act on behalf of the
client
▸
►
Specific delegation
service sharing and
authentication
resource
▸
▸
►
The client hold longlived credentials
Hosting Environment
Service1
Resources
Service2
EPR
Service3
Delegate once, share across services
and invocation
Medium-lived credentials
Authentication protocol independent
approach
Master Ubinet: Large-Scale Distributed Computing (2)
Delegation Service
Delegate
Refresh
Refresh
EPR
Delegate
Client
Johan Montagnat
60
OGSA-DAI Example
►
►
OGSA-compliant Data Access and Integration Service
Objectives: sharing and accessing data resources in a
grid
▸
▸
►
Heterogeneous
Distributed
Scalability and other data-maintenance concerns
prevents the use of a centralized server
FR
Query
Client
FR
data
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
61
OGSA-DAI architecture
►
Service-based framework
▸
▸
▸
access to heterogeneous data resources
provides services for data access, integration, transformation
and delivery
executes data-centric workflows
UK
query
UK
data
ES
query
ES
data
IA
query
IA
data
OGSA-DAI service
Translate and join
FR
query
FR
data
Client
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
62
Query workflow
Convert query
query
Convert
from French
French to
to
from
English
English
Country
Capital
UK
London
France
Paris
Convert data
data
Convert
from
from
English to
to
English
French
French
Run SQL
SQL query
query
Run
SELECT Country, Capital
FROM Countries
Capital
Grande-Bretagne
Londres
France
Paris
Join
Join
the
the
data
data
SELECT Pays,Capital
FROM Pays
SELECT País, Capital FROM
Países
Convert query
query
Convert
from French
French to
to
from
Spanish
Spanish
Pays
Run SQL
SQL query
query
Run
Convert data
data
Convert
from
from
Spanish to
to
Spanish
French
French
País
Capital
España
Madrid
Italia
Roma
Master Ubinet: Large-Scale Distributed Computing (2)
Pays
Capital
Grande-Bretagne
Londres
France
Paris
Espagne
Madrid
Italie
Rome
Pays
Capital
Espagne
Madrid
Italie
Rome
Johan Montagnat
63
Cloud computing
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
64
Cloud computing
►
Delivering computing resources as a service, on demand
▸
►
Three levels of service commonly indentified
▸
▸
▸
►
Computing resources are allocated through a service interface
IaaS: Infrastructure as a Service (compute, data and network
resources)
PaaS: Platform as a Service (infrastructure + facilities for
application development, deployment, maintenance)
SaaS: Software as a Service (complete application stack)
Motivations
▸
▸
▸
▸
Convergence of services industry and infrastructure provision
Externalize management of computing resources to external
resource providers
Scale-effect: benefit from very large data-centers
Business-centric: pay resources as you go
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
65
History of on-demand computing
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
66
Cloud and Grid computing
►
►
►
Clouds focus is on resources provision
Grid focus is on middleware capability
Complementary
▸
▸
▸
Clusters, grid computing and cloud computing: the combination
of multiple computers into larger metacomputers
Clouds focus on infrastructure provision, yet require middleware
to operate
Cloud solutions are more or less flexible in terms of services /
middleware provided
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
67
Virtualization
►
Abstraction of computer resources
▸
▸
▸
▸
►
Make it possible for a single physical computer to emulate
several (different) virtual machines
Separate the OS from the underlying platform resources
Several OSes and software stack can co-exist on a physical
resource
Resources (CPU time, memory, network...) are shared among
virtual machines
Why?
▸
▸
▸
▸
▸
Hardware is underutilized, especially multi-core machines
Some services require permanent operation but limited machine
resources
Grouping services on a single host is much more energy
efficient
Data centers run out of space
Increased flexibility and lower system administration costs
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
68
Hardware virtualization principle
►
Add an hypervisor layer between OS and hardware
App App App
App App
O.S.
O.S.
App App Several Virtual Machines
O.S.
Privileged code
Hypervisor
Hardware
Privileged code
sharing hardware resources
Hardware
►
Different techniques
▸
▸
▸
Full virtualization through binary translation or CPU-assistance
Kernel-level virtualization
Paravirtualization for hypervisor-aware systems
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
69
Unmodified OS virtualization
►
Run native guest OSes on top of a host OS
O.S.
Load &
Translate
Priviledged
instructions
Privileged code
App App App
App App App
O.S.
O.S.
O.S.
Translated code
Translated code
Privileged code
Hypervisor
Privileged code
Hardware
►
Pros/cons
+ Works with any OS unmodified
+ Host OS unaware of virtualization layer
- Complex virtualization process with binary code translation
and CPU calls emulation
- All hardware devices usually not supported
- Performance impact of emulation
►
Example: VMware
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
70
Kernel-level virtualization
►
Homogeneous OSes running on the same hardware
App App App App
App App App App
Compatible O.S.
Compatible O.S.
Custom environment
Custom system libraries
Custom root file system
Custom environment
Custom system libraries
Custom root file system
Host (linux) virtualization kernel
Privileged code
Hardware
►
Pros/cons
+ reduced overhead of virtualization process
- limited to homogeneous Oses
►
Example: KVM
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
71
Para-virtualization
►
Cooperating OSes-hypervisor
▸
Each OS is controlled by a Virtual Machine Monitor (VMM) in
the hypervisor
App App App
App App App
O.S.
O.S.
Hypervisor calls
Hypervisor calls
Hypervisor
VMM
VMM
Privileged code
Hardware
►
Pros/Cons
+ Low hypervisor overhead
- With compliant OSes only
►
Example: Xen
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
72
Hardware-assisted virtualization
►
CPU support for virtualization
▸
Trapping privileged access from OS and notifying hypervisor
App App App
App App App
O.S.
O.S.
Privileged code
Privileged code
CPU call
trapping
Hypervisor
VMM
VMM
Privileged code
Hardware
►
Pros / Cons
+ Combine advantages of para-virtualization and full
virtualization (unmodified OSes support)
+ Reduced privileged code trapping overhead
- Still some performance penalty, several generations of
hardware support
►
Examples: Intel-VT and AMD-V compliant processors
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
73
More virtualization
►
Virtual memory unit virtualization
▸
▸
►
Devices virtualization
▸
▸
▸
►
Memory is the most critical component to be shared between
several guest OSes
Hardware-assist memory virtualization in latest x86 CPUs
Other devices may require virtualization (GPUs, network...)
Network virtualization gives more control / guarantee on
network bandwidth dedicated to applications
Network devices increasingly programmable in software
System manipulations by hypervisor
▸
▸
▸
▸
Suspend / resume / stop system activity
Write VM memory state on disk / restore
Migrate VM from one host to another
Checkpoint and run fail-safe backup VM...
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
74
Use of virtualization technologies
in Clouds
►
Virtual Machine
▸
▸
▸
►
Resources allocation
▸
▸
►
A virtualized machine emulates a physical computing
environment (including OS, libraries, applications) and request
for hardware resources to an hypervisor
VMs can be started (OS boot) / paused / resumed / stopped
VMs are bundled in image files, easily transported
Physical resources can be allocated to a VM on-demand (e.g.
amount of memory available, number of CPU cores...)
A computer can host as many VMs as its physical resources
can bear
Cloud infrastructures use VMs
▸
▸
For user environment installation and deployment
To easily balance load over existing physical servers
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
75
Cloud computing
►
►
Strong industry involvement
Use virtualization to adapt to all customer needs
▸
▸
▸
To adapt hardware environments to software needs rather than
the opposite
Free-up the relationship between software and hardware
Large hypervisor pool to allocate resources on-demand
∙
►
Automation
▸
►
Elastic resources provisioning framework
Virtualized VMs can easily be managed to reach garanteed
quality of services
Service-oriented
▸
▸
The cloud itself is a service accessible / configurable through an
API
Quality of service is usually guaranteed by contract
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
76
Heterogeneous interfaces
Applications for thin clients
Increasing level of complexity
and vendor lock-in
Browser-based, desktop-based, mobile
Client
Services
Applications (SaaS)
Platform (PaaS)
Storage
Infrastructure (IaaS)
(e.g. Google/Yahoo Maps, WS-2.0)
Some standardization of WS 2.0
Online apps (e.g. Google Apps)
Completely vendor specific
Cloud IDE (e.g. AppEngine)
Complete development environment (OS,
compilers, database server, web server...)
Data storage (e.g. S3)
Decouple long-term storage from computing
Computing resources (e.g. EC2)
Physical resources an hypervisors
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
77
Amazon Web Services example
http://aws.amazon.com
►
►
Extensive cloud-oriented services portofolio
Web Service-oriented API
▸
▸
►
SOAP and/or RESTful API
Bindings for several languages
From simple VM delivery to complete business hosting
▸
IaaS
∙
▸
PaaS
∙
∙
∙
∙
∙
∙
▸
Virtual servers, storage servers, VPN...
Multi-platform development environments, integration tools
Identity management, data security
Elastic scaling, workload splitting, load balancing
Platform monitoring
Databases (relational and NoSQL), web server, email server
...
SaaS
∙
∙
Data analysis services, workflow managers
Communication services, data transcoding and streaming
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
78
Conclusions
►
Service-oriented architectures widely adopted
▸
►
WS-RF architectures enable the design and deployment
of Grid middlewares
▸
▸
▸
►
Toolkits available
Interoperability issues addressed
Basics for managing collection of services
Still, tough problems have to be addressed when
deploying a large-scale distributed system
▸
▸
▸
►
Convergence of distributed computing and web communities
Scalability of all services
Fault tolerance
Deployment processes
With the emergence of the Cloud concept, infrastructure
provision also became a service
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
79
References
►
Web Services
▸
►
OGSA
▸
►
http://www.w3.org/2002/ws
“The GRID 2: blueprint for a new computing infrastructure”, Ian
Foster, Carl Kesselman. Elsevier.
Service wrapper
▸
"A Service-Oriented Architecture enabling dynamic services
grouping for optimizing distributed workflows execution". T.
Glatard, J. Montagnat, D. Emsellem, D. Lingrand. Future
Generation Computer Systems (FGCS), 24 (7), pages 720–730,
Elsevier, july 2008.
Master Ubinet: Large-Scale Distributed Computing (2)
Johan Montagnat
80

Download Report